# Lung parenchyma

## Segmentation of lung parenchyma in CT images using CNN trained with the clustering algorithm generated dataset

• 10k Accesses

• 29 Citations

• 1 Altmetric

• Metrics details

### Background

Lung segmentation constitutes a critical procedure for any clinical-decision supporting system aimed to improve the early diagnosis and treatment of lung diseases. Abnormal lungs mainly include lung parenchyma with commonalities on CT images across subjects, diseases and CT scanners, and lung lesions presenting various appearances. Segmentation of lung parenchyma can help locate and analyze the neighboring lesions, but is not well studied in the framework of machine learning.

### Methods

We proposed to segment lung parenchyma using a convolutional neural network (CNN) model. To reduce the workload of manually preparing the dataset for training the CNN, one clustering algorithm based method is proposed firstly. Specifically, after splitting CT slices into image patches, the k-means clustering algorithm with two categories is performed twice using the mean and minimum intensity of image patch, respectively. A cross-shaped verification, a volume intersection, a connected component analysis and a patch expansion are followed to generate final dataset. Secondly, we design a CNN architecture consisting of only one convolutional layer with six kernels, followed by one maximum pooling layer and two fully connected layers. Using the generated dataset, a variety of CNN models are trained and optimized, and their performances are evaluated by eightfold cross-validation. A separate validation experiment is further conducted using a dataset of 201 subjects (4.62 billion patches) with lung cancer or chronic obstructive pulmonary disease, scanned by CT or PET/CT. The segmentation results by our method are compared with those yielded by manual segmentation and some available methods.

### Results

A total of 121,728 patches are generated to train and validate the CNN models. After the parameter optimization, our CNN model achieves an average F-score of 0.9917 and an area of curve up to 0.9991 for classification of lung parenchyma and non-lung-parenchyma. The obtain model can segment the lung parenchyma accurately for 201 subjects with heterogeneous lung diseases and CT scanners. The overlap ratio between the manual segmentation and the one by our method reaches 0.96.

### Conclusions

The results demonstrated that the proposed clustering algorithm based method can generate the training dataset for CNN models. The obtained CNN model can segment lung parenchyma with very satisfactory performance and have the potential to locate and analyze lung lesions.

### Background

In recent years, segmentation has known great successes in various medical images analysis tasks including detection of atherosclerotic plaques [1], pelvic cavity assessment [2, 3], ear image data towards biomechanical researches [4], skin lesions detection [5], etc. This has led to its expansion to lung diseases detection [6, 7] and specifically to lung field extraction [8]. Lung segmentation is an incredibly important component of any clinical-decision support system dedicated to improving the early diagnosis of critical lung diseases such as lung cancer, chronic obstructive pulmonary disease (COPD), etc. [9]. However, it constitutes a very challenging task [10]. Lung segmentation is difficult to achieve due to the fact that lung pathologies present various appearances different from the normal lung tissue [11, 12]. There exist dozens of lung diseases including the ground-glass opacity, consolidation, cavity, tree-in-bud and micro nodules, nodules, pleural effusion, honeycomb, etc., and each of them possesses different shape, texture, and attenuation information at CT images [13].

With the aim of improving the early diagnosis and treatment of lung diseases, numerous studies have been conducted to segment and analyze both normal and abnormal lung from CT images. Generally, according to the study by Mansoor et al. [12], existing methods can be categorized into four classes; and each of them owns specific advantages and disadvantages. The first class is the thresholding-based methods which set a thresholding (or CT number) interval to create binary partitions [14]. These methods are the fastest, but pathological regions are often not included and various morphological operations are required. The second class is referred to as region-based methods and includes the region growing [15], graph cuts [16, 17], random walk [10, 18, 19], etc. This class of methods is fast and works well with more subtle attenuation variations. However, they presence some deficiencies such as over-segmentation problem and may fail if there exist great number of pathological findings in the lung. The third class is shape-based methods and can be further divided into two sub-classes: atlas-based and model-based. As the prior knowledge of lung shape, an atlas or model is aligned to the target images firstly, and then the atlas or model is transformed geometrically to the best segmentation through an optimization procedure [20,21,22,23]. These methods work well only for the lungs with mild and moderate abnormalities, but have difficulties while creating representative model and are computationally expensive. The fourth class is neighboring anatomy guided methods which use the spatial information of surrounding organs (e.g., rib, heart, spine, liver, spleen) to constrain the segmentation [11]. Moreover, a new trend has been becoming obvious for the segmentation of lung, i.e., the combination of different methods generate better results [12, 19]. Other surface-based methods are also available [24], and readers are encouraged to refer to the reviews for further details [12, 25, 26].

Recently, the tremendous success of machine learning techniques has attracted the attention of many researchers resulting in the development of numerous successful machine learning-based lung segmentation methods. For instance, Xu et al. [27] extracted 24 three-dimensional texture features including the first-order, second-order, fractal features, and used Bayesian classifier to discriminate five categories. Yao et al. [28] had extracted 25 features from each 16 × 16 image patch and used support vector machine (SVM) to differentiate normal from abnormal lung regions (pulmonary infection and fibrosis). Similarly, the 130 gray-level co-occurrence features extracted from 21 × 21 × 21 pixel VOI and k-nearest neighbor classifier were used to classify lung parenchyma into normal, ground glass, and reticular patterns [29]. Extracted Mobius invariant shape features and statistical texture features and SVM were also employed to detect and quantify tree-in-bud (TIB) opacities from CT images [30]. Song et al. [31] had extracted 176 texture, intensity and gradient features from each image patch and investigated the performance of four approximative classifiers. Thereafter, to the end of overcoming the difficulties encountered in the processes of features design and selection, some deep learning (i.e., convolutional neural network, CNN) and representation learning methods have been used to address the lung CT images analysis [32, 33].

Nowadays, many machine learning-based algorithms are being used to detect or distinguish various lung abnormalities and the obtained results are combined with the normal lung parenchyma segmented by other traditional methods to detect the complete lung area. Although, these methods can produce satisfactory results, their implementations comprise many processes which may need longer computational time. Moreover, in some machine learning methods especially the deep learning methods, huge amount of image patches need to be labeled or annotated manually; which is a time-consuming process and constitutes a tedious task for the radiologists. Thus, the development of a machine learning-based framework for precise segmentation of lung parenchyma from thoracic CT images will be of great help in analyzing and treating lung diseases. Additionally, an easier and low-cost accurate way for generating massive dataset used in the processes of training and validation of deep learning model is highly needed.

Motivated by the aforementioned, we propose one different strategy to segment lung parenchyma excluding lesions from CT images using a CNN trained with the clustering algorithm generated dataset. This idea originates from the observation that the normal lung parenchyma owns commonalities across subjects, diseases and CT scanners, although lung pathologies present various appearances at CT images. Segmentation of lung parenchyma can help detect and locate neighboring lung lesions, which is of great significance to the early diagnosis and treatment of lung diseases. The contributions of this paper are as follows. First, we proposed a weak supervised approach to generate large amount of CT image patches for the subsequent training and validation of CNN which can be effectively and efficiently used to replace the conventional time-consuming process of determining regions of interest (ROI) manually. Second, we designed and trained a CNN model to identify patches of lung parenchyma generated from CT images. Through this trained CNN model, the fully automatic segmentation of lung parenchyma can be achieved with excellent robustness, efficiency and accuracy. The proposed fully automated machine learning based framework for lung parenchyma segmentation possesses the potential to help researchers and radiologists locate and analyze the neighboring lesions of the lung.

### Methods

Our proposed lung parenchyma segmentation method consists of three stages: (1) the generation of the labeled dataset to be fed into the CNN; (2) the design, training and validation of a CNN model; (3) the segmentation using the trained CNN. A detailed explanation of every stage of the proposed framework is given below.

### The generation of the labeled dataset

Adopting the popular machine learning framework, all the input images with a fixed size of 512 × 512 are split into smaller patches with the same size at first. The size of patches is determined through comparing the clustering results at different settings of 64 × 64, 32 × 32, 16 × 16, 8 × 8, 4 × 4 and 2 × 2. The best patch size is set to 8 × 8, and the reason will be interpreted in the subsection of experiments. Total number of patches is 10,076,160 split from the data of 23 patients.

The complete procedure with corresponding results after each sub-step of generating the labeled dataset is illustrated in Fig. 1. At first, with all the 10,076,160 patches as input, the k-means clustering algorithm with two categories is performed twice using the mean and minimum intensity of the image patch, respectively. As shown in Fig. 1a, the lung parenchyma and the background air outside the human body have been grouped into one class of low intensity. Then, one technique named the cross-shaped verification is used to remove the group of patches of that background air. For each group of low intensity patch obtained in this step, we check whether there is at least a high intensity patch in all of its four directions (left, right, up, and down). Only if this assumption is true, the current low intensity patch will be kept and regarded as one of lung parenchyma, otherwise it will be discarded. The k-means clustering with mean intensity of patch and the cross-shaped verification possesses the ability to generate more accurate lung parenchyma boundaries, but more patches are kept in the gap between human body and the scanner bed. The same process occurs in the opposite way in the case of using the minimum intensity. Hence we apply this kind of k-means clustering algorithm twice and take the intersection of the obtained volumes.

Subsequently, connected component analysis algorithm based on Max-Tree proposed by Fu et al. [34] is applied to extract the lung parenchyma. Furthermore, padding is performed to expand the 8 × 8 patch into 32 × 32 patch without overlapping so as to meet the image input demand of the next CNN training, as shown in Fig. 1b. It is worth mentioning that the expansion of the 8 × 8 patch will result in image patch containing both lung parenchyma and body parts. However, the center of every patch is the lung parenchyma. Finally a total of 60,864 patches of lung parenchyma are generated. Correspondingly, the same number of patches belonging to non-lung parenchyma is selected out randomly for the balance of two classes. The balance of the two classification classes is performed in the aim of eliminating the decline in testing accuracy caused by the imbalance of the training dataset.

### A CNN model

We proposed a simplified CNN model possessing the ability to differentiate the real lung parenchyma image patches from non-lung parenchyma image patches. The structure of this CNN network comprises an image input layer, a convolutional layer, a pooling layer and two fully connected layers with a Softmax layer. A detailed description of the abovementioned CNN model is displayed in Fig. 2. Comparing with the well-known AlexNet structure comprising five convolutional layers, we just preserve one convolutional layer with six convolutional kernels to deal with the input set of image patches of 32 × 32. The single convolution layer is followed by rectified Linear Unit (ReLU) and normalization layers which helps accelerate the convergence of the stochastic gradient descent (SGD) and prevent overfitting as well. Then a MaxPooling layer, the first fully connected layer, a dropout layer, ReLU layer, the second fully connected layer and Softmax layer follow, respectively. The first fully connected (FC) layer includes 120 neurons. The dropout layer helps in avoiding overfitting.

### The segmentation using CNN

After splitting all the CT images for segmentation into patches of 32 × 32 using each voxel as the center point, they are input into the trained CNN. Simultaneously, each patch will be automatically given a label of 1 or 0, denoting lung parenchyma (LP) or non-lung parenchyma (NLP). The maximum connected component detection is done to extract the whole LP volume. Finally the hole in the LP volume is filled to achieve the final segmentation results of lung parenchyma.

### Experiments and image data

In the aim of implementing the machine learning-based framework proposed in this study, four main experiments have been conducted. The first experiment consists of determining the most appropriate patch size among different sizes including 64 × 64, 32 × 32, 16 × 16, 8 × 8, 4 × 4 and 2 × 2; during the stage of generating the labeled dataset of LP and NLP. The characteristics of the image patches and the computational time are assessed to determine the best patch size. The second experiment is the parameters optimization procedure of the CNN model; which will be described shortly in “CNN parameters optimization” section. The performance of the trained CNN model is tested in the third experiment. The experimental dataset used in the second and third experiments consists of 121,728 patches obtained from “The generation of the labeled dataset” section. The fourth experiment is performed using a separate dataset of 201 subjects diagnosed with lung cancer or chronic obstructive pulmonary disease (COPD) to further evaluate the performance of the trained CNN model. The details of the partition of the experimental dataset into training, validation and testing are given in Table 1.

All the experiments of this study were conducted under a Windows 7 on a workstation with CPU Intel Xeon E5-2620 v3 @2.40 GHz, GPU NVIDIA Quadro K2200 and 32 GB of RAM. The proposed CNN was implemented using the simplified AlexNet structure and the procedures of unsupervised clustering generation algorithm were implemented in MATLAB 2017a.

### CNN parameters optimization

To the end of setting the best values of the CNN parameters, numerous experiments were conducted while evaluating its performance through comparing the achieved average value of F-score (Favg) and the computational time of training of the training process. Favg is defined as

$$F_{\text{avg}} = = \frac{{Precision_{nlp} \times Recall_{nlp} }}{{Precision_{nlp} + Recall_{nlp} }} + \frac{{Precision_{lp} \times Recall_{lp} }}{{Precision_{lp} + Recall_{lp} }}$$

(1)

where Precisionnlp and Recallnlp represent the positive prediction rate and the sensitivity of the class of non-lung parenchyma, respectively. Similarly, Precisionlp and Recalllp are he positive prediction rate and the sensitivity of the class of lung parenchyma, respectively. Precision and Recall can be computed as

$$Precision = \frac{TP}{TP + FP}$$

(2)

$$Recall = \frac{TP}{TP + FN}$$

(3)

where TP is true positive, FP is false positive, and FN is false negative.

We varied up to 23 parameters during the training process of our CNN model, however, only the variation of nine of those parameters had a significant effect on the classification results. The default settings of these nine parameters can be determined as: the kernel size (5); the kernel number (6); the local response normalization layer (3); the output size of fully connected layer (120); the dropout probability (0.5); the pooling type (Max); the batch size (128); the number of epochs (50); the learning rate (0.01). Using these default settings as the reference, we adjusted each parameter while keeping the others constant and investigated the variation of Favg and the elapsed time. Specifically, 11 cases were evaluated under the circumstances of the kernel size of 10, the kernel number of 3, the channels of normalization of 1, the output of FC of 240, the dropout probability of 0.2 and 0.1, the pooling type of Avg, the batch size of 256, the epochs of 80, the learning rate of 1 × 10−5 and 1 × 10−4.

### Performance evaluation using cross-validation

A total of 121.728 image patches of 32 × 32 are divided into the training and validation datasets with a ratio of 7:1, and the 8-folder cross-validation is carried out. The relationship between the training accuracy and loss and the number of iterations is investigated. The receiver operating characteristic (ROC) curve is drawn and the area under the ROC curve (AUC) is calculated for the trained CNN model. The confusion matrix and six convolutional kernels are presented at last.

### Performance evaluation using the separate dataset and manual segmentations

One separate dataset containing 201 cases of patients was collected to evaluate the robustness, efficiency and accuracy of the trained CNN model for lung parenchyma segmentation. Among them, nine cases are patients with COPD confirmed by the pulmonary function test, and 192 cases are with lung cancer confirmed by the histopathology examination. For the cases with lung cancer, 174 cases are acquired by CT scanner, 18 cases by PET/CT scanner, whose CT images have a circular field of view. The 19,967 image slices resulted from examining all the 201 patients’ image files have been split into 4.62 × 109 image patches. The robustness is evaluated through the 201 cases data with different diseases (COPD or lung cancer) and acquired by different scanners (CT and PET/CT). The accuracy is calculated through comparing the lung field segmentation results achieved by the automated CNN model with that yielded by the manual and independent annotations of two experienced radiologists as the reference.

For a more comprehensive and clearer performance evaluation of the proposed machine learning based lung parenchyma segmentation method, four evaluation metrics have been considered including: the Dice similarity coefficient (DSC), Hausdorff distance, sensitivity, and specificity.

DSC is defined as

$$DSC\left( {V_{\text{GT}} ,V_{\text{test}} } \right) = 2\frac{{\left| {V_{\text{GT}} \bigcap {V_{\text{test}} } } \right|}}{{\left| {V_{\text{GT}} } \right| + \left| {V_{\text{test}} } \right|}}$$

(4)

where VGT is the reference standard segmentation (ground truth) obtained by the radiologists manually, Vtest is the segmentation by our proposed method.

Haussdorf distance (HD) is to measure how far apart the boundaries of our segmentation and reference (ground truth) are from each other. Let the real lung boundaries (ground truth) obtained by the radiologists and the segmentation by our proposed method be defined by VGT and Vtest, respectively. VGT comprises a set of points $$V_{GT(i)} (i = 1,2 \ldots \ldots n)$$ and Vtest as well comprises a set of points $$V_{test(j)} (j = 1,2 \ldots \ldots n)$$ [35]. Hence, HD is defined as

$$HD\left( {V_{GT} ,V_{test} } \right) = \hbox{max} \left( {\mathop {\hbox{max} }\limits_{{i \in V_{GT} }} \mathop {\hbox{min} }\limits_{{j \in V_{test} }} \left\| {i - j} \right\|,\mathop {\hbox{max} }\limits_{{j \in V_{test} }} \mathop {\hbox{min} }\limits_{{i \in V_{GT} }} \left\| {i - j} \right\|} \right)$$

(5)

Sensitivity is defined as TP/P, where P is the number of voxels in reference and TP is the number of voxels segmented correctly by the proposed method. Specificity is defined as TN/N, where N is the number of voxels not in reference, TN is the number of voxels correctly identified as non-lung parenchyma by the current method.

### The patch size

For different sizes of the image patch, the computational time, characteristics of segmentation, and the assessment are given in Table 2 and Fig. 3. The segmentation results obtained considering six different patch sizes after k-means clustering with the mean intensity of the patches for one subject with 126 slices are illustrated in Fig. 3. It can be found that the segmentation becomes more exquisite with the decrease of the patch size, while the computational time increases. Moreover, the accepted results can be obtained while the size is equivalent to or smaller than 8 × 8. The nodule and regions with high CT number are excluded from the lung parenchyma. Considering the tradeoff between the time-computation cost and the achieved segmentation results, the patch size of 8 × 8 is set as the most appropriate choice.

### Optimization of the CNN parameters

Experiments were conducted considering the default settings of the nine parameters previously mentioned in “CNN parameters optimization” section whose numerical values are shown in the first row of Table 3. The proposed system yielded a Favg validation of 0.9758 and the elapsed time for completing the training and validation was 846.53 s. Using these settings as the reference, each parameter is adjusted while keeping the others constant and the variation of Favg and the elapsed time are evaluated. The optimization of the network parameters through performing experiments while varying their values was conducted in the following manner. Doubling the size of the convolutional kernel from 5 × 5 to 10 × 10 led to the drop of Favg from 0.9758 to 0.9688 with an increase of the elapsed time by 27%. Favg also decreases slightly for both cases of the kernel number of 3 and the local response normalization layer of 1. Doubling the output size of the fully connected layer from 120 to 240 led to a slight increase of Favg from 0.9758 to 0.9765, meanwhile the network training time rises sharply to 1282.9 s. The dropout probability of 0.5 has been chosen to be the optimal because the Favg reached 0.9688 and 0.9541 for a dropout probability of 0.2 and 0.1 respectively. In opposition to its usual effects on the training results, an increase of the batch size from 128 to 256 resulted in a diminishment of the Favg from 0.9758 to 0.9659. Similarly, an augmentation of the epoch’s number to 80 resulted in a decrease of the Favg value. Last but not least, the learning rate is a very critical parameter whose decrease can significantly increase the Favg value. Specifically, the Favg reaches 0.9855 and 0.9917 for the learning rate of 0.00001 and 0.0001, respectively. The final optimized parameters employed in our trained CNN model can be found in the last row of Table 3. In summary, the relationships between Favg and most training parameters are not monotonic, and there exists an optimum condition for exploration.

### Performance evaluated using cross-validation

Using the hyper-parameter settings recorded in the last row of Table 3, the training accuracy and loss as the function of iterations number are presented in Fig. 4a, b respectively. The training accuracy and loss reach 99.08% and 0.0294, respectively, while the number of iteration reaches to 4.15 × 104, indicating the convergence without overfitting during the training process of our designed CNN model. The visualization of the six kernels of the convolutional layer after the lung parenchyma segmentation process is displayed in Fig. 4c. It is easily noticeable that the image patterns are very smooth and there is absence of noises and artifacts. Thus, the network parameters have been appropriately chosen leading to great classification performance. In addition, the ROC curve of the trained CNN is plotted in Fig. 4d represents, which clearly displays that the area under the curve (AUC) is up to 0.9991 for both lung parenchyma and non-lung parenchyma categories. Meanwhile, the confusion matrix shown in Fig. 4e indicates a sensitivity of 98.8%, a positive prediction of 99.5%, a specificity of 99.5%, a negative prediction of 98.9%, and an accuracy of 99.2%.

### Performance evaluated using the separate dataset and manual segmentations

Comparing the segmentation results achieved by CNN model with those yielded by the two radiologists for all the 402 slices (manual segmentations), the averaged DSC, HDavg, sensitivity, specificity reach 0.968/0.966, 1.40/1.48 mm, 0.909/0.906, and 0.999/0.999 (to radiologist I/to radiologist II), respectively. For each slice with 512 × 512 voxels, the averaged computational time to segment the region of lung parenchyma is 10.75 s. For a visual illustration of these performances, Figs. 5, 6 displays the segmentation results achieved by our proposed CNN model and manual segmentation on a separate dataset. From left to right, the first, second and third columns represent the segmentation by the radiologists, the segmentation by our CNN model and the hole-filling result performed on our CNN model result, respectively.

Figure 5a presents three segmentation instances of patients suffering from COPD at axial slices. It is found that most of the lung parenchyma regions have been identified and segmented with satisfactory performance. In the second column of Fig. 5a which represents the result of our CNN model, some patches of pulmonary bulla are not well segmented. After an ordinary hole-filling operation, the complete lung field can be obtained, as shown in the third column of Fig. 5a. Besides, there are other three segmentation results of subjects with lung cancer shown in Fig. 5b. Their lung parenchyma can be well distinguished from lung tumor, pleural effusion and other backgrounds. The tumor region embedded in the lung field can be extracted easily through comparing the results before and after hole-filling. Additionally, Fig. 6a demonstrates three more segmentation cases of subjects with lung cancer where the CT images are acquired by PET/CT scanners. One can find that the present CNN model also produces a good effect on lung parenchyma segmentation though the contrast and lung attenuation coefficient are quite different between images acquired by CT and PET/CT scanners.

Some details on those special cases are shown in Fig. 6b. In the first row, the pleural effusion at the right lung cannot be segmented. The hole-filling result is not able to include these regions, for they are located at the boundary of the lung field. Due to the same reason, the pulmonary bulla near the boundary of lung field also failed to be segmented, as shown in the second row of Fig. 6b. As illustrated in the third row of Fig. 6b, the inflammatory lung cancer cannot be segmented out because its CT features are completely different from those of lung parenchyma. Furthermore, a 3D visualization of some segmentation results is displayed in Fig. 7. For most cases, the segmented lung surface is smooth. There can also be seen some cavities or uncompleted lung fields resulted from the presence of the pleural effusion, pulmonary bulla and lung tumor near the boundary of the lung field.

In sum, from the averaged DSC (0.968/0.966), HDavg (1.40/1.48 mm), sensitivity (0.909/0.906), specificity (0.999/0.999) yielded by our proposed CNN model and the visualization of the segmentation results shown in Figs. 5, 6 and 7, the proposed deep learning based approach achieved very satisfactory segmentation performance.

### Semi-supervised method of generating annotations

One semi-supervised method of generating annotations has been proposed and implemented in the current study. Specifically, dual unsupervised k-means clustering, a cross-shaped verification, an intersection operation, a connected component analysis and a patch expansion are successively executed. It is well known that most of the supervised machine learning (e.g., SVM, random forest, CNN) methods require a huge amount of labeled or annotated data usually produced by experts manually. For example, 12,481 lesion patches and 16,741 normal patches are derived from the manually segmented regions in the study by Yao et al. [28]. In the experiments conducted by Song et al. [31], a total of 2062 2-D annotated ROIs were manually drawn by two radiologists. This manual annotation process is always time-consuming and high-costing. In this study we have proposed a semi-supervised method for generating annotated image patches which could effectively and efficiently help radiologists and researchers get rid of the tedious manual annotations. However, the comparison of this method with the manual annotations conducted by the medical experts remains unexplored.

