key: cord-0847905-svbpzj21
authors: Gite, Shilpa; Mishra, Abhinav; Kotecha, Ketan
title: Enhanced lung image segmentation using deep learning
date: 2022-01-03
journal: Neural Comput Appl
DOI: 10.1007/s00521-021-06719-8
sha: e1cfd9b5a1c60d172424c2fba896320757ec8e83
doc_id: 847905
cord_uid: svbpzj21

With the advances in technology, assistive medical systems are emerging with rapid growth and helping healthcare professionals. The proactive diagnosis of diseases with artificial intelligence (AI) and its aligned technologies has been an exciting research area in the last decade. Doctors usually detect tuberculosis (TB) by checking the lungs’ X-rays. Classification using deep learning algorithms is successfully able to achieve accuracy almost similar to a doctor in detecting TB. It is found that the probability of detecting TB increases if classification algorithms are implemented on segmented lungs instead of the whole X-ray. The paper’s novelty lies in detailed analysis and discussion of U-Net + + results and implementation of U-Net + + in lung segmentation using X-ray. A thorough comparison of U-Net + + with three other benchmark segmentation architectures and segmentation in diagnosing TB or other pulmonary lung diseases is also made in this paper. To the best of our knowledge, no prior research tried to implement U-Net + + for lung segmentation. Most of the papers did not even use segmentation before classification, which causes data leakage. Very few used segmentations before classification, but they only used U-Net, which U-Net + + can easily replace because accuracy and mean_iou of U-Net + + are greater than U-Net accuracy and mean_iou , discussed in results, which can minimize data leakage. The authors achieved more than 98% lung segmentation accuracy and mean_iou 0.95 using U-Net + + , and the efficacy of such comparative analysis is validated.

Tuberculosis (TB) is a primal disease that has probably affected humans since the dawn of humankind. TB is a contagious disease and caused by the bacteria mycobacterium tuberculosis. In 2019, TB was the cause of death of almost 1.4 million people around the globe. It is among the infectious diseases responsible for the death of millions of people [1] . Early diagnosis of TB is essential to save the patients from getting it fatal and life-threatening. TB can be diagnosed using computed tomography scans, magnetic resonance imaging, X-rays, etc. Analysis of X-ray is one of the main tools for TB screening. Individuals suspected of TB need biological and clinical investigation before confirming TB diagnosis, and the medication is prescribed as per the guideline provided by World Health Organization (WHO). Regular screening is necessary for the early and correct diagnosis of TB. The chest is one of the primary tools because of its sensitivity and interpretation [2] . However, it is inevitable to eliminate the chances of intuitive inconsistencies in diagnosing disease from radiography [3, 4] . Chest X-ray of TB is often misdiagnosed with other diseases because of similar radiologic patterns [5, 6] . Misdiagnosis of TB can lead to wrong medication and worsening health conditions and cause other severe side effects. Hence, there is a need for correct lung diagnosis.

Lower-middle-and low-income countries face a scarcity of trained radiologists, especially in the country's rural areas. In these types of outlooks, large-scale screening of pulmonary TB by analyzing CXR images can be done using a computer-aided diagnosis (CAD) system. Recent advancements in GPUs and computer vision and the availability of large-scale chest X-ray-labeled datasets helped in successful image recognition. Convolutional neural networks can learn highly nonlinear functions and hierarchical visual features from appropriate training data, but acquiring datasets in the medical imaging domain as comprehensive annotation as ImageNet is challenging [7, 8] . It can be seen that countries around the globe are investing a significant part of their annual budget in the healthcare industry. However, it is still unable to fulfill society's aspirations [9] . Furthermore, the lack of health workers in the countries puts a significant workload on existing healthcare workers, resulting in fatigue and health issues [10] . Thus, the deep learning applications in the healthcare sector are getting massive attention in the recent decade.

