- Open Access
Segmentation of lung parenchyma in CT images using CNN trained with the clustering algorithm generated dataset
- Mingjie Xu†1,
- Shouliang Qi†1, 2Email authorView ORCID ID profile,
- Yong Yue3,
- Yueyang Teng1,
- Lisheng Xu1,
- Yudong Yao1, 4 and
- Wei Qian1, 5
© The Author(s) 2019
- Received: 23 July 2018
- Accepted: 19 December 2018
- Published: 3 January 2019
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.
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.
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.
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.
- Lung parenchyma
- Convolutional neural network
In recent years, segmentation has known great successes in various medical images analysis tasks including detection of atherosclerotic plaques , pelvic cavity assessment [2, 3], ear image data towards biomechanical researches , skin lesions detection , etc. This has led to its expansion to lung diseases detection [6, 7] and specifically to lung field extraction . 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. . However, it constitutes a very challenging task . 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 .
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. , 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 . 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 , 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 [20–23]. 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 . 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 , 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.  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.  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 . 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 . Song et al.  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.
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.
Subsequently, connected component analysis algorithm based on Max-Tree proposed by Fu et al.  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
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.
The details of the train/validation dataset and the separate dataset
Number of patients
Number of slices
Number of lung parenchyma patches
Number of non-lung parenchyma patches
Total number of patches
4.2040 × 109
4.6196 × 109
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
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.
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
The assessment of patch size through the segmentation characteristics and time consumption
Time consumption for one patient
Characteristics of patch segmentation
64 × 64
Too rough, fast
32 × 32
16 × 16
Including other tissues such as fat, tumor and heart
4 × 4
Exquisite, but computationally expensive
2 × 2
Very exquisite, but computationally expensive
8 × 8
Optimization of the CNN parameters
Performance of the proposed CNNs with different parameters
Channels of normalization
Output of FC
Performance evaluated using cross-validation
Performance evaluated using the separate dataset and manual segmentations
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.
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. . In the experiments conducted by Song et al. , 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  of two convolutional layers, AlexNet  of eight learned layers, and VGG-VD  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 . For instance, Anthimopoulos et al. had proposed one deep CNN with five convolutional layers to do lung pattern classification for interstitial lung diseases . Recently Shin et al.  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. . 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.
Comparison of the proposed method with the traditional methods
DSC (vs observer A) (%)
DSC (vs observer B)
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 .
Pixel-wise or patch-wise segmentation
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 . 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.
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.
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.
The authors would like to thank Mr. Patrice Monkam for his valuable help in the writing of this manuscript.
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.
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).
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.Google Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.MathSciNetView ArticleGoogle Scholar
- Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–77.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- Wang J, Li FQ. Automated segmentation of lungs with severe interstitial lung disease in CT. Med Phys. 2009;36(10):4592–9.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- Grady L. Random walks for image segmentation. IEEE Trans Pattern Anal Mach Intell. 2006;28(11):1768–83.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- Sluimer I, Prokop M, Ginneken BV. Toward automated segmentation of the pathological lung in CT. IEEE Trans Med Imaging. 2005;24(8):1025–38.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.
- 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.View ArticleGoogle Scholar
- LeCun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. IEEE. 1998;86(11):2278–324.View ArticleGoogle Scholar
- 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. 1097–105Google Scholar
- Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. Comput Sci. 2014.Google Scholar
- LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521(7553):436–44.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- Yao J, Bliton J, Summers R. Automatic segmentation and measurement of pleural effusions on CT. IEEE Trans Biomed Eng. 2013;60(7):1834–40.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- 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.View ArticleGoogle Scholar
- Girshick R. Fast R-CNN. Comput Sci. 2015.Google Scholar
- He K, Gkioxari G, Dollar P, et al. Mask R-CNN. In: 2017 IEEE international conference on computer vision (ICCV). IEEE Computer Society. 2017.Google Scholar