### A simplified or deep CNN, and parameter optimization

We have proposed and implemented a simplified CNN model consisting of only a convolutional layer, a pooling layer and two fully connected layers with a Softmax layer. Comparing with LeNet [36] of two convolutional layers, AlexNet [37] of eight learned layers, and VGG-VD [38] of 19 layers, our CNN model is very “shallow”. However, its sensitivity of 98.9% makes the exploration of more deep CNN models not so urgent. The high sensitivity clearly justifies the ability of the proposed framework to segment lung parenchyma without many difficulties. For the more complicated lesions, the deep CNN is required because the network with more depth can better approximate the target function with high nonlinearity and achieve better feature representations [39]. For instance, Anthimopoulos et al. had proposed one deep CNN with five convolutional layers to do lung pattern classification for interstitial lung diseases [32]. Recently Shin et al. [40] have explored three CNN architectures of CifarNet, AlexNet and GoogLeNet using lymph node (LN) detection and interstitial lung disease (ILD) classification.

As done in some previous CNN studies [32, 40], the optimization of the CNN hyper-parameters is critical and inevitable. Most trends of influence of hyper-parameters on the training accuracy observed in the current study accord with previous study. More specifically, as the common choice for most CNNs, the maximum pooling yields high accuracy and is much faster in terms of convergence. It is found that the dropout and normalization are effective to accelerate the convergence. Small kernel size of 5 × 5 is better than 10 × 10. Even smaller kernels have been employed, e.g., 3 × 3 kernel in VGG-net, 2 × 3 kernel in work by Anthimopoulos et al. [32]. A relatively larger number of convolutional kernels and output units of FC, and a relatively smaller batch size and number of epochs lead to high accuracy. It is of great importance to mention that the learning rate is a very important parameter whose value needs to assign with special attention and according to the size of the objects of interest contained in the image patches to be classified.

### The proposed method comparison with the traditional methods

To the end of detecting and analyzing the various lung diseases, numerous lung segmentation methods have been proposed. Given the wide range of lung lesions, the existing methods aimed to solve different problems and they have been implemented on different image types acquired from various databases. Thus, it is quite challenging to reproduce these algorithms as well as to collect the dataset used in their experiments.

Aiming to quantitatively evaluate the performance of our proposed method, extensive experiments have been conducted using the proposed method and four commonly used lung segmentation methods including the iteration, improved Ostu, watershed and region growth methods [12]. The segmentation results achieved by each of the five methods were evaluated considering the Dice similarity coefficient (DSC) and the algorithm self-adaptability measure. The algorithm self-adaptability can be defined as the number of images successfully segmented by the system after inputting a set of images. The computed DSC and self-adaptability values have been recorded in Table 4. A comparison of those values shows that our method achieved a DSC of 0.9671 which is greater than that of the other four methods. In addition, our fully automatic machine learning based method yielded an adaptability of 100% because of the contribution of the CNN model while the iteration, improved Ostu, watershed and region growth methods yielded an adaptability of 83.33%, 83.68%, 62.5% and 91.32%, respectively. The iteration and improved Ostu methods failed to segment all the input images because of the presence of images whose cylinder-shaped background CT value and injected contrast agent value is always lower than − 1024HU and higher than 1024HU, respectively. The watershed method could not segment all the images due to the fact that the boundary of lung parenchyma is not stable and is easier to leak into the human body. Therefore, the proposed method is superior to the state-of-the-art lung segmentation methods both effectively and efficiently.

### Parenchyma at first, then the whole lung analysis in a unified framework

To the best of our knowledge, this is the first study conducted on extracting lung parenchyma from CT images using a fully machine learning-based framework, rather than the whole lung or various lung pathologies. This idea originated from one previously ignored fact that lung parenchyma is quite different from lung pathologies [11, 12, 41]. The lung parenchyma owns commonalities across subjects, diseases and CT scanners although lung pathologies exist under various appearances. An accurate segmentation of lung parenchyma may have potential to help locate and analyze the lung lesions. Current framework is compatible to further segmentation of various lesions, so the whole lung analysis might be done in a unified framework.

Our method naturally belongs to the bottom-up strategy in which only local information of the shape, texture and intensity within a 32 × 32 patch is considered. The ROI is larger than the thresholding and region-based approaches. Multiple scale technique (e.g., using the patches in different sizes simultaneously) may further improve the performance. Comparing to the top-down strategy (the model-based and neighboring anatomy-guided methods), our method presents lower computational time and can successfully segment lung with a certain level of abnormities. Last, not the least, the CNN-based methods get rid of the work of feature engineering. All the features for classification are learned from the training data and no handcrafted feature is necessary [42].

### Pixel-wise or patch-wise segmentation

Using the proposed CNN model, the average lung parenchyma regions segmentation time is estimated at 10.75 s for each slice with 512 × 512 voxels, which is a pixel-wise segmentation. The computational time can be shortened by using a multiple-core or GPU provided computer. Another alternative might be the patch-wise segmentation. For instance, one can split the CT images into 8 × 8 patches, generate 32 × 32 patches, and input them into the trained CNN to realize the segmentation of lung parenchyma, whose average computational time can be shorten to 0.16 s for each slice. Figure 8 gives some relevant examples. The lung parenchyma regions can also be segmented with reasonable accuracy for subjects with COPD (the first and second columns) and lung cancer (the third to sixth columns). The images in the third and fourth columns of Fig. 8 are obtained using CT scanners, and the others in the fifth and sixth columns are obtained from PET/CT scanners. The boundaries between lung parenchyma and lung tumor, pleural effusion, pulmonary bulla and other backgrounds can be depicted well. Therefore, our proposed segmentation framework owns the good feature of suitability of the multiple-resolution strategy.

Although our proposed segmentation method achieved quite satisfactory performance, it presents some limitations that are worth mentioning. First, the size of the patches utilized in our CNN model is fixed as 32 × 32. The effect of the patch size on the CNN performance is not investigated. Moreover, ensemble of CNN models fed with patches of different sizes may help integrate multilevel features [43]. Second, some state-of-art CNN models such as Fast RCNN and Mask RCNN have presented excellent performance for the objection detection and segmentation [44, 45]. However, their performances on the dataset generated by our clustering based method remain unexplored. In other words, these CNN models may do well in the lung parenchyma segmentation. Third, though we had built up one fully machine learning-based framework, some instances including segmentations of pulmonary nodules, consolidation, and pleural effusion need to be developed. Combination of segmentations of lung parenchyma and various lesions will demonstrate the power of our fully machine learning-based framework.

### Conclusions

A novel machine learning-based method has been presented to segment lung parenchyma from CT images automatically. To the best of our knowledge, it is the first study conducted on extracting lung parenchyma from CT images, rather than the whole lung or various lung pathologies using a fully machine learning-based framework. Moreover, a clustering method is used to automatically generate huge amount of annotated data. This clustering algorithm can properly and efficiently replace the tedious manual annotations which could significantly reduce the workload of the radiologists leading to a more accurate and faster diagnosis of the diseases. The CNN parameters have been carefully optimized through extensive experiments. Through the trained CNN, the voxel-wise identification of lung parenchyma can be achieved without any feature engineering work. Besides the cross-validation, an independent dataset of more than 200 subjects with lung cancer or COPD, acquired by CT or PET/CT scanners have been used to evaluate the performances of the CNN model. The quantitative results show that our method can segment lung parenchyma from images acquired through different imaging modalities (i.e., CT and PET/CT) with very satisfactory performance. The proposed machine learning-based framework may have the potential to help locate and analyze the lung lesions.

### References

1. 1.

Jodas DS, Pereira AS, Tavares JMRS. A review of computational methods applied for identification and quantification of atherosclerotic plaques in images. Expert Syst Appl. 2016;46:1–14.

2. 2.

Ma Z, Tavares JMRS, Jorge RN, Mascarenhas T. A review of algorithms for medical image segmentation and their applications to the female pelvic cavity. Comput Methods Biomech Biomed Eng. 2010;13(2):235–46.

3. 3.

Ma Z, Tavares JMRS, Jorge RMN. A review on the current segmentation algorithms for medical images. In: 1st international conference on imaging theory and applications (IMAGAPP), Portugal, 2009, pp. 135–40. ISBN: 978-989-8111-68-5.

4. 4.

Ferreira A, Gentiland F, Tavares JMRD. Segmentation algorithms for ear image data towards biomechanical studies. Comput Methods Biomech Biomed Eng. 2014;17(8):888–904.

5. 5.

Oliveira RB, Filho ME, Ma Z, Papa JP, Pereira AS, Tavares JMRS. Computational methods for the image segmentation of pigmented skin lesions: a review. Comput Methods Programs Biomed. 2016;131:127–41.

6. 6.

Rebouças Filho PP, da Silva Barros AC, Ramalho GL, Pereira CR, Papa JP, de Albuquerque VH, Tavares JMRS. Automated recognition of lung diseases in CT images based on the optimum-path forest classifier. Neural Comput Appl. 2017. https://doi.org/10.1007/s00521-017-3048-y.

7. 7.

Valente IRS, Cortez PC, Neto EC, Soares JM, de Albuquerque VHC, Tavares JMRS. Automatic 3D pulmonary nodule detection in CT images: a survey. Comput Methods Programs Biomed. 2016;124:91–107.

8. 8.

Rebouças Filho PP, Cortez PC, da Silva Barros AC, Albuquerque VHC, Tavares JMRS. Novel and powerful 3D adaptive crisp active contour method applied in the segmentation of CT lung images. Med Image Anal. 2017;35:503–16.

9. 9.

Ju W, Xiang D, Zhang B, Wang L, Kopriva I, Chen X. Random walk and graph cut for co-segmentation of lung tumor on PET-CT-images. IEEE Trans Image Process. 2015;24(12):5854–67.

10. 10.

Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–77.

11. 11.

Mansoor A, Bagci U, Xu ZY, Foster B, Kenneth NO, Jason ME, Anthony FS, Jayaram KU, Daniel JM. A generic approach to pathological lung segmentation. IEEE Trans Med Imaging. 2014;33(12):2293–310.

12. 12.

Mansoor A, Bagci U, Foster B, Xu ZY, Papadakis ZGZ, Folio LR, Udupa JK, Mollura DJ. Segmentation and image analysis of abnormal lungs at CT: current approaches, challenges, and future trend. RadioGraphics. 2015;35(4):1056–76.

13. 13.

Hansell DM, Bankier AA, MacMahon H, McLoud TC, Müller NL, Remy J. Fleischner society: glossary of terms for thoracic imaging. Radiology. 2008;246(3):697–722.

14. 14.

Hu S, Hoffman EA, Reinhardt JM. Automatic lung segmentation for accurate quantitation of volumetric X-ray CT images. IEEE Trans Med Imaging. 2001;20(6):490–8.

15. 15.

Wang J, Li FQ. Automated segmentation of lungs with severe interstitial lung disease in CT. Med Phys. 2009;36(10):4592–9.

16. 16.

Nakagomi K, Shimizu A, Kobatake H, Yakami M, Fujimoto K, Togashi K. Multi-shape graph cuts with neighbor prior constraints and its application to lung segmentation from a chest CT volume. Med Image Anal. 2013;17(1):62–77.

17. 17.

Dai S, Lu K, Dong J, Zhang Y, Chen Y. A novel approach of lung segmentation on chest CT images using graph cuts. Neurocomputing. 2015;168:799–807.

18. 18.

Grady L. Random walks for image segmentation. IEEE Trans Pattern Anal Mach Intell. 2006;28(11):1768–83.

19. 19.

Shi Z, Ma J, Zhao M, Liu Y, Feng Y, Zhang M, He L, Suzuki K. Many is better than one: an integration of multiple simple strategies for accurate lung segmentation in CT Images. Biomed Res Int. 2016. https://doi.org/10.1155/2016/1480423.

20. 20.

Li B, Christensen GE, Hoffman EA, Mclennan G, Reinhardt JM. Establishing a normative atlas of the human lung: intersubject warping and registration of volumetric CT images. Acad Radiol. 2003;10(3):255–65.

21. 21.

Sluimer I, Prokop M, Ginneken BV. Toward automated segmentation of the pathological lung in CT. IEEE Trans Med Imaging. 2005;24(8):1025–38.

22. 22.

Sun S, Bauer C, Beichel R. Automated 3-D segmentation of lungs with lung cancer in CT data using a novel robust active shape model approach. IEEE Trans Med Imaging. 2012;31(2):449–60.

23. 23.

Zhou J, Yan Z, Lasio G, Huang J, Zhang B, Sharma N, Prado K, D’Souza W. Automated compromised right lung segmentation method using a robust atlas-based active volume model with sparse shape composition prior in CT. Comput Med Imaging Graph. 2015;46:47–55.

24. 24.

Pu J, Roos J, Yi CA, Napel S, Rubin GD, Paik DS. Adaptive border marching algorithm: automatic lung segmentation on chest CT images. Comput Med Imaging Graph. 2008;32(6):452–62.

25. 25.

Sluimer I, Schilham A, Prokop M, Van Ginneken B. Computer analysis of computed tomography scans of the lung: a survey. IEEE Trans Med Imaging. 2006;25(4):385–405.

26. 26.

Van Rikxoort EM, Van Ginneken B. Automated segmentation of pulmonary structures in thoracic computed tomography scans: a review. Phys Med Biol. 2013;58(17):R187–220.

27. 27.

Xu Y, Sonka M, McLennan G, Guo J, Hoffman EA. MDCT-based 3-D texture classification of emphysema and early smoking related lung pathologies. IEEE Trans Med Imaging. 2006;25(4):464–75.

28. 28.

Yao J, Dwyer A, Summers RM, Mollura DJ. Computer-aided diagnosis of pulmonary infections using texture analysis and support vector machine classification. Acad Radiol. 2011;18(3):306–14.

29. 29.

Korfiatis PD, Karahaliou AN, Kazantzi AD, Kalogeropoulou C, Costaridou LI. Texture-based identification and characterization of interstitial pneumonia patterns in lung multidetector CT. IEEE Trans Inf Technol Biomed. 2010;14(3):675–80.

30. 30.

Bagci U, Yao J, Wu A, Caban J, Palmore TN, Suffredini AF, Aras O, Mollura DJ. Automatic detection and quantification of tree-in-bud (TIB) opacities from CT scans. IEEE Trans Biomed Eng. 2012;59(6):1620–32.

31. 31.

Song Y, Cai W, Zhou Y, Feng DD. Feature-based image patch approximation for lung tissue classification. IEEE Trans Med Imaging. 2013;32(4):797–808.

32. 32.

Anthimopoulos M, Christodoulidis S, Ebner L, Christe A, Mougiakakou S. Lung pattern classification for interstitial lung diseases using a deep convolutional neural network. IEEE Trans Med Imaging. 2016;35(5):1207–16.

33. 33.

Tulder GV, Bruijne MD. Combining generative and discriminative representation learning for lung CT analysis with convolutional restricted boltzmann machines. IEEE Trans Med Imaging. 2016;35(5):1262–72.

34. 34.

Fu Y, Chen X, Gao H. A new connected component analysis algorithm based on max-tree. In: 2009 eighth IEEE international conference on dependable, autonomic and secure computing, Chengdu. 2009. pp. 843–4. https://doi.org/10.1109/dasc.2009.150.

35. 35.

Yuan S, Monkam P, Zhang F, Luan F, Koomson BA. Robust active contour via additive local and global intensity information based on local entropy. J Electron Imaging. 2018;27(1):013023. https://doi.org/10.1117/1.jei.27.1.013023.

36. 36.

LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. IEEE. 1998;86(11):2278–324.

37. 37.

Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. In: International conference on neural information processing systems, vol. 25. 2012. pp. 1097–105

38. 38.

Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. Comput Sci. 2014.

39. 39.

LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436–44.

40. 40.

Shin HC, Roth HR, Gao M, Lu L, Xu Z, Nogues I, Yao J, Mollura D, Summers RM. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning. IEEE Trans Med Imaging. 2016;35(5):1285–98.

41. 41.

Yao J, Bliton J, Summers R. Automatic segmentation and measurement of pleural effusions on CT. IEEE Trans Biomed Eng. 2013;60(7):1834–40.

42. 42.

Ravi D, Wong C, Deligianni F, Berthelot M, Andreu-Perez J, Lo B, Yang GZ. Deep learning for health informatics. IEEE J Biomed Health Inform. 2017;21(1):4–21.

43. 43.

Dou Q, Chen H, Yu L, et al. Multilevel contextual 3-D CNNs for false positive reduction in pulmonary nodule detection. IEEE Trans Biomed Eng. 2016;64(7):1558–67.

44. 44.

Girshick R. Fast R-CNN. Comput Sci. 2015.

45. 45.

He K, Gkioxari G, Dollar P, et al. Mask R-CNN. In: 2017 IEEE international conference on computer vision (ICCV). IEEE Computer Society. 2017.

### Authors’ contributions

MX and SQ: proposed the idea, performed experiments, analyzed the data, made discussions and composed the manuscript together with YY (Yong Yue), YT and LX. YY (Yudong Yao) and WQ: directed the experiments and made discussions. All authors read and approved the final manuscript.

### Acknowledgements

The authors would like to thank Mr. Patrice Monkam for his valuable help in the writing of this manuscript.

### Competing interests

The authors declare that they have no competing interests.

### Availability of data and materials

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

### Consent for publication

All subjects gave written informed consent in accordance with the Declaration of Helsinki.

### Ethics approval and consent to participate

This study was approved by the Medical Ethics Committee of Shengjing Hospital of China Medical University and was in accordance with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. All subjects gave written informed consent in accordance with the Declaration of Helsinki.

### Funding

This study was supported by the National Natural Science Foundation of China under Grant (Grant number: 81671773, 61672146) and the Fundamental Research Funds for the Central Universities (N172008008).

### Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

### Author information

Author notes
1. Mingjie Xu and Shouliang Qi contributed equally to this work

### Affiliations

1. Sino-Dutch Biomedical and Information Engineering School, Northeastern University, No. 195 Chuangxin Avenue, Hunnan District, Shenyang, 110169, China

Mingjie Xu, Shouliang Qi, Yueyang Teng, Lisheng Xu, Yudong Yao & Wei Qian

2. Key Laboratory of Medical Image Computing of Northeastern University (Ministry of Education), Shenyang, China

Shouliang Qi

3. Department of Radiology, Shengjing Hospital of China Medical University, No. 36 Sanhao Street, Shenyang, 110004, China

Yong Yue

4. Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ, 07030, USA

Yudong Yao

5. College of Engineering, University of Texas at El Paso, 500 W University, El Paso, TX, 79902, USA

Wei Qian

Sours: https://biomedical-engineering-online.biomedcentral.com/articles/10.1186/s12938-018-0619-9

## Parenchyma

Lung parenchyma showing damage due to large subpleural bullae.

Parenchyma ()[1][2] is the bulk of functional substance in an animal organ or structure such as a tumour. In zoology it is the name for the tissue that fills the interior of flatworms.

### Etymology

The term parenchyma is New Latin from the Greek word παρέγχυμα parenchyma 'visceral flesh' from παρεγχεῖν parenkhein 'to pour in' from παρα- para- 'beside' + ἐν en- 'in' + χεῖν khein 'to pour'.[3]

Originally, Erasistratus and other anatomists used it to refer to certain human tissues.[4] Later, it was also applied to plant tissues by Nehemiah Grew.[5]

### Structure

The parenchyma is the functional parts of an organ, or of a structure such as a tumour in the body. This is in contrast to the stroma, which refers to the structural tissue of organs or of structures, namely, the connective tissues.

### Brain

The brain parenchyma refers to the functional tissue in the brain that is made up of the two types of brain cell, neurons and glial cells.[6] It is also known to contain collagen proteins.[7] Damage or trauma to the brain parenchyma often results in a loss of cognitive ability or even death. Bleeding into the parenchyma is known as intraparenchymal hemorrhage.

### Lungs

Lung parenchyma is the substance of the lung that is involved with gas exchange and includes the pulmonary alveoli and respiratory bronchioles,[8] though some authors include only the alveoli.[9]

### Liver

The liver parenchyma is the functional tissue of the organ made up of around 80% of the liver volume as hepatocytes. The other main type of liver cells are non-parenchymal. Non-parenchymal cells constitute 40% of the total number of liver cells but only 6.5% of its volume.[10]

### Kidneys

The renal parenchyma (of the kidney) is divided into two major structures: the outer renal cortex and the inner renal medulla. Grossly, these structures take the shape of 7 to 18[11] cone-shaped renal lobes, each containing renal cortex surrounding a portion of medulla called a renal pyramid.[12]

### Tumors

The tumor parenchyma, of a solid tumour, is one of the two distinct compartments in a solid tumour. The parenchyma is made up of neoplastic cells. The other compartment is the stroma induced by the neoplastic cells, needed for nutritional support and waste removal. In many types of tumour, clusters of parenchymal cells are separated by a basal lamina that can sometimes be incomplete.[13]

### Flatworms

Parenchyma is the tissue made up of cells and intercellular spaces that fills the interior of the body of a flatworm, which is an acoelomate. This is a spongy tissue also known as a mesenchymal tissue, in which several types of cells are lodged in their extracellular matrices. The parenchymal cells include myocytes, and many types of specialised cells. The cells are often attached to each other and also to their nearby epithelial cells mainly by gap junctions and hemidesmosomes. There is much variation in the types of cell in the parenchyma according to the species and anatomical regions. Its possible functions may include skeletal support, nutrient storage, movement, and many others.[14]

### References

1. ^"Parenchyma". Merriam-Webster Dictionary. Retrieved 2016-01-21.
2. ^"Parenchyma". Oxford Dictionaries UK Dictionary. Oxford University Press. Retrieved 2016-01-21.
3. ^LeMone, Priscilla; Burke, Karen; Dwyer, Trudy; Levett-Jones, Tracy; Moxham, Lorna; Reid-Searl, Kerry; Berry, Kamaree; Carville, Keryln; Hales, Majella; Knox, Nicole; Luxford, Yoni; Raymond, Debra (2013). "Parenchyma". Medical-Surgical Nursing. Pearson Australia. p. G–18. ISBN .
4. ^Virchow, R.L.K. (1863). Cellular pathology as based upon physiological and pathological histology [...] by Rudolf Virchow. Translated from the 2nd ed. of the original by Frank Chance. With notes and numerous emendations, principally from MS. notes of the author. 1–562. [Cf. p. 339.] link.
5. ^Gager, C. S. 1915. The ballot for names for the exterior of the laboratory building, Brooklyn Botanic Garden. Rec. Brooklyn Bot. Gard. IV, pp. 105–123. link.
6. ^"What is the Brain Parenchyma? (With pictures)".
7. ^Arachchige, Arosh S Perera Molligoda (2021-03-16). "Collagen proteins are found also within the neural parenchyma in the healthy CNS". AIMS Neuroscience. 8 (3): 355–356. doi:10.3934/Neuroscience.2021019. ISSN 2373-8006. PMC 8222768. PMID 34183986.
8. ^"Lung parenchyma". Retrieved 9 February 2016.
9. ^Suki, B (July 2011). "Lung parenchymal mechanics". Comprehensive Physiology. 1 (3): 1317–1351. doi:10.1002/cphy.c100033. ISBN . PMC 3929318. PMID 23733644.
10. ^Kmieć Z (2001). Cooperation of liver cells in health and disease. Adv Anat Embryol Cell Biol. Advances in Anatomy Embryology and Cell Biology. 161. pp. iii–xiii, 1–151. doi:10.1007/978-3-642-56553-3_1. ISBN . PMID 11729749.
11. ^Ashton, Leah; Gullekson, Russ; Hurley, Mary; Olivieri, Marion (April 1, 2017). "Correlation of Kidney Size to Number of Renal Pyramids in the Goat Kidney". The FASEB Journal. 31 (1_supplement): 899.5. doi:10.1096/fasebj.31.1_supplement.899.5 (inactive 31 May 2021) – via fasebj.org (Atypon).CS1 maint: DOI inactive as of May 2021 (link)
12. ^Walter F. Boron (2004). Medical Physiology: A Cellular And Molecular Approach. Elsevier/Saunders. ISBN .
13. ^Connolly, James L.; Schnitt, Stuart J.; Wang, Helen H.; Longtine, Janina A.; Dvorak, Ann; Dvorak, Harold F. (2003). "Tumor Structure and Tumor Stroma Generation". Holland-Frei Cancer Medicine. 6th edition.
14. ^Conn, D (1993). "The Biology of Flatworms (Platyhelminthes): Parenchyma Cells and Extracellular Matrices". Transactions of the American Microscopical Society. 112 (4): 241–261. doi:10.2307/3226561. JSTOR 3226561.