Solutions based on machine learning (ML) and deep learning (DL) have been suggested for many medical applications, especially in diagnosing a brain tumor, lung nodules, pneumonia, breast cancer, etc. Deep learning, part of machine learning (ML), encourages image classification and segmentation results, hence widely adopted by the research community [11] . The cost associated with the X-ray imaging technique is low, and the abundance of data for deep learning techniques created a favorable condition for computer-aided diagnostic system development. The study shows that classification of lung images after segmentation techniques improves model accuracy [12] . Therefore, the theme of this paper is to make use of four popular segmentation techniques for the lung images. Significant contributions of this paper could be stated as • Review of state-of-the-art techniques in lung segmentation problems. • Implementation and analysis of four primary segmentation techniques, namely FCN, SegNet, U-Net, and U-Net ? ? . • Result analysis of the above-implemented benchmark segmentation architectures and their comparison on different performance measures. • Discussion of these segmentation techniques and their efficacy in TB diagnosis.

The rest of the paper is organized as follows: Sect. 2 states the existing literature review and lung classification techniques. In Sect. 3, the proposed comparison methodology is explained in a stepwise manner. Section 4 presents the results and a discussion based on the results generated by four segmentation techniques. Lastly, Sect. 5 concludes with the findings of this implementation and gives recommendations for future work.

The literature review is divided into two sections, namely segmentation and classification.

Rehman et al. [12] generated lung segments from X-ray images using U-Net with mean_iou of 92.82. Shaoyong Guo et al. [39] proposed a novel automatic segmentation model using radiomics with a combination of handcrafted and automated features. Dice similarity coefficients of 89.42% are achieved on ILD database MedGIFT. Chen Zhou et al. [40] developed an automatic segmentation model by integrating (3D) V-Net and spatial transform network (STN) to segment pulmonary parenchyma in CT images and analyze texture and features from the segmented pulmonary parenchyma regions to assist the radiologist in COVID-19 diagnosis. Mizuho Nishio et al. [41] used U-Net architecture optimized via Bayesian optimization on Japanese and Montgomery and obtained DSC of 0.976 and 0.973 on respective datasets. Ferreira et al. [42] proposed a modified U-Net model for automatic detection of infection caused by COVID-19. Trained and evaluated on the CT database of the actual clinical case from Pedro Ernesto University Hospital of the state of Rio de Janeiro, this model achieved a dice value of 77.1% and an average specificity of 99.76%. Feidao cao [43] improves the traditional architecture of U-Net by introducing variational autoencoder (VAE) in each layer of decoder-encoder to improve the ability of the network to extract the features. The network was tested and trained on NIH and JRST datasets and achieved accuracy and F1 score of 0.9701, 0.9334 and 0.9750, 0.9578, respectively.

Advantages of segmentation:

1. Segmentation of the image is the most important medical imaging process. It extracts the ROI (region of interest) through an automatic process. Segmentation divides the image into an area based on a specific interest, like segmenting body organs/tissue. 2. Implementation of classification neural network algorithms on segmented radiological images can improve the segmentation accuracy significantly. 3. Segmentation can increase the computational cost, but it can significantly decrease the overall cost of disease diagnosis.

Detecting tuberculosis is an arduous job because of discrete manifestations such as cavities, small opacities, large Zak et al. [16] implemented TB classification using pretrained Vgg-16, Vgg-19, ResNet-50, and inception V2 with accuracies 64%, 72%, and 81%, AUC 0.82, 0.76, and 0.87, and sensitivity 0.77, 0.68, and 0.77. Rohan et al. [17] proposed a model comprising three standard architectures: AlexNet, GoogLeNet, and ResNet, with an accuracy of 88.24% and AUC 0.93. Melendez et al. [18] classify tuberculosis using MIL, SVM, and MIL ? AL with pixel-level AUC 0.855, 0.900, and 0.877 and case-level AUC 0.801, 0.878, 0.861. Rahul et al. [19] presented a convolutional neural network model comprising 7 Conv layers and 13 fully connected layers. The optimizer used was Adam, with a validation accuracy of 88.76% from chest X-ray images. Volkov et al. [22] introduced a CNN model with an accuracy of 86.6%. Dao et al. [23] classified the performance of the pre-trained models for the classification of TB using Shenzhen (CHN) and Montgomery (MC). UKloupes et al. [24] worked on the pre-trained model's performance on two datasets (CHN and MC) to distinguish between positive and negative, and the accuracy achieved is 80%.