• The dictionary definition of parenchyma at Wiktionary
Sours: https://en.wikipedia.org/wiki/Parenchyma

## The Lung Parenchyma

Chest CT for Non-Radiologists pp 43-86 | Cite as

• Mary M. Salvatore
• Ronaldo Collo Go
• Monica A. Pernia M.
Chapter

First Online:

### Abstract

The lung parenchyma should be evaluated on axial images using lung window settings (window 1500, level −500) which are recorded on the computer image. Scroll using the mouse through the right lung beginning at the apex and moving toward the lung bases evaluating the lung parenchyma in front of the right major fissure which will include the right upper and middle lobes of the lung. Next scroll from the bases of the lung to the apex reviewing the lung posterior to the right major fissure, the right lower lobe. Repeat this process reviewing the left lung parenchyma. In front of the left major fissure is the left upper lobe which includes the lingual and posterior is the left lower lobe.

### Keywords

Secondary pulmonary lobule Hounsfield unit Ground glass opacity Organizing pneumonia Usual interstitial pneumonitis (UIP) Nonspecific interstitial pneumonitis (NSIP) Chronic hypersensitivity pneumonitis (CHP)

This is a preview of subscription content, log in to check access.

### References

1. 1.

Mata J, Cáceres J, Alegret X, Coscojuela P, De Marcos JA. Imaging of the azygos lobe: normal anatomy and variations. AJR Am J Roentgenol. 1991;156(5):931–7.CrossRefPubMedGoogle Scholar

2. 2.

Seiber W, Karcara N, Pant P. Pulmonary azygos lobe—an anatomical variant. Kathmandu Univ Med J. 2014;46(2):151–2.CrossRefGoogle Scholar

3. 3.

Miller WS. The lung. Springfield, IL: Thomas; 1947. p. 39–42.Google Scholar

4. 4.

Richard Webb W. Thin-section CT of the secondary pulmonary lobule: anatomy and the image—the 2004 Fleischner lecture. Radiology. 2006;239(2):322–38.CrossRefPubMedGoogle Scholar

5. 5.

Weibel ER. Looking into the lung: what can it tell us? AJR Am J Roentgenol. 1979;133:1021–31.CrossRefPubMedGoogle Scholar

6. 6.

Heitzman ER, Markarian B, Berger I, Dailey E. The secondary pulmonary lobule: a practical concept for interpretation of chest radiographs. I. Roentgen anatomy of the normal secondary pulmonary lobule. Radiology. 1969;93:507–12.CrossRefPubMedGoogle Scholar

7. 7.

Osborne DR, Effmann EL, Hedlund LW. Postnatal growth and size of the pulmonary acinus and secondary lobule in man. AJR Am J Roentgenol. 1983;140:449–54.CrossRefPubMedGoogle Scholar

8. 8.

Reid L, Simon G. The peripheral pattern in the normal bronchogram and its relation to peripheral pulmonary anatomy. Thorax. 1958;13:103–9.CrossRefPubMedPubMedCentralGoogle Scholar

9. 9.

Ochs M, Nyengaard JR, Jung A, Knudsen L, Voigt M, Wahlers T, Richter J, Gundersen HJG. The number of alveoli in the human lung. Am J Respir Crit Care Med. 2004;169(1):120–4.CrossRefPubMedGoogle Scholar

10. 10.

Itoh H, Murata K, Konishi J, Nishimura K, Kitaichi M, Izumi T. Diffuse lung disease: pathologic basis for the high-resolution computed tomography findings. J Thorac Imaging. 1993;8:176–88.CrossRefPubMedGoogle Scholar

11. 11.

Hounsfield GN. Computed medical imaging. Nobel lecture, December 8, 1979. J Comput Assist Tomogr. 1980;4(5):665–74.CrossRefPubMedGoogle Scholar

12. 12.

Zach J, Newell J, Schroeder J, et al. Quantitative CT of the lungs and airways in healthy non-smoking adults. Invest Radiol. 2012;47(10):596–602.CrossRefPubMedPubMedCentralGoogle Scholar

13. 13.

Friedman PJ. Imaging studies in emphysema. Proc Am Thorac Soc. 2008;5(4):494–500.CrossRefPubMedPubMedCentralGoogle Scholar

14. 14.

Shin KE, Chung MJ, Jung MP, Choe BK, Lee KS. Quantitative computed tomographic indexes in diffuse interstitial lung disease: correlation with physiologic tests and computed tomography visual scores. J Comput Assist Tomogr. 2011;35(2):266–71.CrossRefPubMedGoogle Scholar

15. 15.

Austin JH, Müller NL, Friedman PJ, et al. Glossary of terms for CT of the lungs: recommendations of the Nomenclature Committee of the Fleischner Society. Radiology. 1996;200:37–1.CrossRefGoogle Scholar

16. 16.

Marten K, Schnyder P, Schirg E, et al. Pattern-based differential diagnosis in pulmonary vasculitis using volumetric CT. AJR Am J Roentgenol. 2005;184(3):720–33.CrossRefPubMedGoogle Scholar

17. 17.

Ficker JH, Brückl WM, Suc J, Geise A. Haemoptysis: Intensive care management of pulmonary hemorrhage. Internist (Berl). 2017;58(3):218–25.CrossRefGoogle Scholar

18. 18.

Gluecker T, Capasso P, Schnyder P, Gudinchet F, Schaller M-D, Revelly J-P, Chiolero R, Vock P, Wicky S. Clinical and radiologic features of pulmonary edema. Radiographics. 1999;19(6):1507–31.CrossRefPubMedGoogle Scholar

19. 19.

Baik JH, Ahn MI, Park YA, Park SH. High resolution CT findings of re-expansion pulmonary edema. Korean J Radiol. 2010;11:164–8.CrossRefPubMedPubMedCentralGoogle Scholar

20. 20.

Beigelman-Aubry C, Godet C, Caumes E. Lung infections: the radiologist’s perspective. Diagn Interv Imaging. 2012;93(6):431–40.CrossRefPubMedGoogle Scholar

21. 21.

Algin O, Gökalp G, Topal U. Signs in chest imaging. Diagn Interv Radiol. 2011;17(1):18–29.PubMedGoogle Scholar

22. 22.

Gharib AM, Stern EJ. Radiology of pneumonia. Med Clin North Am. 2001;85(6):1461–91.CrossRefPubMedGoogle Scholar

23. 23.

Patsios D, Roberts HC, Paul NS, Chung T, Herman SJ, Pereira A, Weisbrod G. Pictorial review of the many faces of bronchioloalveolar cell carcinoma. Br J Radiol. 2007;80(960):1015–23.CrossRefPubMedGoogle Scholar

24. 24.

Walter JM, Stanley M, Singer BD. Metastatic pulmonary calcification and end-stage renal disease. Cleve Clin J Med. 2017;84(9):668–9.CrossRefPubMedPubMedCentralGoogle Scholar

25. 25.

Gietema HA, Muller NL, Fauerbach PV, Sharma S, Edwards LD, Camp PG, Coxson HO. Quantifying the extent of emphysema: factors associated with radiologists’ estimations and quantitative indices of emphysema severity using the eclipse cohort. Acad Radiol. 2011;18:661–71.CrossRefPubMedGoogle Scholar

26. 26.

Smith BM, Austin JH, Newell JD Jr, D’Souza BM, Rozenshtein A, Hoffman EA, Ahmed F, Barr RG. Pulmonary emphysema subtypes on computed tomography: the MESA COPD study. Am J Med. 2014;127(1):94.e7–23.CrossRefGoogle Scholar

27. 27.

Hogg JC. Pathophysiology of airflow limitation in chronic obstructive pulmonary disease. Lancet. 2004;364:709–21.CrossRefPubMedPubMedCentralGoogle Scholar

28. 28.

Tunsupon P, Dhillon SS, Harris K, Alraiyes AH. Saber-sheath trachea in a patient with severe COPD. BMJ Case Rep. 2016;2016. https://doi.org/10.1136/bcr-2016-214648.

29. 29.

Hansell DM, Bankier AA, MacMahon H, McLoud TC, Müller NL, Remy J. Fleischner society: glossary of terms for thoracic imaging. Radiology. 2008;246(3):697–722.CrossRefPubMedGoogle Scholar

30. 30.

31. 31.

Ferrans VJ, Yu ZX, Nelson WK, Valencia JC, Tatsuguchi A, Avila NA, et al. Lymphangioleiomyomatosis (LAM): a review of clinical and morphological features. J Nihon Med Sch. 2000;67(5):311–29.CrossRefGoogle Scholar

32. 32.

Zar HJ. Chronic lung disease in human immunodeficiency virus (HIV) infected children. Pediatr Pulmonol. 2008;43(1):1–10.CrossRefPubMedGoogle Scholar

33. 33.

Tokuyasu H, Watanabe E, Okazaki R, Kawasaki Y, Kikuchi R, Isowa N, et al. Sjögren’s syndrome with multiple bullae caused by lymphocytic interstitial pneumonia. Lung. 2007;185(3):187–8.CrossRefPubMedGoogle Scholar

34. 34.

Girschikofsky M, Arico M, Castillo D, et al. Management of adult patients with Langerhans cell histiocytosis: recommendations from an expert panel on behalf of Euro-Histio-Net. Orphanet J Rare Dis. 2013;8:72.CrossRefPubMedPubMedCentralGoogle Scholar

35. 35.

Sekar T, Swan KG, Vietrogoski RA. A beeline through Sir Peter James Kerley’s life. AJR Am J Roentgenol. 2011;196(4):W375–9.CrossRefPubMedGoogle Scholar

36. 36.

37. 37.

Prakash P, Kaira MK, Sharma A, Shepard JA, Digumarthy SR. FDG PET/CT in assessment of pulmonary lymphangitic carcinomatosis. AJR Am J Roentgenol. 2010;194(1):231–6.CrossRefPubMedGoogle Scholar

38. 38.

Herold CJ, Bankier AA, Fleishmann D. Lung metastases. Eur Radiol. 1996;6:596.CrossRefPubMedGoogle Scholar

39. 39.

Kebbe J, Abdo T. Interstitial lung disease: the diagnostic role of bronchoscopy. J Thorac Dis. 2017;9(Suppl 10):S996–S1010.CrossRefPubMedPubMedCentralGoogle Scholar

40. 40.

Travis WD, Costabel U, Hansell DM, King TE Jr, Lynch DA, Nicholson AG, et al. An Official American Thoracic Society/European Respiratory Society statement: update of the international multidisciplinary classification of the idiopathic interstitial pneumonias. Am J Respir Crit Care Med. 2013;188(6):733–48.CrossRefPubMedPubMedCentralGoogle Scholar

41. 41.

Bagnato G, Harari S. Cellular interactions in the pathogenesis of interstitial lung diseases. Eur Respir Soc. 2015;24(135):102–14.CrossRefGoogle Scholar

42. 42.

Urushiyama H, Terasaki Y, Nagasaka S, Terasaki M, Kunugi S, Nagase T, Fukuda Y, Shimizu A. Role of α1 and α2 chains of type IV collagen in early fibrotic lesions of idiopathic interstitial pneumonias and migration of lung fibroblasts. Lab Invest. 2015;95(8):872–85.CrossRefPubMedGoogle Scholar

43. 43.

Todd NW, Atamas SP, Luzina IG, Galvin JR. Permanent alveolar collapse is the predominant mechanism in idiopathic pulmonary fibrosis. Expert Rev Respir Med. 2015;9(4):411–8. https://doi.org/10.1586/17476348.2015.1067609.CrossRefPubMedGoogle Scholar

44. 44.

Salvatore M, Henschke CI, Yip R, Jacobi A, Eber C, Padilla M, Knoll A, Yankelevitz D. JOURNAL CLUB: evidence of interstitial lung disease on low-dose chest CT images: prevalence, patterns, and progression. AJR Am J Roentgenol. 2016;2016(3):487–94.CrossRefGoogle Scholar

45. 45.

Miller WT, Shah RM. Isolated diffuse ground-glass opacity in thoracic CT: causes and clinical presentations. Chest Imaging. 2005;184(2):613–22.Google Scholar

46. 46.

Piciucchi S, Tomassetti S, Ravaglia C, Gurioli C, Gurioli C, Dubini A, Carloni A, Chilosi M, Colby TV, Poletti V. From “traction bronchiectasis” to honeycombing in idiopathic pulmonary fibrosis: a spectrum of bronchiolar remodeling also in radiology? BMC Pulm Med. 2016;16(1):87.CrossRefPubMedPubMedCentralGoogle Scholar

47. 47.

Arakawa H, Honma K. Honeycomb lung: history and current concepts. AJR Am J Roentgenol. 2011;196:773–82.CrossRefPubMedGoogle Scholar

48. 48.

Raghu G, Collard HR, et al. An official ATS/ERS/JRS/ALAT statement: idiopathic pulmonary fibrosis: evidence-based guidelines for diagnosis and management. Am J Respir Crit Care Med. 2011;183(6):788–824.CrossRefPubMedPubMedCentralGoogle Scholar

49. 49.

Lynch DA, Sverzellati N, Travis WD, Brown KK, Colby TV, Galvin JR, Goldin JG, Hansell DM, Inoue Y, Johkoh T, Nicholson AG, Knight SL, Raoof S, Richeldi L, Ryerson CJ, Ryu JH, Wells AU. Diagnostic criteria for idiopathic pulmonary fibrosis: a Fleischner Society White Paper. Lancet Respir Med. 2018;6:138–53.CrossRefPubMedGoogle Scholar

50. 50.

Montesinos JJ, Laguna MA. Case 1: cryptogenic organizing pneumonia. AJR Am J Roentgenol. 1998;171(3):835. 838–9.CrossRefPubMedGoogle Scholar

51. 51.

Pereira CA, Gimenez A, Kuranishi L, Storrer K. Chronic hypersensitivity pneumonitis. J Asthma Allergy. 2016;9:171–81.CrossRefPubMedPubMedCentralGoogle Scholar

52. 52.

Sherrick AD, Swensen SJ, Hartman TE. Mosaic pattern of lung attenuation on CT scans: frequency among patients with pulmonary artery hypertension of different causes. AJR Am J Roentgenol. 1997;169(1):79–82.CrossRefPubMedGoogle Scholar

53. 53.

Alhamad EH, Al-boukai AA, Al-kassimi FA, et al. Prediction of pulmonary hypertension in patients with or without interstitial lung disease: reliability of CT findings. Radiology. 2011;260(3):875–83.CrossRefPubMedGoogle Scholar

54. 54.

Sreenivasan S, Bennett S, Parfitt VJ. Images in cardiovascular medicine. Westermark’s and Palla’s signs in acute pulmonary embolism. Circulation. 2007;115:e211.CrossRefPubMedGoogle Scholar

© Springer International Publishing AG, part of Springer Nature 2018

### Authors and Affiliations

• Mary M. Salvatore
• Ronaldo Collo Go
• Monica A. Pernia M.
1. 1.RadiologyIcahn School of Medicine at Mount SinaiNew YorkUSA
2. 2.Division of Pulmonary, Critical Care, and Sleep MedicineCrystal Run Health CareMiddletownUSA
3. 3.Internal MedicineNew York Medical College - Metropolitan Hospital ProgramNew YorkUSA
Lung Parenchyma What is Lung Parenchyma Lung Parenchyma

The lung parenchyma is initially abnormal on chest radiographs in about one-half of patients. The percentage is higher on computed tomography which can detect disease when the radiograph is normal. If lung tissue is obtained, however, there is histologic disease in almost all patients, including those who have no visible lung abnormalities on imaging studies. This is why transbronchial biopsy has such a high diagnostic yield and is the favored means of obtaining tissue for diagnosis.

The involvement of the lung parenchyma by sarcoidosis can have very different appearances ranging from reticulonodular lung tissue densities to multiple masses and miliary lesions. These radiographic expressions of sarcoidosis can mimic many other diseases including malignancies and infections which involve the lungs. There are certain radiologic features, however, that are very useful in limiting the differential diagnosis of lung parenchymal disease in these patients. For example, air containing bronchi in mass like lesions and small nodules distributed along lymphatic vessels are important features of sarcoidosis. These findings are illustrated in this section.

The radiologic patterns of lung disease are produced by interstitial granulomas and associated interstitial densities. Chest radiographs most often show well defined linear and nodular densities which are characteristic of lung tissue disease. Each nodule represents mutiple coalescent granulomas since an individual granuloma is microscopic.

RETICULONODULAR PATTERN CLOSEUP OF RETICULONODULAR PATTERN

#### Note well defined linear and nodular densities characteristic of lung tissue (interstitial) disease.

With progession nodules can compress small peripheral airways resulting in poorly defined confluent densities which have an appearance similar to pneumonia and other diseases which fill the alveoli. Larger conglomerations of granulomas can appear as tumor like masses. Miliary and cavitary lesions are rare. Rarely endobronchial nodules or lymphadenopathy obstruct bronchi and cause atelectasis.

Most parenchymal lung disease resolves. A minority of patients develop irreversible pulmonary fibrosis with disability ranging from minimal to death. This Stage IV disease is often most severe in the upper lobes. The radiographic findings can range from minimal to extensive. These findings include: fibrosis with irregular septal thickening, broad bands and masses of fibrous tissue, traction bronchiectasis, upper lobe volume loss, emphysema, bullae, and honeycomb end stage lung.

Pulmonary fibrosis can lead to cor pulmonale and right heart failure. Bullae can be colonized by fungi, most often anAspergillusspecies. This results in a mobile ball of hyphae in the bulla. The formation of a fungal ball is sometimes preceded by thickening of the wall of the bulla or the adjacent pleura.

STAGE IV STAGE IV

### RADIOGRAPHIC PATTERNS OF LUNG DISEASE

• Reticular, nodular, or reticulnodular densities ( common )
• Acinar type poorly defined nodules, pneumonic appearance ( common )
• Pulmonary fibrosis, emphysema. Upper lobe predominance ( 20% )
• Multiple large nodules. Some with air bronchgrams ( unusual )
• Miliary lesions ( unusual )
• Multiple cavitary lesions ( rare )
ACINAR PATTERN PNEUMONIC APPEARANCE

### COMPUTED TOMOGRAPHY

Computed tomography can show disease when radiographs do not and can also better demostrate lymphadenopathy and the characteristic features of sarcoidosis involving the lung parenchyma. This can be important when the diagnosis is uncertain.

The small ( 2-10mm ) nodules are characteristic of sarcoidosis. These nodules are primarily distributed in the lung tissue along lymphatics. Many are located along the bronchovascular bundles, in the subpleural interstitial space beneath the viseral pleura, and in the interlobular septae. These nodules are much easier to see on computed tomography studies than on radiographs. Computed tomography done with thin ( 1-2mm ) sections and the highest spatial resolution possible is known as high resolution CT (HRCT). HRCT is the best way to show details of lung tissue disease, except for small nodules. On 1-2mm sections, blood vessels in cross section are difficult to differentiate from lung nodules.

Several HRCT images following conventional CT imaging are often used to better show lung tissue disease.

Patchy abnormal increased density of the lung with preserved visibility of the underlying anatomy is called ground glass density. Ground glass density is common on HRCT of sarcoidosis but is not specific. It has been hypothesized that in patients with sarcoidosisis this density is due to alveolitis, but there is no convincing proof that sarcoidosis causes alveolitis, and in addition, ground glass density can be produced by lung tissue disease. Ground glass density, however, probably does correlate with active reversible disease.

None of the radiographic or CT findings that can be produced by sarcoidosis are pathognomonic. The small nodules of sarcoidosis are characteristic, but small lung nodules similar to sarcoidosis are also common in lymphangitic lung metastases and silicosis. Lymphangitic metastases to the lung can closely resemble sarcoidosis. Small nodules distributed along lymphatics are common to both. Lymphangitic metastases, however, cause greater septal thickening and less distortion of the lobular anatomy than sarcoidosis. Silicosis is also characterized by small nodules and can look similar to sarcoidosis, but the history and clinical findings are distinctive.

### CT FINDING IN SARCOIDOSIS

• Small nodules: peribronchovascular, subpleural including fissures, lobular septae
• Small nodules producing thickened or nodular vessels, lobular septae and bronchial walls
• Acinar type poorly defined nodules, pneumonic appearance
• Multiple patchy areas of ground glass density
• Multiple large masses, some with air bronchograms
• Late fibrosis, emphysema, bronchiectasis, upper lobe volume loss, honeycomb lung
• Upper lobe predominance
NODULAR PATTERN SUBPLEURAL NODULES

#### The cluster of small nodules here looked like a tumor on a radiograph. Note nodular thickening of the major fissure which is a typical distribution of sarcoid nodules..

ALVEOLAR SARCOIDOSIS ALVEOLAR SARCOIDOSIS

#### Computed tomography shows a mass which has air containing bronchi (arrows) within it.. In addition to sarcoidosis, bronchioloalveolar carcinoma, lymphoma, and pseudolymphoma can present as a mass with air bronchograms.

MILIARY SARCOIDOSIS CAVITARY SARCOIDOSIS

## Parenchyma lung

Open Access

Peer-reviewed

• Xiaolei Liao,
• Juanjuan Zhao,
• Cheng Jiao,
• Lei Lei,
• Yan Qiang ,
• Qiang Cui
• Xiaolei Liao,
• Juanjuan Zhao,
• Cheng Jiao,
• Lei Lei,
• Yan Qiang,
• Qiang Cui

x

### Background

Lung parenchyma segmentation is often performed as an important pre-processing step in the computer-aided diagnosis of lung nodules based on CT image sequences. However, existing lung parenchyma image segmentation methods cannot fully segment all lung parenchyma images and have a slow processing speed, particularly for images in the top and bottom of the lung and the images that contain lung nodules.

### Method

Our proposed method first uses the position of the lung parenchyma image features to obtain lung parenchyma ROI image sequences. A gradient and sequential linear iterative clustering algorithm (GSLIC) for sequence image segmentation is then proposed to segment the ROI image sequences and obtain superpixel samples. The SGNF, which is optimized by a genetic algorithm (GA), is then utilized for superpixel clustering. Finally, the grey and geometric features of the superpixel samples are used to identify and segment all of the lung parenchyma image sequences.

### Results

Our proposed method achieves higher segmentation precision and greater accuracy in less time. It has an average processing time of 42.21 seconds for each dataset and an average volume pixel overlap ratio of 92.22 ± 4.02% for four types of lung parenchyma image sequences.

Citation: Liao X, Zhao J, Jiao C, Lei L, Qiang Y, Cui Q (2016) A Segmentation Method for Lung Parenchyma Image Sequences Based on Superpixels and a Self-Generating Neural Forest. PLoS ONE 11(8): e0160556. https://doi.org/10.1371/journal.pone.0160556

Editor: Yuanquan Wang, Beijing University of Technology, CHINA

Received: March 11, 2016; Accepted: July 21, 2016; Published: August 17, 2016

Copyright: © 2016 Liao et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All data have been uploaded to Figshare. The URL is: (https://figshare.com/s/254e3467efd57a442334).

Funding: The work is supported by the National Natural Science Foundation of China (61373100, 61540007) and the National Key Laboratory Open Foundation of China (BUAA-VR-15KF02, BUAA-VR-16KF13).

Competing interests: The authors have declared that no competing interests exist.

### Introduction

Lung cancer is one of the most common causes of cancer-related death worldwide [1]. Computed tomography (CT) [2] scanning technology has good density resolution for lesions in the human body and is currently the most effective and direct imaging method for the early diagnosis of lung cancer. However, as the accuracy requirements for clinical imaging of lesions increase, the CT scanning thickness decreases, and a large number of CT image sequences need to be produced [3]. The massive amount of image data will inevitably increase the challenge of CT image processing, leading to a slow processing speed and decreased efficiency. In addition, because each pulmonary CT image presents a different morphological structure from the top to the bottom of the lung in CT image sequences, the general segmentation algorithm is not effective. Therefore, determining how to segment the lung parenchyma image sequences quickly without reducing accuracy is of great significance for the subsequent segmentation of pulmonary nodules and benign and malignant diagnoses.

Lung segmentation can be an important component of computer-aided diagnosis (CAD) systems [4]. Geng H’s group used an iterative gray threshold to select seed points automatically and then extract each lung parenchyma image with the region growing method, which is sensitive to background noise [5]. Liming D and colleague present a new form of lung parenchyma segmentation. The optimal threshold value method and the boundary tracking method are used to segment the lung region and can effectively eliminate the influence of background noise but may lose some of the lung parenchyma [6]. Mansoor A and coworkers segmented the lung parenchyma in two steps [7] by using the fuzzy connectedness (FC) image segmentation algorithm to perform the initial lung parenchyma extraction and then texture-based local descriptors to segment abnormal imaging patterns using a near-optimal keypoint analysis. However, this method is not effective for processing irregular images. Wavelet transform has been applied by Shojaii R [8] to decompose an image into several regions, and the regions with low pixel intensities are kept and grown to segment the honeycomb regions. This method can effectively segment irregular lung parenchyma images. Yan-hua and coworkers used several methods, including the optimal iterative threshold, three-dimensional connectivity labeling, and three-dimensional region growing methods, for the initial segmentation of the lung parenchyma and used the morphological method to repair the lung parenchyma [9]. Luo X [10] and others used an improved active contour model, which can obtain better segmentation result with the help of artificial segmentation but is very time consuming. There are also some scholars who use superpixel to segment medical images. Yu N and Weinstein S P [11] proposed a novel automatic segmentation framework for tumor on breast DCE-MRI images by using graph-cuts and superpixel classification, which can achieve a classification accuracy of 96%.

Superpixel was also used for bacteria cell segmentation by Song Y’s group [12]. Features of superpixels are extracted and trained by supervised deep learning method with an accuracy of 99% and a sensitivity of 100% for four types of different bacteria.

In general, lung segmentation methods are based on threshold, region, and mathematical morphology. However, lung CT images are sequential, and the existing methods in the study of lung parenchyma segmentation algorithms are generally for single-image segmentation of CT images and ignore the before-to-after image correlation. A few scholars have studied sequential image segmentation, but this often involves a long processing time, low efficiency and extensibility.

In this paper, we use the position particularity of the lung parenchyma in lung CT images, fully consider the strong correlation between adjacent slices of CT image sequences, and put forward a segmentation method for lung parenchyma image sequences based on superpixels and a self-generating neural forest. The experimental results show that our proposed method can significantly increase the speed of segmentation for four types of lung parenchyma images, which guarantees accuracy and integrity.

### 2.1 Materials

#### 2.1.1 Ethics statement.

This study was approved by the institutional review board (IRB) of the Coal Center Hospital in Shanxi. The study was conducted in accordance with the hospital’s ethics requirements. Informed consent was obtained from all patients for being included in the study.

#### 2.1.2 Datasets.

The CT image datasets used in this study were obtained from a hospital in Shanxi Province, China. All data can be accessed at https://figshare.com/s/254e3467efd57a442334. We used a Discovery ST16 PET-CT scanner from the General Electric Company of America (150 mA, 140 kV, with a slice thickness of 3.75 mm). In the experiment, we select lung CT sequence image datasets from 80 people with a total of 4812 CT images, and the size of each image was 512 × 512. Based on the physician’s prior knowledge and the morphological perspective of lung CT image sequences, the 80 datasets were divided into four categories: without nodules, benign nodules, malignant SPN (solitary pulmonary nodules) and pleural nodules. Each category had 20 datasets and approximately 1200 CT images.

### 2.2 Proposed Method

We propose a segmentation method for lung parenchyma image sequences based on superpixels and a self-generating neural forest that mainly involves a gradient and sequential linear iterative clustering algorithm (GSLIC) to obtain superpixels, clustering of superpixels with a self-generating neural forest (SGNF), and lung parenchyma image sequences segmentation. A block diagram of the lung parenchyma image sequences segmentation is shown in Fig 1.

#### 2.2.1 Gradient and sequential linear iterative clustering (GSLIC).

It has been difficult to obtain better segmentation in a relatively short time in lung parenchyma image sequence segmentation. To solve this problem, we used multiple CT images and drew on the prior knowledge of the physicians to propose a superpixel segmentation algorithm for image sequence. Our method is based on a gradient and sequential linear iterative clustering (GSLIC) algorithm and includes lung ROI image sequence extraction and ROI sequence superpixel segmentation.

1. Lung ROI sequence extraction: Because the position of the lung parenchyma region in CT images is relatively fixed, we increased the running speed by first extracting the lung ROI sequences using a statistical method for ROI extraction that is adopted in this paper. More than 4800 CT lung images of 80 individuals were analyzed, and we determined that the rectangles in the upper-left (100, 60) and lower-right corners (400,420) could include all lung parenchyma regions. Therefore, we can obtain the entire lung ROI sequence based on the two points in the CT image sequence. In Fig 2, the original lung CT image (a) was used to extract its ROI image (b). Extracting the lung ROI sequences can reduce the processing time and simultaneously eliminate some noise.

2. Superpixel segmentation on ROI image sequences: The concept of superpixels was first put forward by Ren [13] in 2003. A superpixel is a collection of pixels with similar characteristics, such as color, brightness, and texture. An image can be composed of a certain number of superpixels that contain multiple combination characteristics of the pixels and can preserve the edge information of the original image. Compared with a single pixel, a superpixel contains rich characteristic information and can greatly reduce image post-processing complexity and significantly increase the speed of image segmentation.

Traditional superpixel segmentation uses a process of simple, linear iterative clustering. After this method was improved by Lucchi, Hammoudi, and Wang J [14–16], it was applied to a single image segmentation. In this paper, an algorithm based on gradient and sequential linear iterative clustering (GSLIC) is proposed to segment image sequences. Each pixel in the sequence of lung CT images can be represented by a six-dimensional feature vector ([l, a, b, x, y, z]T). The similarity between the pixels can be measured by the Euclidean distance between them. A pixel’s feature vector is made up of its color vector [l, a, b] in CIELAB color space and its space coordinate vector [x, y, z], where x and y are the pixel coordinates, and z is the serial number of the image. The GSLIC procedure is shown in Table 1.

In an original CT image with N pixels that needs to be divided into K superpixels, each superpixel contains approximately N/K pixels, and therefore, the average length of each superpixel S is about (1).

(1)

We first take an initial clustering center every S pixels and then select cluster centers using a 3 * 3 nuclear window to the lowest gradient position. When selecting the initial clustering center, a method of gradient descent is adopted to sample pixels at a regular grid so that the edge points are not selected as the cluster centers. By using a 3 * 3 nuclear window, a pixel’s gradient G(x, y) can be defined as (2). (2) The Min {G(x, y)} coordinates in each grid can be chose as the cluster center. Following that, each clustering center can be search for neighboring similar pixels around the search space for 2S * 2S based on the similarity of Ds between the pixels.

In our method, for the same CT image with serial number z, the similarity of Ds between the pixels ([lj, aj, bj, xj, yj, z]T) to the clustering center ([li, ai, bi, xi, yi, z]T) can be calculated by their color feature distance Dlab and space feature distance Dxy. The calculation formulas of Dlab, Dxy and Ds are as follows in (3), (4) and (5).

(3)(4)(5)

In (5), δ is a parameter to adjust the weight of Dlab and Dxy. The larger the value, the bigger the weight of Dxy to calculate Ds will be, which is generally between 1 and 20. The result of superpixels segmentation on ROI image is shown in Fig 3.

The proposed method can segment an image into a series of superpixels, each of which can also be expressed by feature vector ([l, a, b, x, y, z]T). In the process of obtaining superpixels, effective clustering centers can be chose by gradient descent, and the blocks can be obtained through the clustering algorithm. In addition, taking the correlation between the sequences of CT images into account, the coordinate information of the clustering centers in the previous image is directly transmitted to the next image, which can significantly improve the image’s superpixel segmentation speed.

#### 2.2.2 Superpixel clustering based on SGNF and GA.

In this paper, a method using a self-generated neural forest (SGNF) algorithm optimized by a genetic algorithm (GA) was proposed to cluster the superpixels. The genetic algorithm is used to select the optimal clustering centers, which are used to generate the neural trees that form the neural forest. Our method effectively overcomes the instability of the primary SGNN and improves the efficiency and accuracy of clustering.

1. Self-generating neural tree (SGNT): Self-generated neural networks (SGNNs) were developed in 1992 [17] using a competitive learning mechanism for samples learning. A self-generating neural tree (SGNT) is generated by an SGNN using unsupervised learning.

An SGNT includes neurons, weights, and connections. In this paper, we use an ordered pair < {nj}, {lk} > to express an SGNT, where {nj} is the set of neurons, and {lk} is the set of connections. Each neuron also can be represented as ordered pair < w, {nc}>, where w is the weight of the neuron, and {nc} is the set of child neurons of the neuron. Each leaf neuron corresponds to a sample, and each root neuron is a cluster center. All leaf neurons of the root neuron belong to the same cluster, and the weight of every neuron is the average attribute of all the leaf neurons it covers.

Therefore, the structure of an SGNN shows simplicity and a good self-organizing capability and learning speed. It is beneficial to learn clustering that has high performance. Fig 4 shows the structure of an SGNT with five samples. Fig 4(A) lists a clustering sample set where Wj, j = a, b, …, e are the sample attributes. Fig 4(B) is the SGNT generated by following SGNT generating rules [18–20].

2. Self-generating neural forest (SGNF) optimized by GA: Despite its good capacity for clustering, an SGNT is influenced by the input order of the samples [21]. To solve this problem, we propose an adaptive clustering algorithm that is optimized by GA and in which the SGNT is generalized to a self-generating neural forest (SGNF), and the GA is applied to select optimal superpixel seeds as the initial input into the SGNF. The process of clustering with an SGNF optimized by a GA is shown in Table 2.

Each SGNT in an SGNF corresponds to a cluster, and all the leaf neurons in an SGNT belong to the same cluster. In section 2.2.1, we show that each superpixel can be expressed by a feature vector ([l, a, b, x, y, z]T). For given sample {Xi} where i = 1, 2, …, L, the distance between sample Xj and clustering center Xi can be calculated as (6): (6) where k is the sequence number of the element in the feature vector, and wjk is the weight of the first k attribute. When processing the superpixels segmented by GSLIC, the attributes of superpixels, such as color and coordinate feature values, can be used to generate an SGNF. The result of superpixel clustering with GA-SGNF is shown in Fig 5.

Different K seed samples generate a SGNF with different structures. Therefore, the choice of K seed samples can be seen as an optimization problem. To obtain preferable clustering results, the GA is used to search for K seed samples to optimize the clustering results as described in the next section.

3. Genetic Algorithm: The genetic algorithm (GA) is a method based on the probability search technology of population optimization [22–24]. The GA has a good ability for global searching and can search for the optimal solution quickly. In the previous section, each superpixel is expressed by a feature vector that can also correspond to the process of chromosome encoding in the GA. In addition, the sample capacity and clustering numbers are small, making the algorithm converge rapidly. Therefore, the optimal seed points will be found by the GA to obtain the best clustering results in a short time. The process of choosing optimal seed points using the GA is shown in Table 3.

We first define a chromosome structure C = (c1, c2, , cK), where the ci, i = 1, 2, …, K, including initial K superpixels, and K is user-specified. For each superpixel with gene code string Xi= (ai1, ai2, , aip), where in aip, i = 1, 2, …, L, p represents the number of superpixel attributes in section 2.2.1.

For the given superpixels {X:X1,X2,XL}, we finally obtain K classes of superpixels with cluster center R = (r1, r2, , rK). The number of superpixels for cluster ri is ni, and xij are all superpixels in cluster ri. We define a fitness function φ(C) that can be obtained by the between-class variance δ2 to evaluate the goodness of a chromosome. φ(C) and δ2 can be calculated using (7) and (8).

(7)(8)

The higher the value of φ(C), the better the chromosome quality is assumed to be. The chromosome C with the maximum value of φ(C) is considered the optimal one in the population, and the K superpixel seeds are chose to generate the SNGF.

The selection process copies individual strings with high fitness function values into the next population based on the ‘‘roulette wheel” selection approach. The main purpose of crossover is to exchange genetic information of the selected chromosomes. Mutation is the process of a random alteration in the genetic structure of a chromosome, which can introduce genetic diversity into the population. In our method, the probabilities of crossover and mutation are τ and η, respectively, and the termination criteria are as follows:

1. The biggest fitness function value is obtained and the algorithm converges.
2. The fixed number of generations is reached.

#### 2.2.3 Feature extraction and lung parenchyma segmentation.

1. Feature extraction and lung identification: After clustering the superpixel samples using the optimized SGNF algorithm, four superpixel sample sets that include the left and the right lung parenchyma images, pleural tissue and extrathoracic area are obtained. We still need to identify and segment the lung parenchyma from the image sequences. Because the average greyscale value of each superpixel sample set is equal to the average value of all superpixels, the sample set with the highest value should be the pleural tissue, and the value for the lung parenchyma is close to that of the extrathoracic area. As the distribution of superpixel coordinates of the lung parenchyma is relatively concentrated, the sample set with the highest value of coordinate variance should be the extrathoracic area. And the left two sample sets should be the left and the right lung parenchyma. In this paper we will mainly extract two features of superpixel samples: the average grayscale value and the coordinate variance. An overview of the detection method is shown in Table 4.

As mentioned previously, each superpixel can be expressed by feature vector ([l, a, b, x, y, z]T), Xi= (ai1, ai2, , ai6). For a sample set {X}, the formulas to determine the average grayscale value, centroid coordinate and coordinate variance are as follows: (9)(10)(11)

The sample set with the smaller coordinate variance should be the left and the right lung parenchyma images, which are assumed to be S1 and S2. We still need to traverse all superpixel samples in S1 and S2. As the superpixels in the same image will have the same attribute value z, all superpixel samples can be connected according to the value of attribute z. By sequentially outputting all images, the sequential coarse lung parenchyma images were fully segmented.

2. Removing the trachea/bronchus and refining the lung contour: After coarse segmentation of lung CT image sequences, there are still the trachea/bronchus at the top of the lung image. To ensure the integrity of the lung parenchyma segmentation, we adopt an improved region growing method [25] to remove them. The description of improved region growing is given as follows.

1. Step 1: Binarization for the coarse lung image sequences
2. Step 2: Extract the minimum bounding rectangle of the lung.
3. Step 3: Select seed points by using LRS algorithm.
4. Step 4: Refine the lung contours with erosion and dilation.
5. Step 5: Acquire the final lung mask sequences.

We first use adaptive threshold method for image binarization and extract the minimum bounding rectangle of the lung. And then LRS algorithm is employed to select left and right seed points. In LRS, scan the minimum bounding rectangle image along the left and right sides simultaneously until there are more than 5 consecutive points on y direction with the pixel value 255, and record the middle (third) one’s ordinate value as the seeds. Next we adopt the improved region growing method based on these seeds to discard disconnected trachea, bronchus and other noise and extract lung out. Finally, dilation and erosion are used to smooth the contour and eliminate some vessels, small nodules as well as bones. Thus we acquire the final lung mask sequences with which to segment lung parenchyma image sequences accurately. The process of removing the trachea/bronchus and refining the lung contour is shown in Fig 6.

Fig 6. The process of trachea/bronchus removing and lung contour refining.

(a) Binarization of the coarse lung image; (b) Extraction of minimum bounding rectangle; (c) Select seed points with the LRS algorithm; (d) Final lung mask.

https://doi.org/10.1371/journal.pone.0160556.g006

### Results

To verify our method’s validity and universality on lung parenchyma segmentation for these four types of lung CT image sequences, we compare the results of our method with those of some existing algorithms, such as the active contour model (ACM) [26], the watershed (Watershed) [27], region growing (RG) [28] and the level set (Level Set) [29], and with manual segmentation by two experts. All our implementations were programmed in the Microsoft Visual Studio 10.0 environment and executed on a personal computer equipped with a 3.40GHz Intel Core i7-3770 processor with 8 GB RAM. The software packages we used for medical image processing and 3D visualization are ITK 4.4.2 and VTK 6.1.0.

### 3.1 Qualitative evaluation

A solitary pulmonary nodule (SPN) is one of the most common types of pulmonary nodules. In this paper, for a series of lung CT image sequences with solitary pulmonary nodules, we use our algorithm and the ACM, watershed, RG and level set algorithms for lung segmentation. We must set up some of the necessary parameters to ensure the accuracy and effectiveness of the segmentation method. The values of these parameters are shown in Table 5.

Because of the large number of image sequences, we select five lung CT images from the top to the bottom of the lung in a dataset with SPN and then use one image out of every twelve to demonstrate the process and the results of lung image segmentation.

The process and results of our method are shown in Fig 7. Column (b) is the ROI extraction result of the original lung CT image sequences column (a); column (c) is the result using the SGLIC algorithm for superpixel segmentation; column (d) is the result of using the SGNF algorithm for clustering, which is optimized by the genetic algorithm; column (e) and (f) are the coarse and the final lung parenchyma mask; and columns (g) and (h) are the final segmentation results of our method and the artificial segmentation.

Fig 7. The segmentation results of our proposed method.

Column (a) and (b) are five original lung CT images and ROI images from the top to the bottom; (c) and (d) shows the results of GSLIC and SGNF; (e) and (f) are the coarse and the final lung parenchyma mask; (g) and (h) present the final results of the proposed method and manual segmentation.

https://doi.org/10.1371/journal.pone.0160556.g007

When using the RG, watershed and active contour model algorithms for sequence image segmentation, we set up the left and right lung seed points and select five lung CT images based on the result of the experiment. In this paper, the coordinates of these seed points from top to bottom are (235, 272) and (206, 276); (212, 277) and (331, 260); (187, 249) and (341, 251); (194, 245) and (343, 265); and (195, 304) and (333, 327), which will be used to segment the left and the right lung parenchyma images. The process and results of using the watershed and RG segmentation algorithms are shown in Fig 8 and Fig 9. We have observed that when the watershed algorithm is used to segment the images, different level values will have different results (Fig 8, Column (b)—(d)). Compared with the segmentation results of the level values of 0.05, 0.1 and 0.15, the best level value of the best segmentation results is 0.15, and the coarse and final segmentation results are shown in Fig 8, Column(e) and (f). In addition, when using the RG algorithm, the threshold of the best segmentation result is between 100 and 120 (Fig 9, Column (f)).

Fig 8. The segmentation results of using the watershed method.

Column (a) is the five original lung CT images from top to bottom; (b)-(d) show the results of the level values of 0.05, 0.1 and 0.15, respectively; and (e) and (f) present the final results using the watershed algorithm and manual segmentation.

https://doi.org/10.1371/journal.pone.0160556.g008

Fig 9. The segmentation results using the RG (region growing) method.

Column (a) is the five original lung CT images from top to bottom; (b)-(d) show the mask of the left, right and whole lung; (e) are the coarse lung parenchyma mask; (f) and (g) present the final results of using RG algorithm and manual segmentation.

https://doi.org/10.1371/journal.pone.0160556.g009

In Fig 10, for the CT images with SPN, we give the comparison of the artificial segmentation results of the 5 methods. The best segmentation result is obtained by using the ACM to segment, the image spacing we select is 5 pixels, and the expansion coefficient value is 2.0 (Fig 10, Column (d)). When we use the level set algorithm for segmentation, the best segmentation result is obtained when the time threshold is 100 and the stop time is 500 (Fig 10, Column (g)). Fig 11 shows the front and back of the lung in the 3D reconstruction of the lung parenchyma image sequences segmentation results with our proposed method using VTK.

Fig 10. The comparison of the final segmentation results lung parenchyma scans with malignant SPN.

Column (a) is the five original lung CT images from top to bottom, and (b)-(g) show manual segmentation; our proposed method; and the ACM, watershed, RG and level set methods, respectively.

https://doi.org/10.1371/journal.pone.0160556.g010

For the other three types of lung sequence images, we also compare the segmentation results of these five segmentation methods without nodules (Fig 12), with benign nodules (Fig 13), and with pleural nodules (Fig 14).

Fig 12. The comparison of the final segmentation results of lung parenchyma scans without nodules.

Column (a) is the five original lung CT images from top to bottom, and (b)-(g) show manual human segmentation; our proposed method; and the ACM, watershed, RG and level set methods, respectively.

https://doi.org/10.1371/journal.pone.0160556.g012

Fig 13. The comparison of the final segmentation results of lung parenchyma scans with benign nodules.

Column (a) is the five original lung CT images from top to bottom, and (b)-(g) show manual human segmentation; our proposed method; and the ACM, watershed, RG and level set methods, respectively.

https://doi.org/10.1371/journal.pone.0160556.g013

Fig 14. The comparison of the final segmentation results of lung parenchyma scans with pleural nodules.

Column (a) is the five original lung CT images from top to bottom, and (b)-(g) show manual human segmentation; our proposed method; and the ACM, watershed, RG and level set methods, respectively.

https://doi.org/10.1371/journal.pone.0160556.g014

When the experimental results are compared with the results of manual segmentation, it is shown that the method in this paper has the best segmentation results, particularly for the segmentation of irregular lung images, such as the top and the bottom of the lung and the pleural nodules in CT image sequences, and the advantages are more obvious.

For the lung sequences images without nodules (Fig 12), the segmentation results of our method (Fig 12, Column (c))and the region growing method (Fig 12, Column (f)) have good segmentation results, but the segmentation results of the watershed (Fig 12, Column (e)) and level set algorithms (Fig 12, Column (g)) are relatively poor.

For the lung image sequences with benign nodules (Fig 13, Row 2), the ACM and watershed algorithm will lose some of the lung parenchyma (Fig 13, Column (d) and (e)); the RG and level set algorithms will lose some of the pulmonary nodules (Fig 13, Column (f) and (g)); and the method in this paper has the best segmentation results (Fig 13, Column (c)).

For the lung sequence images with malignant SPN (Fig 10, Row 2), the algorithm in this paper can ensure the integrity of the segmentation of the lung parenchyma (Fig 10, Column (c)), while ACM will lose some of the lung parenchyma (Fig 10, Column (d)) and the watershed, RG and level set algorithms (Fig 10, Column (e)—(g)) will lose some of the pulmonary nodules.

For the lung sequence images with pleural nodules (Fig 14, Row 4), the ACM, watershed, RG and level set algorithms will miss the retraction part of the pulmonary and pleural nodules (Fig 14, Column (d)—(g)), while the method in this paper can guarantee much of the retraction part of the pulmonary and pleural nodules (Fig 14, Column (c)) and is the closest to the manual segmentation results (Fig 14, Column (b)).

In addition, for all CT images from the top to the bottom of the lung, the method in this paper is the only one that can ensure the integrity of the segmentation. Therefore, our proposed method has a better segmentation result and a higher generality on segmentation of lung parenchyma images.

### 3.2 Quantitative comparisons

Quantitative evaluation has significant importance in objectively assessing the effectiveness of an algorithm. The probabilistic rand index (PRI) [30], variation of information (VoI) [31] and Jaccard similarity coefficient [32] (Kim et al., 2005) are used to objectively assess the performance of the proposed algorithm.

Assuming that the original lung image S contains M pixels, the referential and actual segmentation results are expressed as Ss and Sr, respectively, and the following conditions theoretically should be have met (12): (12) where K and N are the number of segmented regions in the referential and the actual segmentation results, respectively.

The probabilistic rand index (PRI) is a parameter to evaluate the consistency of attribute symbiosis between the actual segmentation results and the reference.

For a pixel pair (xi, xj) in the original lung image S marked (si, sj) with the same attributes in the referential segmentation result Ss, which should be the same in Sr, the value of PRI [30] can be calculated as (13): (13) where I is a discriminant function that is used to determine whether the pixel pair has the same label. The value of PRI is in the range of [0, 1], and the larger the value, the better the result.

The variation of information (VoI) [31] is a measure of information content that depicts how much one segmentation reflects the information of the other segmentation. It is the conditional entropy among the distributions of the segments labels. Therefore, the VOI value can be calculated as: (14) where H(Ss) and H(Sr) represent the entropy, and I(Ss, Sr) represents the mutual information. H(Ss) and I(Ss, Sr) also can be calculated as (15) (16) and (17).

(15)(16)(17)

The VoI values lie in [0, ∞). The 0 indicates that the two segmentations match perfectly. The smaller the value of VoI, the less information changes and the better the results will be.

The Jaccard similarity coefficient (Jaccard) is a measure to compare the similarity between the sample sets, which can indicate the coincidence degree of two images. The value of Jaccard can be calculated as (18).

(18)

The Jaccard values lie in [0, 1], and a higher Jaccard similarity coefficient indicates a better segmentation result.

Table 6 shows the average scores of the PRI, VoI and Jaccard measures for the five algorithms on four types of lung CT image sequences. It is clear from Table 6 that the proposed method outperforms the other state-of-the-art algorithms in terms of PRI, VoI and Jaccard.

For the lung sequences images of without nodules, five methods have better segmentation results. The average PRI and Jaccard values of the RG algorithm are close to the proposed method but the average VoI values are much greater. The watershed algorithm has a minimum PRI value of 0.9474, while the level set algorithm has a maximum VoI value and the lowest Jaccard value. Moreover, for lung sequence images with benign nodules, malignant SPN and pleural nodules, the RG algorithm’s performance drops rapidly with the highest VoI value, while the watershed and level set algorithms are relatively stable. Our method has better results in term of PRI, VoI and Jaccard values.

In general, our proposed method has the best segmentation performance in terms of the parameter comparison of the five types of image segmentation. From lungs without nodules to lungs with pleural nodules, our method’s performance declines slightly. The main cause perhaps is that the lung parenchyma images contain lung nodules, and the types of lung nodules are becoming more and more complicated. It is clear from Table 6 that the RG algorithm is the most sensitive to lung nodules. Unless otherwise stipulated in the image sequences without nodules, the RG algorithm is very close to the proposed method, but once lung nodules are included, there is a drastic decline in the indicators.