Four different CNN models (Vgg-16, Vgg-19, ResNet-50, and GoogLeNet) were explored, and the results generated by these models were analyzed (Yaakob et al. [25] ). For TB detection, Ahsan et al. [26] presented a general pre-trained CNN, and accuracies achieved are 81.25% and 80%. Yadav et al. [27] transfer learning technique was used along with a deep learning framework which shows an accuracy of 99.98%. Ardila et al. [28] introduced a deep learning architecture to detect lung cancer and achieved an AUC of 94.4%. Thriach et al. [29] proposed a pre-trained Conv neural network for lung cancer classification with a mean accuracy of 74.43%, mean specificity of 74.96%, and mean sensitivity of 74.68%. Hua et al. [30] used CNN and DBN to classify lung cancer for achieving 82.2% and 73.4% of specificity and sensitivity. Islam et al. [31] presented the architecture, a combination of LSTM and CNN, to classify COVID-19 by analyzing chest X-ray and achieved AUC of 99.9%, specificity of 99.2%, sensitivity 99.3%, and an F1-score of 98.9%.

Based on the above extensive literature review, it is found that plenty of CAD architectures used ML and traditional DL architectures to detect TB and other infectious diseases and achieved accuracy up to 90% and more. In medical applications, robust and flexible algorithms or methods can increase the accuracy of the CAD system to diagnose TB from chest X-ray images and make the system reliable. The utilization of the latest and different architectures or assembling of benchmark algorithms can increase classification accuracy. Usually, the whole chest X-ray is used to diagnose lung disorder using convolutional neural networks. Although lungs are used to detect infectious diseases like TB and pneumonia, CXR also contains other regions of the chest cavity. Therefore, focusing only on lungs from CXR throughout training and classification can increase the accuracy significantly. Thus, to isolate lungs from X-ray images, segmentation techniques are used. So in this study, four benchmark segmentation techniques are explored, and their results are analyzed comprehensively. Based on the extensive literature review, along with detailed overview of the models and their accuracy in diagnosis of the lung related diseases as presented in Table 1 .

This section contains comprehensive information about the dataset, preprocessing techniques, and segmentation models. Later in this section, the details about evaluation metrics like accuracy, dice coefficient, mean_iou, recall, specificity, sensitivity, and precision are discussed and are used to compare four models employed in this study. In this study, the authors explored four broadly used segmentation algorithms in multiple fields, including medical diagnoses, and tried to analyze which type of algorithm works well with limited medical images. The attention net has an additional mechanism that adds more parameters to the model, resulting in increased training time. It requires powerful graphical processing units (GPUs) to train, which is not very cost-effective. These are the few reasons why authors considered U-Net ? ? over attention net.

a. Montgomery County X-ray set:

Montgomery is a labeled dataset consisting of X-ray images with a frontal view. This dataset contains 138 X-ray images; 80 of these chest X-ray images contain no disease and 58 chest X-ray images show infection caused by TB. This dataset has been acquired by the Department of Health and Human Services of Montgomery County, MD, USA. Dataset also contains manually generated lung segment masks of every X-ray image of the dataset and is in DICOM format [36] . b. Shenzhen Hospital X-ray set:

Shenzhen dataset is a labeled dataset consisting of X-ray images with a frontal view. This dataset contains 662 X-ray images; 326 of which are regular chest X-rays and 336 of these X-ray images show the presence of infection caused by TB. The Hospital in Shenzhen, China, has collected this dataset in JPEG format [36] .

First, X-ray images of Montgomery and Shenzhen datasets were converted to PNG format because X-ray images of Montgomery and Shenzhen datasets are in DICOM and JPEG format, so it will be simple to train CNN models. Then, in the next level, standardization of images was performed by resizing because both datasets had different sizes. In this study, 512 9 512 pixel size was considered for FCN, SegNet, U-Net, U-Net ? ? as the input size for different convolutional architecture is different, and normalization of data was done using Z-score normalization using mean and standard deviation.

After preprocessing, the datasets were randomized and divided into 80% training and 20% test data, where 80% of the training data were used to provide experience to the segmentation models regarding lung segments. The remaining 20% of the data were used to evaluate the segmentation models.