Moreover, the performance of each of the five algorithms in terms of PRI, VoI and Jaccard values on each image with malignant SPN (a) and without nodules (b) is graphically represented in Fig 15, Fig 16 and Fig 17. It is clear from Fig 15 that the proposed method performs better than the other methods in terms of the PRI value, the RG algorithm is close to the proposed method and better than the other algorithms. Fig 16 shows that our proposed method can always keep the global minimum value of VOI, while the level set value nearly approaches ours. Both the Watershed and RG algorithm have higher VoI values. It is obvious that our Jaccard value is higher than that of RG and is far better than for the other methods.

In consequence, based on the comparison and analysis of the three measures PRI, VoI and Jaccard, our method is the most close to the artificial segmentation results, which can further reflect our method’s high preformation and wide generality in the segmentation of lung parenchyma images.

For the four types of lung parenchyma images, we also analyzed the time performance of the five methods, as shown in Table 7, and the average processing times for the five methods are shown in Fig 18. The “Average Dataset size” row signifies the average number of lung CT images in a dataset. It is clear from Fig 18 that the average processing time for each dataset using our method is 42.21 seconds; i.e., it will take 0.71 seconds to process a single slice, which is far better than in other four methods. Therefore, the proposed method has obvious advantages over the other methods in terms of segmentation speed of lung CT images.

### Conclusion

Our work indicates that our proposed method can segment various types of lung parenchyma image sequences effectively. This method is more accurate and universally applicable than any of the traditional methods. Based on the segmentation of the four different types of sequences of lung CT images, which included 4812 images from 80 datasets, we compare the results of our method and those of the existing algorithms with manual segmentation. The experimental results show that our method can achieve accurate segmentation of the lung parenchyma and in particular accurate segment the lung CT images, which have complex morphological structures such as the top and bottom of the lung and contain pulmonary nodules. Our method can achieve an average volume pixel overlap ratio of 92.22 ± 4.02% for the four types of lung parenchyma image sequences. Moreover, our method is less time consuming, with an average processing time of 42.21 seconds for each dataset, meaning it takes approximately 0.71 seconds to process a single slice. Therefore, in the segmentation of lung parenchyma image sequences, taking the high correlation between adjacent slices of CT image sequences into consideration can significantly improve the speed of segmentation, while superpixels can guarantee the quality and the post-processing of image segmentation, and the SGNF optimized by the GA can be more effective at maintaining the integrity of the lung parenchyma segmentation.

### Acknowledgments

The work is supported by the National Natural Science Foundation of China (61373100, 61540007) and the National Key Laboratory Open Foundation of China (BUAA-VR-15KF02, BUAA-VR-16KF13).

### Author Contributions

1. Conceived and designed the experiments: XL JZ.
2. Performed the experiments: LL XL.
3. Analyzed the data: CJ QC.
4. Contributed reagents/materials/analysis tools: YQ QC.
5. Wrote the paper: XL.

### References

1. 1. Siegel R, Naishadham D, Jemal A. Cancer statistics, 2012[J]. CA: a cancer journal for clinicians, 2012, 62(1): 10–29.
2. 2. Hsieh J. Computed tomography: principles, design, artifacts, and recent advances[C]. Bellingham, WA: SPIE, 2009.
3. 3. Saghir Z, Dirksen A, Ashraf H, Bach KS, Brodersen J, Clementsen PF, et al. CT screening for lung cancer brings forward early disease. The randomised Danish Lung Cancer Screening Trial: status after five annual screening rounds with low-dose CT. Thorax. 2012;67(4):296–301. pmid:WOS:000301889800006.
4. 4. Zhao JJ, Ji GH, Qiang Y, Han XH, Pei B, Shi ZH. A New Method of Detecting Pulmonary Nodules with PET/CT Based on an Improved Watershed Algorithm. Plos One. 2015;10(4). ARTN e012369410.1371/journal.pone.0123694. pmid:WOS:000352478400114.
5. 5. Geng H, Bian Z, Yang J, Tan W. Fully automatic extraction of lung parenchyma from CT scans[C]//Intelligent Control and Automation (WCICA), 2014 11th World Congress on. IEEE, 2014: 5626–5630.
6. 6. Liming D, Caiming Z, Shanshan G. A new algorithm of automatic lung parenchyma segmentation based on CT images[C]//2011 International Conference on Consumer Electronics, Communications and Networks (CECNet). 2011: 976–979.
7. 7. Mansoor A, Bagci U, Mollura D J. Near-optimal keypoint sampling for fast pathological lung segmentation[C]//Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE. IEEE, 2014: 6032–6035.
8. 8. Shojaii R, Alirezaie J, Babyn P. Automatic segmentation of abnormal lung parenchyma utilizing wavelet transform[C]//Acoustics, Speech and Signal Processing, 2007. ICASSP 2007. IEEE International Conference on. IEEE, 2007, 1: I-1217-I-1220.
9. 9. Yan-hua R, Xi-wen S, Sheng-dong N. A 3D segmentation method of lung parenchyma based on CT image sequences[C]//Information Networking and Automation (ICINA), 2010 International Conference on. IEEE, 2010, 2: V2-332-V2-336.
10. 10. Luo X, Tian J, Lin Y. An algorithm for segmentation of medical image series based on active contour model[J]. Journal of Software, 2002, 13(6): 1050–1059.
11. 11. Yu N, Wu J, Weinstein S P, Gaonkar B, Keller B M, Ashraf A B, et al. A superpixel-based framework for automatic tumor segmentation on breast DCE-MRI[C]// SPIE Medical Imaging. International Society for Optics and Photonics, 2015.
12. 12. Song Y, Ni D, Zeng Z, He L, Chen S, Lei B, et al. Automatic Vaginal Bacteria Segmentation and Classification Based on Superpixel and Deep Learning[J]. Journal of Medical Imaging & Health Informatics, 2014, 4(5):781–786(6).
13. 13. Ren X, Malik J. Learning a classification model for segmentation[C]//Computer Vision, 2003. Proceedings. Ninth IEEE International Conference on. IEEE, 2003: 10–17.
14. 14. Lucchi A, Smith K, Achanta R, Lepetit V, Fua P. A Fully Automated Approach to Segmentation of Irregularly Shaped Cellular Structures in EM Images. Medical Image Computing And Computer-Assisted Intervention—Miccai 2010, Pt Ii,. 2010;6362:463–71. pmid:WOS:000287828300057.
15. 15. Hammoudi AA, Li FH, Gao L, Wang ZY, Thrall MJ, Massoud Y, et al. Automated Nuclear Segmentation of Coherent Anti-Stokes Raman Scattering Microscopy Images by Coupling Superpixel Context Information with Artificial Neural Networks. Lect Notes Comput Sc. 2011;7009:317–25. pmid:WOS:000306302100039.
16. 16. Wang J, Wang X. VCells: Simple and efficient superpixels using edge-weighted centroidal Voronoi tessellations[J]. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2012, 34(6): 1241–1247.
17. 17. Wen W X, Liu H, Jennings A. Self-generating neural networks[C]//Neural Networks, 1992. IJCNN., International Joint Conference on. IEEE, 1992, 4: 850–855.
18. 18. Inoue H, Narihisa H. Efficiency of self-generating neural networks applied to pattern recognition[J]. Mathematical and computer modelling, 2003, 38(11): 1225–1232.
19. 19. Inoue H, Narihisa H. Efficient pruning method for ensemble self-generating neural networks[J]. Journal of Systemic, Cybernetics and Informatics, 2003, 1(6): 72–77.
20. 20. Feng S, Tan A H. Self-organizing neural networks for behavior modeling in games[C]//Neural Networks (IJCNN), The 2010 International Joint Conference on. IEEE, 2010: 1–8.
21. 21. Zhao JJ, Ji GH, Xia Y, Zhang XL. Cavitary nodule segmentation in computed tomography images based on self–generating neural networks and particle swarm optimisation[J]. International Journal of Bio-Inspired Computation, 2015, 7(1): 62–67.
22. 22. Law T Y, Heng P A. Automated extraction of bronchus from 3D CT images of lung based on genetic algorithm and 3D region growing[C]//Medical Imaging 2000. International Society for Optics and Photonics, 2000: 906–916.
23. 23. Dehmeshki J, Ye X, Lin XY, Valdivieso M, Amin H. Automated detection of lung nodules in CT images using shape-based genetic algorithm[J]. Computerized Medical Imaging and Graphics, 2007, 31(6): 408–417. pmid:17524617
24. 24. Khan A, Jaffar M A. Genetic algorithm and self organizing map based fuzzy hybrid intelligent method for color image segmentation[J]. Applied Soft Computing, 2015, 32: 300–310.
25. 25. Zhao JJ, Ji GH, Han X, Qiang Y, Liao XL. An automated pulmonary parenchyma segmentation method based on an improved region growing algorithm in PET-CT imaging[J]. Frontiers of Computer Science, 2016, 10(1):189–200.
26. 26. Bellotti R, De Carlo F, Gargano G, Tangaro S, Cascio D, Catanzariti E, et al. A CAD system for nodule detection in low-dose lung CTs based on region growing and a new active contour model[J]. Medical Physics, 2007, 34(12): 4901–4910. pmid:18196815
27. 27. Masoumi H, Behrad A, Pourmina MA, Roosta A. Automatic liver segmentation in MRI images using an iterative watershed algorithm and artificial neural network[J]. Biomedical Signal Processing and Control, 2012, 7(5): 429–437.
28. 28. Dehmeshki J, Amin H, Valdivieso M, Ye X. Segmentation of pulmonary nodules in thoracic CT scans: a region growing approach[J]. Medical Imaging, IEEE Transactions on, 2008, 27(4): 467–480.
29. 29. Li C, Huang R, Ding Z, Gatenby JC, Metaxas DN, Gore JC. A level set method for image segmentation in the presence of intensity inhomogeneities with application to MRI[J]. Image Processing, IEEE Transactions on, 2011, 20(7): 2007–2016.
30. 30. Unnikrishnan R, Hebert M. Measures of similarity[C]//Application of Computer Vision, 2005. WACV/MOTIONS'05 Volume 1. Seventh IEEE Workshops on. IEEE, 2005, 1: 394–394.
31. 31. Meila M. Comparing clusterings by the variation of information. Lect Notes Artif Int. 2003; 2777:173–87. pmid:WOS:000185937100013.
32. 32. Prastawa M, Bullitt E, Ho S, Gerig G. A brain tumor segmentation framework based on outlier detection[J]. Medical image analysis, 2004, 8(3): 275–283. pmid:15450222

• Pulmonary imaging
• Imaging techniques
• Computed axial tomography
• Algorithms
• Neurons
• Lung and intrathoracic tumors
• Pleurae
• Genetic algorithms
Sours: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0160556
Anatomy and physiology of the respiratory system

## Lung Parenchymal Mechanics

Béla Suki,1,*Dimitrije Stamenovic,1 and Rolf Hubmayr2

### Béla Suki

1Department of Biomedical Engineering, Boston University, Boston, Massachusetts

Find articles by Béla Suki

### Dimitrije Stamenovic

1Department of Biomedical Engineering, Boston University, Boston, Massachusetts

Find articles by Dimitrije Stamenovic

### Rolf Hubmayr

2Department of Internal Medicine, Mayo Clinic College of Medicine, Rochester, Minnesota

Find articles by Rolf Hubmayr

See other articles in PMC that cite the published article.

### Abstract

The lung parenchyma comprises a large number of thin-walled alveoli, forming an enormous surface area, which serves to maintain proper gas exchange. The alveoli are held open by the transpulmonary pressure, or prestress, which is balanced by tissues forces and alveolar surface film forces. Gas exchange efficiency is thus inextricably linked to three fundamental features of the lung: parenchymal architecture, prestress, and the mechanical properties of the parenchyma. The prestress is a key determinant of lung deformability that influences many phenomena including local ventilation, regional blood flow, tissue stiffness, smooth muscle contractility, and alveolar stability. The main pathway for stress transmission is through the extracellular matrix. Thus, the mechanical properties of the matrix play a key role both in lung function and biology. These mechanical properties in turn are determined by the constituents of the tissue, including elastin, collagen, and proteoglycans. In addition, the macroscopic mechanical properties are also influenced by the surface tension and, to some extent, the contractile state of the adherent cells. This article focuses on the biomechanical properties of the main constituents of the parenchyma in the presence of prestress and how these properties define normal function or change in disease. An integrated view of lung mechanics is presented and the utility of parenchymal mechanics at the bedside as well as its possible future role in lung physiology and medicine are discussed.

### Introduction

The lung is an organ with complex internal structure that evolved to serve the gas exchange needs of the organism. For efficient gas exchange, the internal surface area should be maximized, while the distance traveled by O2 and CO2 between alveolar air and capillary blood should be minimized (274, 275). Furthermore, the functional needs of an organism also require a significant reserve capacity of the lung and hence a large surface area for gas exchange. These constraints, together with the shape of the thorax, place severe limitations on the internal structure of the lung. First, the gas exchanging regions need to be connected with the airway opening that defines the airway structure. Second, for efficient overall gas exchange, different regions should be supplied uniformly with fresh air (135). The three-dimensional structure that can satisfy such constraints is a fractal branching tree (136). The terminal branches of the tree supply air to the acinus that is composed of thousands of alveoli where the actual gas exchange occurs via diffusion (215). The thickness of the septal walls is only 4 to 5 μm and the diameter of the alveoli in the human lung at total lung capacity (TLC) is approximately 200 μm. The parenchymal structure is thus a huge collection of tiny and fine balloons that pack an enormous surface area (close to that of a tennis court) into the chest cavity (275).

Gas exchange in the lung is maintained via the rhythmic process of inspiration and expiration. The acini are connected to the airway opening and consequently, the pressure in the alveoli is near atmospheric most of the time. Since the thin-walled compliant alveoli easily collapse, they must be held open by a positive transmural pressure. This pressure, called transpulmonary pressure (Ptp), is generated by a negative pressure around the lung in the thoracic cavity. Because of the mechanical connectedness of the lung parenchyma, the distending action of Ptp produces a tensile stress, or prestress, throughout the parenchyma. The extent to which the alveoli become distended by the cyclic variation of prestress during breathing depends upon the mechanical properties of the parenchyma. Thus, the efficiency of gas exchange is inextricably linked to three fundamental features of the lung: the structural organization of the parenchyma, the mechanical properties of its components, and the prestress.

There are several important consequences of the fact that the parenchyma is prestressed. First, the mechanical forces generated by the prestress will ultimately be transmitted to pulmonary cells that adhere to the parenchymal tissue, affecting thereby many essential cell functions and hence general lung biology (76, 264). Second, the prestress also contributes to lung deformability itself, which influences a host of other macroscopic and microscopic processes such as local ventilation, regional blood flow, smooth muscle contractility, surface tension, fluid balance, and alveolar stability (234). The main pathway of stress transmission is, however, through the extra-cellular matrix (ECM). Thus, the mechanical properties of the ECM also play a key role both in lung function and biology. These mechanical properties in turn are determined by the constituents of the tissue including elastin, collagen and the “ground substance” composed mostly of proteoglycans (252). In addition to the prestress and the ECM, the macroscopic mechanical properties are also influenced by adherent cells, the surface tension of the air-liquid interface and the organization of the parenchyma. The field of biomechanics attempts to uncover how the elementary properties of biological constituents and their organization determine the specific microscopic and macroscopic properties of an organ or tissue (80).

This article focuses on the biomechanical and structural properties of the lung parenchyma in the presence of prestress and how these properties define function in the normal and diseased lung. First, we present a brief review of the biomechanics of soft tissues. Next, we summarize the constituents of the connective tissue of the lung and their structural organization. We will then describe how the mechanical properties of the parenchyma at the macroscale arise from the properties of its constituents that include fibers, ground substance, surface film and cells, as well as their structural organization. In places where there is a lack of experimental data on specific ECM properties of the lung, we will use results from other tissues or organs. The importance of prestress is highlighted throughout. Next, we examine how the most important load-bearing elements of the parenchyma transmit the prestress down to the level of cells since these stresses are critical in determining the homeostasis and cellular responses to lung injury. We also discuss lung stability since it is an important determinant of normal gas exchange. Finally, we provide an integrated view on lung mechanics and speculate on the utility of parenchymal mechanics at the bedside as well as on its possible future role in lung physiology and pulmonary medicine.

### Biomechanics of Soft Tissues

Biomechanics can be defined as the application of the principles of mechanics to biology and physiology. According to Fung (80), “Biomechanics aims to explain the mechanics of life and living. From molecules to organisms, everything must obey the laws of mechanics.” Beyond its power to explain the mechanical behavior of the living, biomechanics is now recognized to be part of mechanobiology (192, 264) in that the mechanical properties of biological tissues, also called biomechanical properties, play fundamental roles in the normal functioning of virtually all connective tissues, organs, and organisms. Indeed, these biomechanical properties are critical determinants of how mechanical interactions of the body with the environment produce physical forces at the cellular level. Mechanobiology is particularly relevant for the lung since it is an open system cyclically stretched by external forces generated by the respiratory muscles (76, 252, 264).

In the lung, mechanical forces can directly influence physiological function via cellular signaling (283) such as during lung development (260), surfactant release by alveolar epithelial cells (282), contraction of airway smooth muscle cells (74) and tissue remodeling (144). It is now well recognized that mechanical interactions between cells and the ECM have major regulatory effects on cellular physiology and cell-cycle kinetics, which can lead to the reorganization and remodeling of the ECM (30, 43). This in turn influences the macroscopic biomechanical properties and hence the functioning of the lung.

### Linear Elastic Behavior

Traditional biomechanics (80) has focused on characterizing the macroscopic structural and mechanical properties of living tissues and organs by establishing mathematical relations, called the constitutive equations, that describe how mechanical stresses (force per unit area) change in response to a change in the size and/or shape of a body usually given in terms of strain (relative change in dimension). The simplest constitutive equation is Hooke's law that relates a small uniaxial length change to the corresponding stress (σ) in the material via the following equation:

where Y is the Young's modulus of elasticity. Here the strain ε is defined as the relative change in length, ε = (ll0)/l0 with l and l0 being the deformed and resting length, respectively, and the stress is equal to the stretching force F applied to the material divided by the initial cross-section A0, σ = F/A0. Equation 1 describes a static linear relationship between stress and strain in simple uniaxial elongation that has often been used in parenchymal tissue strip experiments (3, 21, 31, 65, 67, 69, 73, 122, 129, 154, 172, 177, 182, 208, 213, 256, 286). Note that Eq. 1 presumes that prior to the application of F, the material is stress-free. However, many biological tissues, including lungs, are not stress-free but rather are prestressed and their elastic properties depend on the level of the prestress, as discussed in the next section.

The mechanical properties of an ideal isotropic linearly elastic material, are completely characterized by two elastic moduli. One of them can be Y and the other can be one of the following: the shear modulus, the Poisson's ratio or the bulk modulus. The shear modulus (μ) describes the material's ability to resist shape distortion without volume changes, and is defined as the ratio of the shear stress (shear force divided by the area it applies to) and the corresponding shear strain. During uniaxial stretch of a tissue strip, the lateral dimensions of the strip decrease. This decrease can be characterized by the Poisson's ratio that is defined as the negative ratio of the strains perpendicular and parallel to the elongation. Finally, the bulk modulus (κ) measures the resistance of the material to uniform volume change and is given by the change in pressure divided by the relative change in volume of the sample.

The various moduli are a function of the constituents and their organization in the material tested. For parenchymal tissue strips measured in tissue bath, Y depends on the elastin, collagen, and proteoglycan content of the strip (3, 31, 45, 64, 67, 69, 177, 255, 287), alveolar diameters (31), anatomic makeup (214), tonicity of the bath (45, 172), to some extent cellular contractile state (65, 73, 182, 286), and even genetic make-up (8). As we shall see below, small mechanical strains can be superimposed on a prestressed state; hence the moduli can also depend on the prestress. These dependencies will be discussed later in the article.

### Nonlinear elastic behavior

While the simple linear constitutive equation of Eq. 1 can be useful in various applications particularly when the modulus can be related to composition and microstructure, in reality constitutive equations of biological materials are invariably nonlinear within the physiological ranges of stress and strain (80). Figure 1A shows the stress-strain curves of lung parenchymal tissue strips from a normal rat and one that had been treated with elastase that causes functional changes in the lung akin to pulmonary emphysema (138). It can be seen that with increasing strain, the relationship deviates from linearity, especially in the normal lung; the slope dσ/dε increases with increasing ε, which is called strain-hardening. Clearly, Eq. 1 is not applicable to describe such a behavior. However, for small variations around a given strain, Eq. 1 can still be a useful descriptor and Y becomes the incremental modulus of the tissue that is a function of the operating point around which the slope of the curve is evaluated. To illustrate this, consider a general nonlinear elastic constitutive equation:

where f is a nonlinear continuous function of its argument that often takes an exponential form (242). It is useful to approximate f with a series expansion around a given operating strain, ε0:

σf(ε0) + f(ε0)(ε − ε0) + (1 ∕ )f(ε0)(εε0)2

(3)

Open in a separate window

Figure 1

(A) Stress-strain curves of parenchymal tissue strips from a normal rat and a rat that had been treated with elastase-mimicking pulmonary emphysema. (B) Pressure-volume curves measured by injecting 2 ml of air starting from functional residual capacity in a normal and an elastase-treated rat. Adapted from Ref. (138) with permission.

Here f′ (ε0) and f″(ε0) are the first and second derivatives of f, respectively, evaluated at ε0. The first term on the right-hand side of Eq. 3, σ0 = f0), represents the prestress. It is thus evident that for small deviations of ε from ε0, the higher order terms become negligible and the stress around the operating point, σ* = σ − σ0, becomes a linear function of the strain ε* = ε − ε0. Thus, we obtain σ* = f′ (ε0)ε* which is analogous to the linear stress-strain relation in Eq. 1. The main difference is that f′ (ε0) now represents the incremental modulus of the material that depends on the operating strain ε0. By inverting the relation σ0 = f0), that is, ε = f−10), the incremental modulus can also be obtained in terms of the prestress σ0.

For isotropic inflation of the lung, a nonlinear constitutive equation similar to Eq. 3 is obtained by replacing σ and ε with Ptp and the fractional change in lung volume, respectively. In isolated lungs, Ptp = 0 cmH2O corresponds to zero gas volume in the lung, whereas in situ, it is more convenient to choose residual volume (RV) as the reference lung volume. The operating point in lung mechanics is often chosen to be a specific value of Ptp that can be experimentally set by the positive end-expiratory pressure (PEEP) that is superimposed on the functional residual capacity (FRC). Figure 1B compares the pressure-volume (P-V) curves of a normal and an emphysematous lung with inflation starting from FRC as the operating point (138). In this case, PEEP 0 cmH2O and both P and V are taken to be zero at FRC. Consequently, the first term on the right-hand side of Eq. 3 now corresponds to Ptp at FRC and f′ in the second term is the incremental bulk modulus that can be related to the more familiar static lung elastance (EL,S) as the bulk modulus normalized by absolute lung volume. The value of EL,S can be determined simply as the change in Ptp divided by the volume slowly injected into the lung. It can be seen from Figure 1 that emphysema reduces the ability of the lung parenchyma to resist stretching both at the tissue strip level during uniaxial strain (Fig. 1A) as well as in the whole lung during isotropic three-dimensional expansion (Fig. 1B).

### Viscoelasticity

When the constitutive equation includes at least one term that depends on the rate of change of deformation or the rate of change of stress, the tissue is referred to as viscoelastic. All living tissues display viscoelastic behavior (80), which is characterized by time- and frequency-dependent behavior of the material responses. Unlike elastic materials, which attain equilibrium instantaneously following the application of external loading, in viscoelastic materials this process is delayed and impeded by internal viscous stresses. Consequently, viscoelastic materials exhibit creep (i.e., continuous deformation in response to a constant stress) and stress adaptation (i.e., continuous change in stress in response to a constant strain). During cyclic loading, a phase lag develops between stress and strain due to the impeding effects of viscous forces that in turn leads to hysteresis between the loading and unloading limbs of the stress-strain curve.

One approach to mathematically deal with systems that exhibit hysteretic behavior is to define a separate elastic-like stress-strain relation for the loading and unloading limbs of the hysteresis loop (80). A more general description of nonlinear viscoelastic materials is given by a functional series expansion of the output (stress or pressure) in terms of the time history of the input (strain or volume) (246). In the frequency domain, the Fourier transform of the output P is related to the Fourier transform of the input V via a series expansion involving convolutions (246):

P(ω) = E1(ω)V(ω) + ∫E2(ωω − ω)V(ω)V(ω − ω)dω

(4)

Here ω is the circular frequency, E1 is the first order kernel and E2, … are the higher order kernels incorporating nonlinear viscoelastic effects. In general, these kernels are complex numbers and depend on ω. The expansion in Eq. 4 is not valid if the system displays static hysteresis or discontinuities. For simplicity, we will assume that the kernels are smooth functions of ω, which allows us to linearize Eq. 4 around an operating point. If P denotes the pressure measured above a given PEEP and V is the corresponding change in lung volume, then E1 is the complex modulus of the lung. The linear impedance Z of the lung is defined as the pressure divided by the flow and is given by Z = E1/(jω) where j √−1 is the imaginary unit indicative of the out-of-phase behavior between pressure and flow.

Lung tissue was recognized to be viscoelastic as early as 1939 by Bayliss and Robertson (22) and later by Mount (179) in 1955. Subsequently, Hildebrandt and co-workers demonstrated in a series of studies (102–104,108,109, 244) that the frictional component of stress in the lung tissue depends on the amount but not the rate of expansion, a finding that appears to contradict the notion that frictional losses are caused by viscous dissipation. Importantly, the relationship between the frictional and elastic stresses in the lung tissue turns out to be nearly invariant; the frictional stress is invariably between 10% and 20% of the elastic stress. The ratio of viscous and elastic stresses is referred to as the structural damping coefficient, or “hysteresivity” (77). This fixed relationship holds at the level of the whole lung (100), isolated lung parenchymal tissue strips (73), isolated smooth muscle strips (75), and even isolated living cells (68). The coupling between frictional and elastic stresses is known in the structural mechanics literature as the structural damping law (55), whereas in the physiological literature it is often referred to as the constant phase model (100), which is described below. The structural damping law is an empirical relation which implies that frictional energy loss and elastic energy storage are tightly coupled (77). In the context of linear viscoelasticity theory, it simply means that viscoelastic responses are characterized by a broad spectrum of time constants (80, 248).

If we limit the analysis to the lung tissues, the bulk mechanical properties of the parenchyma can be well described by the so-called constant phase model of tissue impedance (Zti) proposed by Hantos et al. (100,101):

The parameters G and H are the coefficients of tissue damping and elastance, respectively, and being incremental moduli, they depend on the operating point defined either by the PEEP or lung volume. The parameter α is a dependent quantity, given by 2/π) tan−1(H/G). It is noteworthy that this model also provides a good description of tissue behavior at the tissue strip level during cyclic uniaxial deformation (286). Both in tissue strips and whole lungs, the first term in Eq. 5, Rti = Gα , is the tissue resistance and the second written as Eti = Hω1 − α is the dynamic lung tissue elastance. The power-law dependence of Rti and Eti on ω is consistent with a power-law type of stress relaxation of the tissue that is detailed in the article “Complexity and emergence in Comprehensive Physiology.” Furthermore, as indicated above, the constant phase model is closely related to the structural damping law; the tissue hysteresivity can be obtained from Eq. 5 simply as η = G/H which shows that the phase angle of Zti is independent of frequency and consequently, the dissipative portion of Zti is a constant fraction of its elastic portion. Experimental studies confirm these model predictions and show that viscosity changes very little with frequency and tidal volume in whole lungs or tissue strips indicating that damping and elastic stresses are tightly coupled (17, 77, 177, 182, 287). However, several more recent studies indicated that at higher PEEP levels viscosity decreases possibly due to the increased contribution of collagen (31, 229) and that subtle changes in the coupling between viscous and elastic stresses reflect biological remodeling of the ECM of the parenchyma (3, 31, 64, 67, 69, 128, 197).

Numerous studies have evaluated the PEEP dependence of Eti or H and G in a variety of conditions including emphysema (31, 124, 126, 127), fibrosis (64, 181), and adult respiratory distress syndrome (ARDS) (5, 132, 141, 258). Such linear analyses around an operating point can be useful since the way Eti and H depend on PEEP carries information about the underlying pathology. For example, in anesthetized animals, lung volume decreases significantly in the supine position. Consequently, alveoli become unstable and collapse, a phenomenon that will be discussed later. Therefore, in healthy anesthetized animals H decreases when PEEP is increased from 0 to about 6 cmH2O due to alveolar recruitment in several species (94) including mice (99, 127, 128). Beyond a PEEP of about 10 cmH2O, H increases with PEEP due to the dominating contribution of stiff collagen. However, in various mouse models of emphysema such as the tight skin mouse (126), the rate of increase of H with PEEP is much stronger than in normal mice. Indeed, despite the much lower H in the tight skin than in the normal mice at low PEEP, H is similar in magnitude at a PEEP of 9 cmH2O in the two groups (Fig. 2), implying abnormal ECM organization and collagen function in the tight skin mouse (126).

Although the incremental analysis discussed above is useful, the full nonlinear dynamic stress-strain curve or P-V relation carries more information about the structure and composition of the parenchyma than the incremental modulus. Indeed, the incremental analysis is limited to mapping the values of the moduli as a function of the operating point and neglects all higher order components. While powerful, such nonlinear analysis is complicated and only a few studies have applied it to the lung (155, 157, 246, 250, 288). Nevertheless, the natural way the lung functions is similar to the incremental analysis since breathing involves relatively small amplitude tidal oscillations superimposed on an operating point, Ptp at FRC. Furthermore, some information about the nonlinearity can be extracted from forced oscillatory data when breathing amplitudes are employed (289). In this case, nonlinear distortion of the waveforms can be analyzed which can provide information on collagen function (126, 128).

The precise link between the viscoelastic behavior of the lung tissue and its microstructure is still poorly understood, although various theories have been proposed. For example, stress relaxation may occur through cascades of microruptures within the tissue (18) or as a result of slow undulation of fibers (248) or fiber-fiber kinetic interaction (173). In any case, the constitutive equations are commonly determined from measured dynamic stress-strain or P-V curves. These relations generally reflect behavior that emerges from the mechanical properties of the individual constituents as well as their structural arrangement in the tissue. This article focuses mostly on the elastic behavior of the lung tissue. Viscoelasticity will be treated only briefly and mostly for the purposes of identifying the roles or the changes in the roles of certain components of the tissue following a given intervention or disease condition. Some possible mechanistic insight into the viscoelastic behavior of the ECM of the normal lung tissue is presented in the article on “Complexity and Emergence,” whereas the rheological behavior of cells is treated in the “Material Properties of the Cytoskeleton” article in Comprehensive Physiology.

### Structure-function relations

One of the primary goals of biomechanics is to develop quantitative relationships between the biochemical composition and microstructure of the tissue and its functional properties, usually characterized by the constitutive equation. For example, functional properties can be given by the elastic modulus such as Y at the tissue strip level or lung elastance at the organ level. The dependence of the functionality on the composition and microstructure is called the structure-function relation. Quantitative understanding of structure-function relations in the normal lung can help identify structural defects from noninvasive functional measurements at the organ level in diseases. To develop such structure-function relations, it is necessary to determine the bulk composition of the tissue and to gain understanding of the structure and interaction of the components.

Generally, connective tissues are composed of cells and ECM that includes water and a variety of biological macromolecules. The macromolecules that are most important in determining the mechanical properties of these tissues are collagen, elastin, and proteoglycans. Among these macromolecules, the most abundant and perhaps the most critical for structural integrity is collagen. One might expect therefore that the amount of collagen in a tissue is the primary determinant of its mechanical properties. However, different connective tissues with similar collagen content can exhibit different viscoelastic behavior (80). During the last decade, the advent of novel imaging techniques (54) and quantitative computational modeling (205) have allowed the study of micromechanics of specific components of tissues and hence improved our understanding of the relationships between tissue composition, microstructure, and macrophysiology. In particular, it has become evident that macrophysiology reflects both the mechanical properties of the individual components of the tissues, as well as the complexity of its structure (19, 252). Consequently, an understanding of lung parenchymal mechanics and function requires the integration of the physical properties of the constituent molecules with their organization into fibrils, fibers, ground substance, and cells. The next section provides an overview of the most important components of the ECM of the lung and other determinants of lung mechanical function including interstitial cells and surfactant.

### Main Constituents of the Parenchyma

The lung parenchyma consists of a large collection of near spherical gas exchanging units, the alveoli. The internal surface of the alveoli is lined by a layer of cells, the epithelium, which is covered by a thin liquid film. There is a surface tension at the air-liquid interface that contributes to lung elastic recoil. The alveolar septal walls are composed of interstitial cells and the ECM. Cells can modulate the local tension on the ECM. Alternatively, mechanical forces of breathing are transmitted to the cells via the ECM fibrils and fibers such as collagen and elastin that are embedded in a soft gel called the proteoglycan matrix. Figure 3 depicts the general organization of the parenchyma from the scale of tens of alveoli surrounding an alveolar duct to cell-ECM interactions within the septal wall. Below, we describe in detail the structure and basic mechanical properties of the constituents of the alveolar wall tissue.

Open in a separate window

Figure 3

Structure and complexity of the parenchyma at three length scales. The top panel shows a terminal bronchiole (TB) leading to an alveolar duct (AD). The bottom left is a zoom into a single air-filled alveolus (A) with type I (E1) and type II (E2) alveolar epithelial cells covered by a thin liquid layer. The dots represent surfactant (S) molecules at the air-liquid interface. Secretion of lamellar bodies (LB) by the E2 cell is also shown. The right panel is a schematic representation of the extracellular matrix of the alveolar septal wall with various components including amorphous elastin (El), wavy collagen (C), complex proteoglycans (PG), basement membrane (BM) and fibroblast cells (F). (Drawing by E. Bartolák-Suki).

### Extracellular matrix

#### The collagen system

There are nearly 30 different types of collagen molecule. Most interstitial collagens (I, II, III, V, and XI) are helical that provides them with a basic structure supporting role. The helices consist of three polypeptide chains each of which is a left-handed coil of approximately 1000 amino acids, with the three chains forming a right-handed super helix (34). These helical molecules are rod-like rigid structures with length and diameter of about 300 nm and 1.5 nm, respectively, and capable of spontaneous fibril forming (226). The helical subunits are first assembled in the endoplasmic reticulum of the cell in precursor forms called procollagens that have amino and carboxyl terminal globular regions known as propeptides. These serve to solubilize the procollagen, and correctly align the individual peptide chains to facilitate helix formation (226). Following secretion, the propeptides are enzymatically cleaved, which allows the collagen molecules to associate both axially and laterally and start forming fibrils. Apparently, type I collagen is thermally unstable at body temperature and folding of the least stable microdomains can trigger self-assembly of fibrils where the helices are protected from complete unfolding (146). The fibril structure itself also shows tremendous hierarchical complexity. For example, the lateral packing of molecules can exhibit significant fluid-like disorder (106, 117). The collagen fibrils can further organize into thicker fibers through cross-linking of lysine and hydroxylysine residues present within the overlapping terminal helical and teleopeptide regions of the molecules (226). Figure 4A summarizes the hierarchical complexity of the collagen system. These fibrils and fibers may be arranged either in a randomly oriented manner (e.g., in lung tissue or cartilage) or as quasi-deterministic networks (e.g., in tendon) within an organ. In the lung, the collagen fibers are wavy at low inflation and become straight at higher lung volumes (Fig. 4B).

Open in a separate window

Figure 4

(A) Structure of collagen. Top left: single alpha helix; bottom left: collagen molecule comprising a triple helix; top right: cross-linked collagen; bottom right: schematic view of 5 molecules; with permission from Ref. (117). (B) Collagen network in the rat lung is wavy at low transpulmonary pressure (left) and at a medium inflation level (right). AE denotes alveolar entrance. Scale bar is 10 μm. Adapted from Ref. (261) with permission.

The interstitium of the lung parenchyma contains mostly types I and III collagen that provide the structural framework for the alveolar wall. Fiber thickness ranges from several hundred nanometers to well over a micron (232). The distribution of fiber thickness is skewed, and has a long “tail” (232) similar to a power law (245), indicating broad variability of fiber structure. This blend of deterministic order (exact amino sequence and axial packing) and random disorder (from fluid-like lateral packing to random networks) may partly be responsible for the existence of a broad range of time constants that characterizes the viscoelastic properties of the connective tissue of the lung (21, 248). These collagen fibers in the parenchyma are further organized to form an axial fiber network extending down from the central airways to the alveolar ducts, a peripheral fiber network extending centrally from the visceral pleura, and a parenchymal interstitium that connects the two (273). Variations in the collagen content of the parenchyma during development (255, 256), in fibrosis (64) or following in vitro digestion (287) have suggested an important role for these protein fibers in the biomechanical properties of the parenchyma. In addition to the fibrillar types I and III collagen, type IV collagen is more sheet like and is part of the basement membrane to which epithelial cells adhere. The role of type IV collagen in the mechanical properties of the ECM of the lung is currently unknown.

The Young's modulus of the type I collagen molecule has been estimated to be between 3 and 9GPa (217, 226). The elasticity of a single collagen molecule has been attributed to the existence of amino acid sequences along the triple helix that lack proline and hydroxyproline (227). These regions are more flexible than other regions of the helix. Such variation of rigid and flexible regions likely has a significant effect on the fibril-forming ability and hence the elasticity of the fibrils. Additionally, the unfolding of thermally activated molecular kinks or “crimps” along the molecule may also contribute to elasticity (175). The stress-strain curve of fibrils appears reasonably linear up to 3% to 5% strain with a modulus in the order of 0.5 to 5 MPa (226).

#### The elastic fiber system

Elastin is another essential load bearing component of the ECM. Elastin is known for its resilience over a large range of strains, and hence its ability to provide elasticity to tissues. Consequently, tissues and organs that need elasticity because they undergo cyclic stretching throughout the life time of an organism can generally be expected to contain a significant amount of elastin. Elastin is synthesized in a soluble precursor form called tropoelastin with a molecular weight of 72 kDa by smooth muscle cells, endothelial cells, and fibrob-lasts (40, 167, 185). In contrast to collagen, which is rich in hydrophylic amino-acid residues, the amino acid composition of elastin is rich in hydrophobic residues including glycine and proline (211). The hydrophobic residues together with the dense interchain crosslinking make elastin highly stable and insoluble. Tropoelastin is capable of self-assembly under physiological conditions (24) to form insoluble fibrils and fibers with a half-life of 70 years (201). Because heparan sulfate, a glycosaminoglycan (GAG) and an essential component of the proteoglycans, interacts with tropoelastin, it also plays a role in elastic fiber formation (88).

Elastic fibers are composed of elastin and microfibrils. The three most important groups of microfibrils that are closely associated with elastin include fibrillins, fibulins, and the microfibril-associated glyoproteins (56, 206, 211). The microfibrils also play a role in elastogenesis by regulating the deposition of tropoelastin onto the developing elastic fiber (176). The elasticity of microfibrils is controversial and their role in lung elasticity has not been studied. Values of the Young's modulus of microfibrils have been reported to be as low as 0.2 MPa (259), which is about 3 to 5 times lower than the stiffness of elastin (80, 223), and as high as 96 MPa (224), which is closer to that of collagen (80). The microfibrils often form a fibrous outer mantle surrounding the more amorphous elastin. While the 3D molecular structure of elastin fibers is not as well understood (117, 139), elastin organizes itself into easily extensible fibers and has a linear stress-strain relation up to 200% strain (80). The elastin molecules organize themselves into cross-linked fibers. The distributions of the diameters and lengths of elastin fibers in the lung are skewed with long “tails,” and appear similar to the distribution of collagen fiber properties (232). Thus, the elastic fibers exhibit significant structural heterogeneity and are also known to be mechanically connected to the collagen (35) via microfibrils and/or proteoglycans (117, 133, 204).

It is notable that the stiffness of elastin is at least 2 orders of magnitude smaller than that of collagen (80). This is likely a result of the amorphous nature of elastin compared to the more regular organization of collagen fibers. The elastic resistance of elastin is thought to be of entropic origin. Figure 5A shows the stretch-induced conformational changes in elastic fibers. However, this is not simply the entropic elasticity of a random chain. The tropoelastin has two major types of alternating domains, the hydrophilic helical domains and the hydrophobic lysine-rich domains. These nonrandom, regularly repeating structures exhibit dominantly entropic elasticity by means of a damping of internal chain dynamics on extension (266). In the lung, elastin forms a complete network of fibers (Fig. 5B). Traditionally, elastin is thought to dominate parenchymal elasticity at normal breathing lung volumes (221), while collagen becomes progressively more important as volume approaches TLC. However, comparing the effects of elastin and collagen digestion on the constitutive equation of parenchymal strips suggests that collagen and elastin may be equally important even at lower lung volumes (287).

Open in a separate window

Figure 5

(A) Longitudinal sections of elastin-rich extracellular matrix sheet stained with acid Orcein at 0% strain (top) and 30% uniaxial strain in the horizontal direction (bottom). Note the straightening and thinning of the elastin fibers with increased strain. The scale bar denotes 10 μm. With permission from Ref. (29). (B) Structure of elastin in the parenchyma. V, AS, and AD denote vessel, alveolar sack, and alveolar duct, respectively. The scale bar is 200 μm. Adapted with permission from Ref. (261).

#### The proteoglycans

Within the lung, collagen and elastin fibers of the connective tissues are embedded in a hydrated gel, the ground substance. The composition of the matrix and the ratio of fiber to gel vary among tissues (117) and change during maturation and with disease states (131). Critical constituents of this matrix are the GAGs that are long chains of repeating disaccharide units that are variably sulfated and highly charged (117). There are several different types of GAGs (e.g., hyaluronic acid, chondroitin sulfate, heparan sulfate, dermatan sulfate, and keratan sulfate) whose molecular weights vary over three orders of magnitude implying that the polymer chains can contain as many as 104 units with a huge variability in size and structure (36, 220). Within the lung parenchyma, the most abundant GAGs are heparan sulfate and chondroitin sulfate. Except for hyaluronic acid, GAGs covalently attach to a PG core protein via a link tetrasaccharide to form proteoglycans. Similar to collagen, GAGs can also have secondary and tertiary structures by forming helical and randomly organized regions depending on the ionic environment and pH of the matrix (220). Images of the proteoglycans obtained by electron microscopy reveal an extraordinarily complex structure (36, 71). Proteoglycans can also associate ionically with one another to form large aggregates that exhibit an even higher level of hierarchical organization. Usually, the hyaluronic acid forms a long core to which various proteoglycans attach (Fig. 6).

Proteoglycans have a number of important biological roles (41, 209). For example, they can act as receptors on the cell surface and hence influence intracellular signaling (42). Proteoglycans on the cell surface can bind to growth factors and various other proteins and this binding can regulate the secretion of proteins, such as proteolytic enzymes that are involved in cell migration and tissue remodeling (125). Furthermore, the lateral and axial growth of collagen fibrils appears to be, in part, determined through interactions with the proteogly-cans (117,118, 204). Proteoglycans have also been reported to influence elastic fiber assembly (112). Thus, through a variety of indirect mechanisms, proteoglycans can significantly alter lung mechanics. As we shall see below, their elastic behavior can also directly influence the macroscopic stress-strain curve.

The majority of studies on proteoglycan mechanics have been carried out in cartilage. The elastic modulus of the cartilage tissue measured using the indentation atomic force microscopy appears to depend on the size of the probe (241). At the millimeter to micrometer scale, values in the order of 2 MPa were obtained, whereas at the nanometer scale a 100-fold lower modulus similar to that of simple agarose gel was seen. The reason is that at the millimeter scale, the elastic modulus includes contributions from both collagen and proteoglycans and the value is dominated by the properties of the collagen. In contrast, at the nanometer scale individual elements, mostly GAGs, appear to dominate elasticity (241). Little is known about the mechanics of the proteoglycans in the lung. It is likely that their role in lung function has been underestimated. Indeed, only a few studies have examined their role in lung mechanics (3, 45, 84).

### Lung cells

The lung parenchyma, prestressed by the Ptp, is a habitat of up to 40 different cell types. The prestress within the parenchymal tissue that is transmitted to the cells through the ECM adhesions influences cellular mechanical responses and mechanotransduction. On the other hand, forces generated within cells directly and indirectly affect the mechanical properties of the parenchyma.

From the point of view of mechanics, the most important cells are the contractile cells. Although recent findings suggest that both epithelial (263) and endothelial (137) cells can also contract, their contractile force is small. The stronger contractile cell types include smooth muscle cells in the alveolar duct and mouth and blood vessel walls and the myofibroblasts and fibroblasts (147). Beside airway and blood vessel walls, the protein α-actin associated with contractile force generation has been found in septal ends and bends but not in septal walls (187). Stimulation of the contractile machinery of these cells with different agonists induces local internal stresses in the fiber network of the ECM that can lead to changes in the viscoelastic properties of the lung tissue (65, 73, 182, 286). However, during contractile challenge, the mechanical properties of excised parenchymal tissue strips have also been found to vary with the number of medium-size airways in the sample (214). Thus, it is possible that part of the previously observed changes in the mechanical response during agonist challenge were in fact related to smooth muscle contraction and airway-parenchymal interaction. Nevertheless, the viscoelastic properties of the lung parenchyma are only moderately affected by the active tone of the interstitial cells (65, 286). A more important function of the interstitial cells is to actively remodel and repair the connective tissue during growth or after injury. For example, transforming growth factor beta (TGF-β), which is the main cytokine that stimulates fibroblasts to produce and secrete ECM molecules, is upregulated by mechanical stretch (97). Tensile forces also regulate the connective tissue growth factor that is able to stimulate ECM protein release through a TGF-β-independent pathway (218). As a result of such cellular processes, the nonlinear viscoelastic properties of the lung tissue can significantly change both at the organ and the alveolar wall levels (31). Thus, while cellular mechanical properties contribute little to the mechanical properties of the parenchyma in response to physical (e.g., deformation) or chemical stimuli (e.g., histamine challenge) over a short time period, they are responsible for the longer term homeostatic maintenance as well as the remodeling of the composition and structure of the ECM. Below we provide a short summary of the mechanical properties of cells and the reader is referred to a more complete account in the “Mechanics of the Cell” section.

#### Mechanical properties of cells

The cellular response to mechanical perturbation such as the prestress in the lung is governed by the cytoskeleton (CSK), an intracellular molecular network composed of filamentous biopolymers (actin filaments, microtubules, and intermediate filaments) and a number of actin binding and crosslinking proteins. The CSK is the site of the actomyosin contractile machinery that generates mechanical forces through ATP-dependent processes. Furthermore, the CSK also transmits forces across the cytoplasm. Many signaling molecules are immobilized in the CSK network and are very highly sensitive to mechanical deformation. In other words, the deformable cytoskeletal network provides a physical scaffold for mechanotransduction. To understand how mechanical forces regulate cellular function, it is necessary to know how cells develop mechanical stresses as they deform under applied mechanical forces. This mechanical response of the CSK is determined by the passive material properties of the molecules of the CSK, by the contractile forces generated within the CSK and by changes in biochemistry that modify cytoskeletal composition and structure through remodeling. It was found that the actin network has the bulk contribution to cell stiffness (>50%), whereas the contributions of microtubules and intermediate filaments are substantially lower (236, 270, 271). It was also found that measured cell stiffness (0.1–10 kPa) is much lower than the Young's modulus of any of the cytoskeletal filament or stress fibers (16, 68, 70), indicating that the network properties of the CSK play an important role in cell deformability. The roles of cytoskeletal filaments and cytoskeletal architecture on cellular mechanical behaviors are discussed in detail in articles “Stress Transmission within the Cell” and “Material Properties of the Cytoskeleton” in Comprehensive Physiology.

Microrheological measurements on living cells have shown that their mechanical behavior is governed by two major principles: (1) the CSK exists in a state of tension, that is, the prestress, which is critical for stabilizing cell shape and regulating cell rigidity; and (2) cellular rheological behavior is driven by very slow dynamics such that global viscoelastic responses of cells scale with time and frequency of loading according to a weak power law. Various theories have been proposed to explain the experimentally observed mechanical behaviors of cells, but there is a lack of consensus between those theories about the biophysical mechanisms that govern cell mechanical behaviors. Currently, none of those theories can provide a complete description of cellular mechanical responses that include both prestress-dependent and power-law behaviors. These models are discussed in detail in the articles “Stress Transmission within the Cell” and “Material Properties of the Cytoskeleton in Comprehensive Physiology

#### Cell-ECM interactions

Force transmission between the CSK and the ECM is not continuous. Forces are transmitted across the cell membrane at discrete sites of focal adhesion composed of integrin receptors with their extracellular domain attached to the ECM and their cytoplasmic domain to a cluster of proteins (α-actinin, paxillin, vinculin, talin, zxyin, etc.) that are physically linked to the cytoskeletal actin (52, 130). Therefore, the mechanical properties of the ECM and the focal adhesions also play an important role in mechanical behaviors of cells. More importantly, these mechanical properties together with biochemical cues influence intracellular signaling that in turn determines the type and amount of remodeling enzymes and load bearing ECM components produced and secreted by interstitial cells (47). Thus, cells maintain the composition and architecture of the ECM of the lung, while the prestress related to Ptp and the local mechanical properties of the ECM are key regulators of essential cellular processes (249). Figure 7 shows a small region of the lung tissue fixed in formalin and immunohistochemically labeled for types I and III collagen as well as cell nuclei. It can be seen that some nuclei appear rounded (green arrow), whereas some appear to be stretched in a direction parallel with the alveolar wall (blue arrow). One may speculate that forces in the wall primarily carried by the collagen at the high fixation pressure were transmitted to the nucleus.