There are many other algorithms which are used for medical image segmentation, including DeepLab v1, DeepLab v2, DeepLab v3, DeepLab v3 ? , 3D U-Net, V-Net, Res-U-Net, DenseUNet, H DenseUNet, GANs, SegAN, SCAN, PAN, AsynDGAN, etc. The authors have chosen four segmentation models, namely FCN, SegNet, U-Net, and U-Net ? ?, because of their features, for example, SegNet requires low memory for training and testing, FCN is fast and uses pixel-wise classification to produce segments, U-Net is effective with fewer data, and U-Net ? ? is a modified version of U-Net and thus utilizes attributes such as redesign skip connection and deep supervision to produce perfect segments. FCN is among the first segmentation algorithms, so it is used as a benchmark in this paper for the other three algorithms (SegNet, U-Net, and U-Net ? ?). The authors also tried to track the improvement in segmentation algorithms and their performance in medical image segmentation, especially lung segmentation. There are very few papers that tried to track and analyze the improvement in segmentation algorithms comprehensively. These are among the most used semantic segmentation architectures shown by their Google Scholar citation scores presented in Fig. 1 .

These four segmentation models are called ''modified'' in their architecture diagrams as per the finalized dataset and the problem statement.

Fully convolutional networks use CNNs for a pixelto-pixel transformation, as shown in Fig. 2 . However, unlike CNN, the weight and height of all intermediate layers feature maps are brought back to the original size through convolutional transpose in FCN, allowing localization and skip connections to be implemented to recover the fine spatial information lost downsampling [37] .

FCN model architecture: FCNs use locally connected layers like convolution, upsampling, and pooling. Each layer is a 3-D array of size h*w*d where h and w are spatial dimensions and d represents no channels. Dense layers are avoided in FCN, which means fewer parameters make the network faster and easier to train. The downsampling path of the model is responsible for the interpretation and extraction of the context, and upsampling is used to model for localization. FCNs employ skip connections to recover fine-grained information lost in downsampling. This improvement allows the model to have pixel-wise predictions. b. SegNet SegNet was developed by the University of Cambridge and primarily used for semantic segmentation [35] . This segmentation CNN model incorporates two halves. The first half is an encoder, and the second half is a decoder followed by a pixel-wise classification. The network architecture of the encoder is identical to Vgg-16, and low-resolution encoder feature maps are converted into a full input feature map by the decoder network of the architecture for pixel-wise classification. The pooling indices of max-pooling layers are computed during downsampling to perform nonlinear upsampling, as shown in Fig. 3 .

SegNet Architecture: SegNet architecture can be divided into two halves, followed by a pixel-wise classification layer. c. Encoder network. d. Decoder network. e. Encoder:

It performs convolution with a filter to produce a set of feature maps. It has 13 convolutional layers that are not fully connected, max-pooling layers, and these are used to achieve translation invariance. Combining it with subsampling leads to pixels governing large input feature maps. These methods achieve better classification accuracy and a reduction in the size of the feature maps. This is also responsible for the lossy image presentation with faded boundaries, which are unsuitable for image segmentation. The output should have the exact image resolution as the original image. This is achieved by using upsampling in the decoder. To achieve the exact image resolution in output as the input image, it is essential to store and capture the details of the edges in the encoder feature map before subsampling. SegNet accumulates only the max-pooling indices. f. Decoder:

For each encoder, the corresponding decoder input Neural Computing and Applications feature maps are upsampled by memorized maxpooling indices from the corresponding encoder feature map and convolved with decoder filter banks to produce a dense feature map. The feature maps produced by decoders are of the same size and channels as their encoder inputs. The trainable classifier is fed with the higher-dimensional feature representation present at the final decoder as output. The classifier classifies each pixel and produces a channel image of probability at the output. g. U-Net