Open in a separate window

Figure 7

Double-label immunohistochemistry of mouse lung tissue. The blue labels type I collagen, the brown corresponds to type III collagen, and the pink is cell nucleus. It can be seen that some fibers comprises almost exclusively type I or type III collagen (black arrows), whereas at several locations, the two collagen types also appear to colocalize suggesting that they mix and form composite fibers (red arrow) where the color is intermediate between blue and brown. Green arrow shows a round nucleus, whereas the blue arrow points to an elongated nucleus suggesting that the nucleus is under mechanical tension. From Ref. (249) with permission.

Although the mechanical properties of cells contribute less to the macroscopic mechanical properties of the parenchyma than those of the ECM, there are striking similarities between cellular and parenchymal tissue mechanics. Indeed, the mechanical properties of both cells and the parenchyma are characterized by a weak-power-law viscoelasticity and by strain hardening with increasing levels of the prestress. While these similarities may reflect common biophysical mechanisms that are reiterated at different length scales (76), it is not clear whether they occur by mere chance or have some functional advantages. For example, matching the viscoelastic responses of cells and the parenchyma would ensure that cells and the tissue matrix deform synchronously during breathing. This in turn would facilitate efficient stress transmission from the tissue to adhering cells which is important for mechanotransduction. On the other hand, there are phenotypic differences in alveolar epithelial cell stiffness that likely contribute to the observed heterogeneity of alveolar cell deformation during lung inflation and hence help better understand stretch-induced surfactant release (13).

### Surfactant and surface tension

The airways and alveoli are lined with a thin liquid film containing pulmonary surfactant which derives from type II epithelial cells. During volume excursions between FRC and TLC, the P-V curves of air-filled and liquid-filled lungs have remarkably different hysteretic behaviors; on the inflation limb, the recoil pressure is substantially higher in the air- than in the saline-filled lungs, whereas along the deflation limb this difference is much smaller (15, 90, 166, 231, 240, 269). The greater hysteresis in the air-filled lung is attributed to the hysteretic behavior of the alveolar surface film and associated phenomena such as airway opening as discussed later. The specific mechanical behavior of the surface film arises from changes in film composition and from asymmetry in the adsorption-desorption kinetics during the expansion-compression cycle (49, 92, 231). In addition to its direct effects on lung recoil, the surfactant also influences lung macrophysiology by ensuring alveolar stability, and preventing collapse at low lung volume by reducing the surface tension (11). Among the various components of the surfactant, phospholipids, and low-molecular weight hydrophobic surfactant proteins play a critical role in determining its biophysical properties that help maintaining low surface tension (123). The amount and composition of surfactant released by the type II epithelial cells into the air-liquid interface are largely determined by the dynamic stretching pattern of the lung parenchyma (10, 184, 282). Type II cells respond to the proximity of an air-liquid interface by a graded calcium response, which results in lamellar body secretion, implying that the physical environment in the vicinity of cells can initiate a complex biological response which, in turn, modulates the physical environment of the cell via a feedback loop (200).

Pulmonary surfactant must adsorb rapidly to the air-liquid interface to form a surface active film. Recent models suggest that phospholipid adsorption is mediated by pores that bridge the gap between the vesicular bilayer and the air-liquid interface (111). Proposed mechanisms are akin to vesicle-plasma membrane fusion during neurotransmission and are thought to require hydrophobic surfactant proteins to aid in bending of the outer lipid leaflet (212). Moreover, it appears that the adsorption and transformation of lamellar body lipids to the air-liquid interface is surface tension dependent and does not require tubular myelin as intermediary structure (26). Classic models assume that alveolar stability at low lung volumes requires the presence of a highly DPPC (Dipalmitoyl-phosphatidylcholine)-enriched monolayer, implying the “squeeze out” of unsaturated phospholipids during film compression (49). More recent studies combining fluorescence microscopy and atomic force microscopy have invoked the formation of multilayers during film compression, which may act as a surface-associated surfactant reservoir (290).

Surface tension has various effects on lung mechanics. First, surface tension directly contributes to the overall recoil stress of the parenchyma. Second, it distends the alveolar ducts and by distorting duct geometry, it indirectly alters the elastic properties of the associated connective tissues (234, 281). For small deformations, similar to those that occur during normal tidal breathing, the hysteresis of the surface film is negligible and surface film viscoelasticity may be less important than lung tissue viscoelasticity (219). Indeed, tissue hysteresivity, as defined above, was found to be very similar in isolated lungs with an intact air-liquid interface and in lung tissue strips that lack an air-liquid interface (213). While interfacial phenomena may not contribute greatly to energy dissipation during quiet breathing, they do exert a profound effect on lung recoil under conditions when lung expansion is limited by neuromuscular disease, obesity, or chest wall restriction (243). Moreover, the invariable need to call upon the lungs' reserve capacity during activities of daily living and the devastating consequences of impaired surfactant function in disease, underscore its critical role in lung biology as well as lung mechanics. A more detailed review of surface tension and surfactant biology may be found in the articles “Air-liquid interface and Alveolar” surface tension and lung surfactant as well as in the section on alveolar duct mechanics below.

### Mechanical Properties of the Normal Lung

In this section, we will overview the elasticity of the normal lung parenchyma at several length scales. We will start with the elasticity of collagen molecules, fibrils, and fibers followed by elastin fibers. We will then examine the elastic behavior of the alveolar wall and how these fibers fold within the ground substance of the wall. The parenchyma is a network of alveolar walls and hence it is also important to understand its network behavior. Next, we present various models of the tissue strip. Even though the tissue strip and the uniaxial stretching condition are not physiological, the tissue strip is a viable preparation and many important physiological and biological questions can be and have been answered at the level of the tissue strip. Finally, we present a comprehensive picture of the mechanics of the entire parenchyma of the normal lung during breathing.

### Molecular, fibril, and fiber elasticity

The structure of the lung is largely determined by the connective tissue network. The complex organization and the nonlinear viscoelastic properties of these tissue components lead to complex mechanical behavior. One of the main load-bearing components of the parenchyma is collagen. As discussed above, molecular kinks or crimps contribute to collagen elasticity. Crimps also exist at the fibril and fiber level (226). When thicker fibers in the tissue are stretched, it is the crimps along the fibers that first unfold followed by an unfolding of the crimps in the fibrils (175). Further stretching the fibers results in stretching of the triple helices and the cross-links which also raises the possibility of slipping of molecules and fibrils within the fiber (72). In addition to the elasticity of a single molecule, collagen fiber stiffness may depend on the number of fibrils through a given cross section, that is, the diameter as well as the type of cross-linking between molecules and fibrils. Both increasing diameter and cross-linking tend to increase fiber stiffness in normal collagen (7, 226). Furthermore, fibril length as well as small proteoglycan bridges between fibrils can contribute to the stiffness of collagen fibers (205).

The stress-strain curve of tendon composed of many fibrils arranged in parallel is nonlinear with a toe and a steep region (216, 242). The toe region is usually attributed to the crimps along the fibrils which, upon stretching, become straight (91) (Fig. 8). The composition of the fibrils and fibers is also important because fibers can contain a mixture of different collagen types (see Fig. 7). It has been argued that type I collagen is stiffer than type III (227) implying that fiber stiffness can depend on the relative amounts of type I and type III collagen within the fiber. Furthermore, there are notable species-related differences. A small amount (5–10%) of variation in amino acid composition between bovine and equine collagen can lead to a 2- to 3-fold difference in elastic modulus of in vitro cross-linked collagen gels (7). All these factors can give rise to significant inter- and intra-species variability in the mechanical properties of the alveolar walls. However, as both mechanical and biochemical factors contribute to collagen production and assembly, it is likely that there is a significant regional variability in collagen fiber properties within the lung. For example, even though not studied systematically, it is possible that fiber properties are different near the sharp edges of the lobes, where increased stability is required, compared to regions deeper inside the parenchyma. This variability is in addition to the different fiber content and mechanical properties of alveolar ducts and alveolar walls (60, 168). The former has been argued to be stiffer and hence contribute more to lung elasticity at higher lung volume based on relative volume changes along the P-V curve (170) as well as the larger concentration of fibers in ducts than in septal walls (168).

As discussed earlier, elastin is thought to behave as a linearly elastic material. However, thin ECM sheets containing elastin and proteoglycans do exhibit some mild nonlinear behavior during uniaxial stretch (29). The reason for this mechanical behavior is as follows. Elastin fibers in the tissue are not parallel. There is a distribution of angles relative to the direction of macroscopic strain. When the tissue is stretched, elastin fibers change shape and gradually reorient into the direction of macroscopic strain. This is a recruitment-like process in which more and more fibers become aligned with the strain and contribute more to stiffness. Nevertheless, elastin is still significantly softer than collagen and hence in a tissue containing both fiber types, the recruitment of collagen would ultimately dominate macroscopic elasticity. Figure 9 shows the stretch-induced shape change and recruitment of elastin together with the recruitment of collagen toward the direction of macroscopic strain in thin ECM sheets (28).

### Elasticity of the alveolar wall and the tissue strip

The stress-strain curve of individual tissue fibers may become linear once the crimps are unfolded (Fig. 8). However, for larger lung tissue samples that contain many fibers, the stress-strain curve often exhibits exponential-like stiffening (79, 172, 174, 182, 287). During uniaxial stretching, the non-linear stress-strain curve (Eq. 2) of the tissue strip can be written as:

where a is the amplitude of the stress-strain curve and can be related to the incremental elastic modulus near zero strain. Indeed, with ε0 = 0 it follows from Eqs. (3) and (6) that the first term disappears and we obtain:

The parameter c = ab characterizes thus the strength of the nonlinearity which is thought to reflect the progressive recruitment of collagen fibers. In other words, at low strains, most of the stress in the tissue is borne by the relatively compliant elastin fibers and hence a in Eq. (6) would primarily be a function of the volume fraction of elastin. As the applied strain increases, the initially flaccid collagen fibers start to straighten (Fig. 4B), fold into the direction of strain and gradually take up the load-bearing role. As progressively more of the stiff collagen fibers are recruited in this way, the bulk stiffness of the tissue increases commensurately. The parameter c is therefore a function of the volume fraction of both elastin and collagen as well as their ability to fold toward the direction of strain. As we shall see below, the folding and the recruitment of the fibers is related to the third major ECM component, the proteoglycans. Macroscopic tissue stiffness thus arises largely because of the way in which the various constituent fibers are organized with respect to each other, rather than being a reflection of the constitutive properties of any particular fiber type. Maksym and Bates (154) modeled this behavior analytically in terms of a linear chain of pairs of parallel spring and string units. If the springs (representing extensible elastin fibers) are all identical then the overall stress-strain behavior of the model is determined by the length distribution of the strings (representing inextensible collagen fibers). Alternatively, the model can be constructed using strings of equal length and springs with distributed stiffness. They found that the exponential nonlinearity in Eq. (6) could be explained by a power-law distribution of collagen fiber properties, which is compatible with morphometric assessment of collagen in the lung (168, 232).

Maksym et al. (156) also extended their model into two dimensions by using a triangular network of line elements each containing a parallel combination of a crimped collagen fiber and an elastin fiber. The force-transmission through such a model displays an interesting heterogeneous spatial distribution similar to the phenomenon known as percolation suggesting a complex network behavior (see article “Complexity and Emergence” in Comprehensive Physiology). When the microstrain in a material follows the macrostrain, the deformation is called affine. This is the case in a homogeneous continuum. In a network, however, the heterogeneous viscoelastic properties of neighboring segments can significantly contribute to a given segment's microscopically observed mechanical behavior. The deformation and folding of an isolated segment can thus behave differently from a segment that is part of a network. Using diffuse light scattering, Butler et al. (39) estimated that the microstrain followed the macrostrain during externally imposed isovolumetric uniaxial deformation of the lung. In contrast, Brewer et al. (31) directly imaged individual alveolar walls of lung tissue strips during uniaxial stretching and found that alveolar walls did not follow the macroscopic deformation (Fig. 10). The former study superimposed the uniaxial deformation on a uniformly prestressed state of the air-filled lung, which homogenizes the system, whereas in the latter study, tissue strips submerged in fluid were stretched starting from the relaxed state. Additionally, Butler et al. (39) considered length scales of 1 to 2 cm sampled by an optical probe, and their data may correspond to the average behavior of several hundred alveoli. In the study by Brewer et al. (31), the authors examined the mechanics of individual alveolar walls at length scales of about two orders of magnitude smaller than those of Butler et al. (39). The results in Figure 10 suggest that continuum analysis cannot be used to evaluate the configurations of individual alveolar walls and that network models must be used to describe the behavior instead. As we shall see in the next section, such network effects can have significant impact on the focal development of lung diseases. Here it is sufficient to point out that these network effects also represent recruitment processes at the level of alveolar walls at a length scale of about 2 orders of magnitude larger than those in Figure 9.

Open in a separate window

Figure 10

Fluorescent images of the same alveolar region labeled for collagen in a normal rat lung. Left: before deformation; Right after 30% uniaxial stretching vertically. The black lines show alveolar walls and the red lines are their new length and orientation after stretching. The yellow arrow points to the same septal wall junction. Note the significant change in angle between the two septal walls. Scale bar denotes 100 μm. From Ref. (31) with permission.

To describe the deformation of individual alveolar walls, it is thus necessary to include a realistic geometry such as a hexagonal network in a model as originally proposed by Mead et al. (165). Such hexagonal network models of the lung parenchyma have since been developed (31, 45, 277). Simulations using the hexagonal network to mimic the microscopically observed deformation of the alveoli in normal, hypotonic, and hypertonic solutions suggest that the folding of the alveolar wall and collagen during uniaxial stretching is elastically limited by the proteoglycan matrix (45). The reason is that the proteoglycans are highly sensitive to the osmolarity of the bath, whereas collagen and elastin are much less sensitive. In hypertonic solution, the negatively charged proteoglycans collapse resulting in low stiffness, whereas in hypotonic solution they become inflated and their stiffness increases. Thus, the proteoglycans by their compressive resistance hinder the folding of fibers into the direction of the macroscopic strain and so contribute to the elastic behavior of the tissue. A simple hexagonal model without prestress would be unable to mimic the exponentially increasing stress-strain curve of lung tissue strip [e.g., Eq. (6)] because without external constraints such a model is unstable and angles of the hexagons collapse upon stretching that leads to zero elasticity. However, this is not observed experimentally in the tissue strip as Figure 1A demonstrates.

Accordingly, a structurally reasonable model of the tissue strip needs to include both line element elasticity and a mechanism that hinders the collapse of angles. This can be achieved by using an angular spring or bond bending that resists the folding of two neighboring line elements (9). The analysis of such a system needs to include the energy associated with stretching the springs and bending the angle θ between two springs:

(8)

where ε is the macroscopic strain and Δθ indicates the local change in angle between adjacent springs. The first and second summations in Eq. (8) go through all the springs and all the nodes, respectively. The ai and ci are the linear and nonlinear parameters of an individual spring based on Eq. (7) and the qj are the bond-bending constants. The equilibrium configuration of the network following deformation corresponds to minimum of U. Such a model is shown in Figure 11 for two network configurations corresponding to two values of q (taken to be the same throughout the network) while the line element parameters were kept constant (45). Notice that the minimum energy configuration of the same exact structure is very different when the bond-bending is low (left panel) or high (right panel). This means that (1) the Poisson's ratio of the network is significantly influenced by the mechanical interaction between collagen and proteoglycans; and (2) the deformation pattern of the alveolar wall network does not follow the macroscopic deformation, that is, it is not affine. However, by incorporating this mechanical interaction between collagen and proteoglycans into the hexagonal network model and comparing its stress-strain curve to measured data, the average Young's modulus of a single alveolar wall could be calculated and it was estimated to be about 5 kPa. Furthermore, by taking into account the volume fraction of collagen fibers in the alveolar walls, a lower limit of collagen fiber stiffness in the alveolar wall was also estimated and a value of 300 kPa was obtained when tissue was stretched to 30% uniaxial macroscopic strain (45).

Open in a separate window

Figure 11

Effects of the bond-bending parameter q on the configu-ration of the elastic network model at 30% strain in the vertical direction. A: stiff network with bond-bending constant q 100. B: soft network with q = 0.01. Color is proportional to energy carried by the springs. The maximum energy values corresponding to dark red on A and B are different. From Ref. (45) with permission.

The tissue strip preparation is simple and popular partly because of its easy manipulation and partly because the complexities associated with surface tension and circulation are eliminated. The lack of circulation does not pose a technical problem. The tissue strip can be placed in cell culture media that keeps the cells alive. The lack of surface tension is probably also not a serious limitation because the recoil due to surface tension can be restored by applying a prestress, albeit uniaxial in the strip as opposed to the uniform three-dimensional deformation in the lung. Despite these limitations, mechanobiology of lung cells can be conveniently studied in the tissue strip preparation since the cells are in their native ECM. While the study by Cavalcante et al. (45) offered novel insight into the mechanics of the tissue strips at multiple scales, it is still based on a two-dimensional analysis, and the mechanical influence of the alveolar ducts were neglected. Hence a more realistic three-dimensional model of the tissue strip integrating the detailed mechanics of the alveolar wall ECM with the mechanobiology of the interstitial cells embedded in a network of alveoli will eventually be needed.

### Mechanics of the lung parenchyma

The mechanical properties of the parenchyma in situ are different from those of the tissue strip due to the presence of prestress related to Ptp and surface tension at the air-liquid interface. Both of these factors also influence the three-dimensional architecture of the parenchyma and the corresponding mechanical properties that are considered next.

#### Mechanics of the alveolar ducts

Pressure-volume measurements on air- and liquid-filled lungs (15, 90, 166, 231, 240, 269) led to the conclusion that the lung tissue and alveolar surface tension are major stress-bearing components of the lung. Orsós (189) and Weibel and Gil (273) described the lung tissue as comprising three interconnected tissue systems (Fig. 12): (1) a peripheral tissue system consisting of the pleural membrane and the interlobular membranes and their extensions into the parenchyma; (2) an axial tissue system comprising sheaths enveloping airways and pulmonary arteries into the acini, where they form a network of the alveolar duct tissue; and (3) a parenchymatous tissue system of delicate septa that links the axial and the peripheral tissue systems. The appearance of the tissues obtained from scanning electron micrographs of air-filled, saline-filled, and detergent-rinsed air-filled lungs suggested that at a given lung volume, the bore of the alveolar duct is the greatest in the detergent-rinsed lungs and the smallest in the saline-filled lungs. Furthermore, the alveolar septa of air-filled and detergent-rinsed lungs appear taut and tensed, whereas in the saline-filled lungs they are flimsy and slack (14, 90). These differences were attributed to the variation in the alveolar surface tension among the conditions that was the highest in the detergent-rinsed and the smallest (zero) in the saline-filled lungs.

Open in a separate window

Figure 12

Schematic drawing of the connective tissue systems in the parenchyma according to the Wilson and Bachofen model showing the alveolar duct with its axial tissue fibers organized in a helical structure, as well as the septal and peripheral fibers. The heavy arrows indicate the distending action of surface tension that exerts radially outward pull on the axial fibers of the alveolar duct. Adapted from Ref. (281) with permission.

Modeling the lung parenchyma as a two-compartment system composed of tissue and surface film acting mechanically “in parallel” has not been satisfactory. For example, Bachofen et al. (15) assumed that the difference in the recoil pressures Ptp of air-filled and saline-filled (Ps) lungs can be entirely attributed to surface forces, that is, PtpPs = 2γ S/3 V, where γ is the surface tension, S is the alveolar surface area, and V is the lung volume. If this were the case, then during uniform lung expansion, S would be proportional to V2/3, and hence (PtpPs)/γ ∝ V−1/3. However, experimental data on lungs in which γ was maintained constant showed that (PtpPs)/γ is nearly independent of V below γ = 20 dyn/cm and increases with V for γ > 20 dyn/cm (231, 238). This discrepancy suggested that during lung expansion, S does not change uniformly with V2/3 and that the reason for this behavior may be due to the distortion of S caused by surface tension. Subsequently, several investigators pointed out that γ acts in two ways: directly, by providing a recoil pressure equal to 2γ S/3V, and indirectly, by distorting parenchymal geometry and thus providing an additional tissue component of the recoil pressure (153, 279). These observations as well as the morphometric studies by Weibel and co-workers (14, 90) led to the development of a novel microstructural model of the alveolar duct by Wilson and Bachofen (281).

According to this model, the peripheral tissue network predominantly provides Ps of the saline-filled lung. Increasing γ at a given V has a minor influence on this network. The axial tissue network of the alveolar duct and the septal tissue network are not affected by volume changes in the saline-filled lung, but they are distended by γ in the air-filled lung. Wilson and Bachofen (281) assumed that the tension carried by the alveolar septal tissue is much smaller than the tension carried by the alveolar duct tissue and thus they neglected the contribution of the septal tissue. An analysis of the mechanics of the model yielded the following relation between recoil pressure, alveolar surface area, surface tension, and lung volume

(9)

where n and L are the number and the length of fibers forming the alveolar duct lattice and F is the force on the fibers. The axial fibers are described as a system of intersecting helices (Fig. 12). The first term on the right-hand side of Eq. (9) is due to surface tension, whereas the second term is indicative of the indirect contribution of surface forces through distortion of the alveolar duct lattice. It was found (281) that Eq. (9) can fairly accurately predict data for PPs versus S relationships for lung volumes up to 80% of TLC, which were obtained from P-V and morphometric measurements on air-filed, saline-filled, and detergent-rinsed lungs (14, 90). For higher lung volumes, the authors suggested that the contribution of the alveolar septal tissue might need to be included.

While the Wilson and Bachofen model was very successful because it was capable of predicting the quasi-static P-V behavior of the lung, it also has several limitations. First, more recent electron microscopic imaging results predict larger changes in epithelial basement membrane surface area with lung volume (265) than the original studies (14, 90) which the model was based on. Second, the model does not include separate mechanical properties of the collagen and elastic fibers and their contribution cannot be understood during lung inflation. Third, the model considers only the average properties of the parenchyma and it cannot be used to infer the effects of regional heterogeneities on overall mechanics. Finally, the model is elastic and hence it is unable to predict lung mechanics during breathing. With regard to structure, several newer models of the alveolar duct were proposed subsequently. In all these models, alveolar geometry was described as a collection of space filling 14-sided regular polyhedra (tetrakaidecahedron) and the ducts and alveoli are formed by opening specific common faces between the polyhedra (58–62, 82). While these models provided a description of the alveolar duct morphology which is consistent with average morphometric data, they are limited to a single acinus and hence do not incorporate the variability in structure. A stochastic model of a large collection of irregular polyhedra that incorporated viscoelastic effects was also proposed (145). With regard to the mechanical properties of the structural elements, a refined deterministic model of the acinus was introduced more recently by Denny and Schroter (61). These authors developed a three-dimensional model of the acinus that also included separate and realistic viscoelastic properties of the collagen and elastin fibers together with surface tension properties of the air-liquid interface. They used the structure as a representative model of the acinus and calculated its resistance and elastance as a function of frequency. The model predicted little difference in Rti between the air-filled and lavaged lungs. Additionally, surface tension contributed significantly to both Rti and Eti in the air-filled lung. Interestingly, the surface tension tended to amplify any existing tissue hysteresis despite the fact that the surface area-tension curve itself exhibited very little hysteresis during small amplitude tidal oscillations. Future efforts could incorporate the effects of proteoglycans, interstitial cells and, given the ever increasing computational power, multiple acini and gravity to explore the effects of heterogeneity and mechanical interaction among the acini.

Experimentally, different studies appear to provide a somewhat controversial picture of the role of surface tension. For example, Stamenovic and Barnas (235) found that the dissipative properties of the lung's ECM are the major determinants of Rti, whereas the elastic properties of both the tissue and surface film are important determinants of Eti. In contrast, Navajas et al. (183) found that Eti was smaller in isolated liquid-filled lungs than in air-filled lungs, whereas η was similar in both lungs. Hence, they concluded that surface tension accounts for a considerable part of both elastance and resistance of the air-filled lung within the volume range of normal breathing. Sakai et al. (213) reported similar values of viscosity in isolated lungs and tissue strips despite different deformation and the lack of surface film in the latter. Additionally, for breathing frequencies, the energy loss in the lung parenchyma is a much smaller fraction of the stored elastic energy in shear than during isotropic volumetric oscillations and it is Ptp and not the properties of the surface film that primarily determines the lung's dynamic properties in shear (53). Furthermore, dynamic oscillations invoking small changes in surface area of a surfactant monolayer have shown negligible hysteresis in the surfactant itself (219). Others also argued that parenchymal mechanics are largely dominated by the connective tissues for tidal-like oscillations (286). It is also possible that the contribution of the surface film to parenchymal mechanics during breathing is species dependent since the average size of the alveolus shows a moderate increase with body mass while minimum surface tension appears to be independent of size suggesting that the role of surface tension may be less important in larger species (150). On the other hand, septal wall thickness and the amounts of collagen and elastin fibers in the wall increase linearly with alveolar diameter (171) suggesting a species-specific fine balance between surface and tissue forces. Nevertheless, it appears now that during breathing, surface film mechanics contribute much more to Eti than to Rti. More experimental studies are needed to better understand the role of surface film and surfactant and their interaction with tissue elasticity in the lung during tidal breathing.

#### Parenchymal interdependence

Mead and co-workers (165) introduced the concept of mechanical interdependence into pulmonary mechanics to explain how the parenchyma resists nonuniform deformation. They pointed out that Ptp acting on the pleural surface is transmitted into deeper regions of the lung by the interconnected parenchymal network. Specifically, they showed that if the parenchymal network is homogeneous and alveolar gas pressure is uniform, then the stress at any point in the parenchyma equals Ptp and thus “all distensible regions could be thought of as being exposed to transpulmonary pressure.” If, however, gas pressure in one region of the lung is lower than in the surrounding tissue, the net inward pressure would cause the lower pressure region to shrink. Through mechanical interdependence, however, this shrinkage would be resisted by the distortion of the surrounding parenchyma that produces an outward stress and hence counterbalances the net inward pressure. The authors identified three mechanisms by which the parenchymal network resists deformation: change in the spacing of the network elements, change in their orientation and change in the forces carried by those elements. The first two mechanisms are purely kinematic and are determined by the topological arrangement of the microstructure. The third mechanism depends on the material properties of tissue elements and surface film. These notions are consistent with the network modeling of Cavalcante et al. (45) discussed above.

As described earlier, the bulk modulus represents the lung's ability to resist uniform volume change and is defined as κ = VdPtp/dV. The value of κ can be obtained from the local slope of the P-V curve (238, 240) and was found to increase faster than linearly with increasing Ptp, for example, from 3 Ptp to 6 Ptp between 40% and 90% TLC for dog lungs (142). The shear modulus micron representing the lung's ability to resist shear deformation cannot be inferred from P-V measurements but rather from measurements where a nonuniform deformation is applied to the lung. Values of micron were obtained from local indentations applied to the lung surface via a cylindrical punch (98, 142, 143, 238, 240). It was found that micron increases approximately linearly with increasing Ptp; for example, for dog lungs μ ≈ 0.7 Ptp over a wide range of volumes (142). Two conclusions can be drawn from these results. First, since κ is much larger than micron, the lung responds to a force applied locally by changing shape more than by changing volume, and therefore the gas volume per unit mass of tissue remains uniform in the deformed region. Second, since micron depends only on Ptp, the material properties of the parenchymal tissue and the surface film seem to have little effect on micron. This, in turn, suggests that the lung resists shape distortion primarily by reorientation and changes in spacing of its tissue and surface film elements, and not by deforming those elements.

The linear shear modulus versus prestress relationship is characteristic of the so-called stress-supported structures. In the absence of prestress, these structures become unstable since rigidity and connectedness of their structural elements are insufficient to fully constrain their freedom to move and deform. Consequently, these structures do not have intrinsically stable shape (like, e.g., rocks, beams, or rubber), and thus they require the prestress for shape stabilization. When an external force is applied to a prestressed structure, its taut structural elements undergo primarily reorientation and change in spacing, and to a lesser degree extension/shortening, until a new equilibrium configuration is attained. The greater the prestress, the smaller the configurational changes are, that is, the greater the stiffness of the structure. Indeed, in such structures micron can remain finite and small even if the stiffness of its members becomes very large. This explains why micron is small despite the relatively stiff connective tissue fibers that comprise the alveolar walls and why it must be a nearly constant fraction of Ptp (37, 78, 234). Microstructural models of the lung parenchyma that incorporated the above principles successfully predicted the observed linear dependence of micron on prestress, regardless of whether they used very simple line-element networks (38, 277) or morphologically more realistic surface-element networks (59, 134). Thus, the structural connectedness and the stabilizing role of the prestress appear to be key features of parenchymal mechanics that lead to the linear μ-Ptp relationship, whereas the geometrical and material properties of the structural elements may be secondary.

The situation, however, appears to be different at low lung volumes. Cavalcante et al. (45) studied the contribution of proteoglycans to parenchymal elasticity as described earlier. The authors found that proteoglycans contribute to the stress-strain properties as well as to the macroscopic shear modulus of parenchymal tissue strips. Microscopic imaging measurements showed that at the level of individual alveoli, bath tonicity, which only influences proteoglycans (and pulmonary cells) but not elastin or collagen, does change the configuration of the alveolar wall network. Importantly, they showed that stretching of relaxed strips causes an immediate buildup of stress, that is, the tissue exhibits a finite elastic modulus at zero stress. This finding appears to be at odds with the idea that in the whole lung resistance to shear is conferred to the parenchyma by Ptp and that if Ptp = 0, resistance to shear should be lost. If the parenchyma indeed behaved as an unstable structure in the absence of prestress, then uniaxial stretching of a relaxed parenchymal strip should develop no stress until all alveolar walls become aligned with applied strain and individual walls begin to stretch. The results of Cavalcante et al. (45) therefore suggest that the parenchymal structure is not unstable even if Ptp = 0 and that proteoglycans provide stability and resistance to shear at low lung volumes. The authors argued that at the scale of the alveolar wall ECM, proteoglycans, due to their large negative surface charge density, resist compression and shear and hence they serve as a bond-bending spring producing elastic resistance against alignment of fibers and alveolar walls with the direction of macroscopic strain (Fig. 11).

Data from punch indentation experiments show that at low Ptp, the relationship between micron and Ptp deviates from the observed linear relation at medium and higher Ptp (237, 240). While this could be a result of air-trapping that occurs at low lung volumes, it can also indicate the increasing contribution of proteoglycans as discussed above. In fact, this deviation is even more pronounced in liquid-filled lungs where there is no effect of air trapping (240). Taken together, the picture that emerges is that at low Ptp, where surface tension is low and alveolar architecture is not fully recruited, proteoglycans in the ECM may play a major role in resisting lung distortion. With increasing Ptp, the alveolar network becomes fully recruited and the contribution of structural mechanisms becomes predominant.

#### Contributions of non-parenchymal and cellular structures

It should be also mentioned that non-parenchymal structures may play role in pulmonary mechanics. For example, using indentation tests, it was determined that the pleural membrane contributes ~20% to recoil pressure during uniform expansion of the lung (98, 188) and at least 30% to the resistance to shear (119, 233). However, tissue hysteresivity of the pleura was estimated to be small (251). The deformation of the parenchyma associated with the axial forces along the airways may locally distort the parenchyma. It was inferred that the parenchyma stabilizes airway length against transmitted axial forces, and perhaps even forces on airway plugs (107). Interestingly, the elasticity of the bronchial and vascular trees was found to contribute to lung recoil pressure only up to 10% (230).

Other potentially important contributors to lung elasticity include cellular structures. Initial studies found that activation of the airway smooth muscle can cause a decrease in lung compliance in vivo (51, 284, 285). Among the four major alveolar cell types, alveolar type I and type II epithelial cells, endothelial and fibroblasts, the latter cells are the most contractile and have the highest elastic modulus (13). Additionally, there are smooth muscle and perhaps contractile myofibroblast cells in the alveolar ducts (147, 187) that are highly contractile. Therefore, it is possible that the contraction of these cells modulates lung elasticity. However, at the whole lung level, it is important to partition the bronchoconstriction response into separate airway and tissue properties because airway heterogeneities can produce apparent increases in estimated lung properties. Indeed, using gases of different viscosities, it has been demonstrated that airway smooth muscle activation mainly causes time constant heterogeneities and little change occurs in Rti and Eti at least in small animals (152). This finding is also in agreement with tissue strip studies that reported only a modest 5% to 20% increase in stiffness during contractile stimulation depending on the concentration of the agonist (65, 182, 286). Interestingly, a recent study reported that the overall mechanical behavior of the lung may be sensitive to a specific cytoskeletal intermediate filament protein, desmin, which is expressed in smooth muscle cells of the lung. Specifically, desmin-null mice showed lower lung stiffness and recoil pressure both at baseline and following induced airway constriction than control mice (222). However, as discussed earlier, it is generally thought that the most important structural element of the CSK that dominates cell stiffness is actin (236, 270, 271). Furthermore, it is unclear whether the ECM of the desmin knockout mice is comparable to that of the wild-type mice. Nevertheless, this area at the interface of lung biology and parenchymal mechanics warrants more attention and further studies are needed to determine the specific role of various proteins as well as the effect of different biological interventions. Before physiological conclusions can be drawn from any biological intervention, it is crucial to understand the role of heterogeneities in the whole lung response. Besides using gases of different viscosities, another possibility is model-based partitioning of heterogeneities from physical changes in tissue properties (127, 254). Alternatively, the tissue strip can be studied in isolation as it is minimally sensitive to airway heterogeneities.

### Stability of the Lung

Generally, a thermodynamic system responds to an applied load by developing a restoring force. For example, when air is pumped into a rigid chamber, the pressure inside the chamber increases monotonically. However, during inflation of the collapsed lung, the lung does not always develop an increasing restoring force; instead, the pressure inside the lung can decrease intermittently due to the sudden opening of airways and alveolar airspaces. Thus, the local slope of the P-V curve of the lung can become negative indicating the presence of instabilities. Lung parenchymal stability is closely associated with collapsing of the airways and air spaces known as atelectasis. In normal lungs, atelectasis is present only during abnormal conditions such as the first inflation of degassed lungs, ventilation with pure oxygen, pressing on lung surface at low lung volumes or ventilation of anesthetized subjects. Figure 13 shows a series of images of a region of an isolated lung during inflation from the collapsed state. It can be seen that the atelectatic regions gradually disappear as Ptp increases. Conversely, when the regional distending pressure of the parenchyma decreases below some critical value, the region can collapse hindering gas exchange. In lungs with abnormally high surface tension, atelectasis can occur spontaneously with the lungs inflated by small but uniformly positive Ptp (50, 231). The occurrence of atelectasis in prematurely born infants (32, 33) is associated with abnormally high and nearly constant surface tension (11). In this section, we will examine instabilities associated with surface tension and with the dynamics of airway opening and closing.

Open in a separate window

Figure 13

Images of a region of an isolated lung at successive inflation pressures. The inflation was started from the collapsed state and the bottom, middle, and top images correspond to transpulmonary pressure of approximately 25, 27, and 30 cmH2O. Dark red corresponds to collapsed regions. Notice that as inflation progresses, the pink aerated regions gradually penetrate into the atelectatic region by pulling the underlying alveoli open (Z. Hantos and B. Suki; unpublished data).

### Analysis of static instability

The earliest models of parenchymal stability depicted the alveolar structure as a collection of independent spherical balloons connected to a common airway tree (203, 269). The tension in the balloon wall was assumed to be the sum of the surface tension γ and tissue tension T. Using the law of Laplace, the transmural pressure p that keeps each balloon inflated is

where r is the radius of the balloon and T and possibly γ depend on r. As the lung expands, r increases and, according to Eq. (10), p tends to decrease unless the total wall tension increases faster than linear with increasing r. Since the alveolar septal tissue exhibits exponential-like nonlinear tension-extension behavior (79, 182, 287), it is reasonable to assume that T/r increases with increasing r. If γ were constant, then γ/r would decrease with increasing r and if γ were sufficiently high, then p would reach a maximum and start to decrease as r further increases. The first balloon with decreasing p would set up an unstable situation: all balloons with smaller r and higher p would empty into the balloon with a bigger r and lower p. This implies that if the alveoli had slightly different size, constant and elevated γ would make the lung intrinsically unstable. In the normal lung, this instability is, however, avoided because type II alveolar epithelial cells secret surfactant that reduces γ and makes it dependent on r such that γ/r increases with increasing r (81, 239).

In his conceptual model of stability, Mead (164) identified three mechanisms of parenchymal stability which he referred to as geometrical stability, surface film stability, and tissue stability. As a criterion of stability, Mead employed the notion that an increase in Ptp must be accompanied by an increase in lung volume, that is, a positive slope of the P-V curve and thus a positive κ. Fung (81) and Wilson (278, 280) formalized Mead's idea as follows. Because the lung's recoil pressure equals the sum of all tissue and surface forces transmitted across any plane through the parenchyma per unit area of the plane, the rate of increase of these forces (i.e., tissue and surface stability) during lung expansion must exceed the rate of increase in spacing between the force-bearing components of the parenchyma (i.e., geometrical stability) to have increasing recoil pressure. Because the contribution of γ to the recoil pressure depends on the alveolar surface-to-volume ratio (S/V) [the first term on the right-hand side of Eq. (9)], Wilson (279) concluded that if γ were constant and high enough, this term would decrease with increasing V leading to a decrease in Ptp and thus κ < 0, which implies instability. Stamenovic and Smith (237) studied the P-V behavior of lungs with constant and elevated γ and did not observe negative P-V slopes. Instead, they observed very small slopes accompanied by regional atelectasis at low volumes. They argued that these atelectatic regions occur as V decreases, with most of the expiring volumes from the collapsing regions being transferred to the remaining expanded regions. Fung (81, 83) analyzed regional collapsing of the air spaces and concluded that if a region that tends to collapse is entirely surrounded by the parenchyma, then its collapsing will be prevented through the stabilizing effect of mechanical interdependence. A planar regional collapse would be marginally stabilized by the surrounding parenchyma and would be susceptible to collapse if a small local compression force is applied. However, this analysis did not take into account the effect of alveolar surface film properties. Similar conclusions were also reached by Graves et al. (95). Stamenovic and Wilson (239) attempted to reconcile the different results obtained from the various models. Using both continuum and microstructural analysis, they showed that in a homogeneous lung, the parenchyma would be stable and local atelectasis would not occur at any positive Ptp. If, however, the alveolar S/V ratio is inhomogeneous, regions of higher S/V ratio would collapse if surface tension is elevated and constant. This may explain the observed atelectatic regions at low volumes of lungs where surface tension is high and insensitive to changes in alveolar surface area.

### Dynamic instabilities

As described in the previous section, alveolar stability during a small deformation is determined by the relation among the local shear and bulk moduli and surface tension. The analysis of stability can be extended to include airways and gas flows during slow inflation as follows.

Figure 14A depicts the P-V curve of an isolated rat lung during inflation from the collapsed and degassed state to TLC (4). The lung's recoil pressure Ptp was measured with respect to atmospheric pressure as a function of the volume displacement V of a piston that slowly inflated the lung. It can be seen that with increasing V, Ptp also increases, but this increase is not monotonic and Ptp intermittently decreases. Therefore, the elastance E = dPtp/dV can take both negative and positive values. Successive values of E were estimated as the local slope of a straight line fit within a short moving window along the P-V curve. The magnitudes Eneg of elastance with negative values show fluctuations and the distribution Π of Eneg was found to be exponential:

where E0, the characteristic value of Eneg, increased with inflation rate (Fig. 14B). It was found that the pattern of Eneg varied from inflation to inflation, but the distribution in Eq. (11) was reproducible. This excludes the possibility that Eneg arises from tissue rupture.

Open in a separate window

Figure 14

(A) P-V curve during the inflation of a degassed rat lung. The inset shows a magnification of a region with many local negative elastance patterns. (B) Distributions of negative elastance from 10 inflations at rates of 2.0 ml/s (triangles) and 0.5 ml/s (circles). The regression line fits to the measured distributions are shown by dashed lines. The solid lines correspond to the distributions of negative elastance from 1000 simulated inflations of an 18-generation symmetric binary tree. (C) An example of the P-V curve from the inflation of the model. The inset shows a magnification of a region with many local negative elastance patterns similar to those in Fig. 14A. The red line in the inset traces an avalanche shock. Adapted from Ref. (4).

To understand the existence and fluctuations of the negative elastance, a dynamical model of a system consisting of a piston chamber and a lung was developed (4). The lung was modeled as a binary tree terminating in elastic alveoli (4). The root of the tree, the trachea, was connected to the chamber containing air at atmospheric pressure, corresponding to the experimental conditions. Inflation was then simulated by increasing V at a constant rate. At the beginning of the inflation, all airways were closed and no alveoli are connected to the root, so a small increase in V increased Ptp. A closed airway opens when Ptp exceeds the airway's critical opening threshold pressure (86) Pth, which was assumed to be a random variable uniformly distributed between 0 and 40 cmH2O (the pressure at TLC). When a segment opens, the pressure propagates deeper into the lung. If the Pth of either daughter airway is also smaller than Ptp, then this daughter opens together with its parent. This process leads to an avalanche of openings involving a large number of airways and alveoli (247). The newly opened airways and alveoli increase the volume of the lung, the gas suddenly redistributes in a larger volume and consequently the pressure decreases according to Boyle's law. The reduction in Ptp, however, can terminate the propagation of the avalanche and Ptp increases again due to the steady influx of air. This phenomenon of the decrease in Ptp following a single avalanche and the subsequent rise in Ptp was termed an avalanche shock (Fig. 14C inset), which is not a traditional propagating shock wave. The time course of an avalanche shock is smoothed by relaxation processes due to flow resistance of the airways such as regional airway resistance and viscoelasticity of the parenchymal tissue. The decreasing part of an avalanche shock corresponds to instabilities characterized by negative elastance. The continuous increase in Ptp is intermittently interrupted by avalanche shocks of different sizes until all air in the piston chamber is injected into the lung.

Simulations using the above model with an 18-generation airway tree ending in elastic alveoli produced P-V curves (Fig. 14C) that are similar to the P-V curves observed experimentally (Fig. 14A). The local Eneg was then determined in the same way as for the measured data. The distributions Π (Eneg) from the simulations fit the experimentally ob tained distributions for both slow and fast inflation rates (Fig. 14B). These results indicate that inflating the lung from a state in which a considerable part of the gas exchange region is collapsed is a complex, nonequilibrium dynamical process characterized by regions of instabilities that manifest macroscopically as negative elastance. For slow inflation, the system has enough time following each avalanche to reach equilibrium, so the individual avalanche shocks are distinct and nonoverlapping. With increasing inflation rate, however, avalanche shocks due to separate avalanches increasingly overlap, resulting in smoother P-V curves with fewer regions of negative elastance. Thus, the mechanism for the paradoxical negative elastance and its distribution arises from the avalanche shocks involving the sudden recruitment and subsequent relaxation of a large number of airways and alveoli.

It can be concluded that the instabilities that were attributed to a combination of the static properties of the alveolar liquid lining (50) and the nonuniform deformation of the lung tissue (239) can be dynamic in nature. These simulations also suggest that both airways and alveoli are inextricably involved in the processes producing macroscopic instabilities of the lung during inflation. Since the topological properties of the airway tree are also involved in these instabilities, the associated phenomena can become very complex due to the avalanches (see article on “Complexity and Emergence” in Comprehensive Physiology). Furthermore, it appears that the rate of deformation plays an important role in the formation of instabilities and their contribution to the global mechanical properties of the lung. Next, we examine the influence of taking into account the rate of deformation of individual airway opening and closing on dynamic lung elastance.

### Instabilities during cyclic deformations

Opening and closing of lung units are not processes that occur instantaneously once critical pressure thresholds have been crossed. A closed airway takes a finite amount of time to open when it is subjected to sufficient pressure (86), and the same is true for an open airway that closes when pressure is reduced enough (44). This is also evident from the progressive decrease in lung compliance seen during mechanical ventilation following a recruitment maneuver (6). Both the opening and closing processes together with the time to open and close are significantly influenced by the presence of surfactant. The precise biophysical mechanisms governing these processes are beyond the scope of this article and here we concentrate on a simple and more phenomenological description that can be used to simulate whole lung behavior. Such a model was developed by Bates and Irvin (20). The temporal dynamics of recruitment and derecruitment in the lungs have been simulated by representing the lungs as a parallel collection of compliant units each served by an airway that connects to a common entry point. Each lung unit is associated with a virtual trajectory, parameterized by the quantity x that can take values between 0 and 1. Each unit can exist in one of two states, either fully open or fully closed, with the transition between the two states governed by the behavior of x, as follows. If the pressure, P, applied to an airway exceeds the airway's critical opening pressure, Po, then x increases toward 1 with a speed proportional to the difference PPo and a constant of proportionality so. Similarly, movement toward 0 occurs with speed sc(P 7#x2212; Pc), where Pc is a critical closing pressure, sc is another constant of proportionality, and Pc < Po. It is assumed that x does not change for PcxPo. If an airway is closed and its value of x reaches 1, the airway immediately opens. Conversely, if the airway is open and x reaches 0, the airway closes. The quantity x thus does not correspond to anything physical in particular; the virtual trajectory is merely an empirical mechanism that allows each airway to exhibit hysteretic opening and closing as the pressure applied to it varies in time. Even so, the airway behavior produced by the virtual trajectory mechanism shows similarities to that observed experimentally in in vitro models of elastic conduits that open and close as a result of bridging across the lumen by a fluid lining layer (86, 190, 194,195).

This computational model has been fit to data obtained in mice with acute lung injury (ALI) caused by intratracheal instillation of hydrochloric acid (161). The mice were given a recruitment maneuver (deep lung inflation) to recruit closed lung regions and then ventilated with a normal tidal volume and frequency for several minutes during which time lung elastance was monitored. This procedure was repeated at three different levels of PEEP. The model was able to fit the experimental data closely (Fig. 15) when the values of Po and Pc were normally distributed. While Po and Pc had the same variance, the mean Po was about 4 cmH2O greater than mean Pc. Furthermore, the speeds of opening and closing were hyperbolically distributed, with the speed of opening being about an order of magnitude greater than the speed of closing. The best-fit model parameters for the acid-injured mice were the same as for control (uninjured) mice except that the mean values of Po and Pc were both increased by about 5 cmH2O, which is consistent with an increase in surface tension at the air-liquid interface in the lung (86, 194, 195). An analytical version of this model was also used to account for the derecruitment dynamics of rats injured by saline lavage (6). Again, the model simulated the lung stiffness data accurately, but the rate of rise of elastance following a recruitment maneuver in the rats was much greater than in the acid-injured mice in Figure 15. This shows that recruitment and derecruitment dynamics may be species dependent and are greatly influenced by the type and degree of lung injury.

Open in a separate window

Figure 15

Fits of a computational model of recruitment and derecruitment in the lung (lines) to experimental measurements of respiratory elastance (symbols) in mice with various degrees of acid-induced injury ventilated at three different PEEP levels. Elastance was measured as a function of time following a recruitment maneuver. From Ref. (161) with permission.

In summary, we have reviewed three different approaches to lung stability. The corresponding mechanisms are not exclusive. For example, the local shear and bulk moduli of the parenchyma could be incorporated into the model of dynamic instabilities via avalanches by making the critical opening pressures dependant on the local moduli. Also, the velocity of opening could be taken into account in avalanche dynamics. Alternatively, gas redistribution and possibly avalanche dynamics likely occur during cyclic breathing-like deformation of the injured lung during mechanical ventilation. Furthermore, the bulk and shear moduli of the parenchyma should also influence the speed of opening and closing. Understanding these processes may become critical in determining the nature of ventilator-induced lung injury such as repetitive opening and collapse induced by high shear and normal stresses on the epithelium during mechanical ventilation.

### Elastin, collagen, and mechanics in interstitial diseases

Using scanning electron microscopy, the three-dimensional microstructure of normal lung parenchyma has recently been visualized in extraordinary detail (see also Fig. 4B and Fig. 5B)

Sours: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3929318/

### Similar news:

.

532 533 534 535 536