The U-Net is a CNN architecture for solving segmentation problems in the biomedical field and other image transformation tasks. U-Net is more successful than other convolutional models in pixelbased image segmentation because it is very effective with limited data. This unique model was developed by Olal Ronneberger et al. [33] , as shown in Fig. 4 . U-Net model architecture: To segment biomedical images, the U-Net architecture has two paths. The first path is a contraction (also called encoder). The encoder captures context via a compact feature map. The encoder is a stock of max-pooling and convolution layers like Vgg-16. The other half of the architecture is a uniform expanding path (also known as a decoder) which is also the second path. The second path did the precise localization using transposed convolution. The encoder section is made of many contraction blocks. The encoder follows the classic architecture of ConvNet. The network uses a repeated implementation of two 3 9 3 convolutions (ReLU) and a 2 9 2 maxpooling operation with stride 2 for contraction. The numbers of feature channels double with every downsampling step. The expansive path contains a feature map's upsampling and 2 9 2 convolution (''up-convolution''), which halves the number of feature channels, concatenation with the corresponding feature map from skip connection and two 3 9 3 convolutions followed by ReLU. At the final layer, there is a 1 9 1 convolution used to map the component feature vector.

In total, the network has 23 convolutional layers. h. U-Net ? ?

The U-Net ? ? was proposed by Zhou et al. [34] . U-Net ? ? is the modified version of U-Net. U-Net ? ? uses the dense block ideas from DenseNet to improve U-Net. U-Net ? ? has three additional features to the original U-Net. i. Redesigned skip pathways. j. Dense skip connections. k. Deep supervision.

U-Net ? ? model architecture: U-Net ? ? architecture can be divided into three parts, as mentioned in the overview that distinguishes U-Net ? ? and U-Net.

1. Redesigned skip pathways: The U-Net ? ? consists of redesigned skip connections, as shown in Fig. 4 . These are used to connect the semantic gap between encoder and decoder. The semantic gap of the feature map between encoder and decoder is reduced by skip connection Conv layers mentioned above. The direct connection of feature maps between encoder and decoder in U-Net results in semantic dissimilar feature maps fusion. In U-Net ? ? , the output from the convolutional of the previous layers is concatenated with the output of the corresponding upsampled output of the low, dense block. This helps bring the feature maps of the encoders closer to the feature maps waiting in the corresponding decoder and helps in optimization quickly. 2. Dense skip connections: The U-Net ? ? has dense skip connections, as shown in Fig. 5 , inspired by the DenseNet and the purpose of the dense skip connection to implement skip pathways between the encoder and decoder. This helps in improving the accuracy of segmentation and gradient flow. In addition, deep skip connection is responsible for accumulating prior feature maps and delivery to the right node due to dense convolution blocks along the skip pathways. This results in the generation of a full resolution feature at multiple semantic levels. 3. Deep supervision: Deep supervision in U-Net ? ? , as shown in red in Fig. 5 , is implemented to adjust the model complexity to balance the speed and performance of the architecture. It is a must for an accurate CNN model to average the output from all segmentation branches.

After discussing the segmentation models above, the proposed methodology diagram (Fig. 6) gives a brief idea about the implementation flow adopted for this research. The dataset of chest X-ray is preprocessed first, and then the data split of 80-20% training and testing, respectively, takes place. Then, after applying four popular segmentation Fig. 4 Modified U-Net architecture [33] Neural Computing and Applications techniques on the image dataset, results are generated, compared, and analyzed in depth.

Evaluation of models focuses on estimation of the performance of the model on unseen data. Thus, evaluating the performance of the neural network models for lung segmentation was done after finishing the training and validation phases, and models were compared using loss, accuracy, intersection over union (IoU), dice, sensitivity, specificity, specificity, recall, and precision. The equations were used to calculate loss, accuracy, IoU, dice, sensitivity, and specificity:

Recall ¼ TP=ðTP þ FNÞ: ð7Þ 

In this study, four neural network architectures of lung segmentation are evaluated on the Montgomery and Shenzhen datasets. This study uses a dataset of 704 images taken from these two datasets to check the model performance.

FCN is trained and validated on the datasets, and then the overall performance is evaluated using the test set. Best performing training and validation results are stated in Table 2 .

Their results are discussed separately: Sect. 4.1 presents the results generated by four segmentation techniques separately, and 4.2 presents the comparison among them.

The performance of the FCN in the study is represented in Figs. 7, 8, 9 , and 10. Accuracy is the way to measure how often an algorithm or architecture classifies positive and negative. The specificity is the measure of true positives identified by the model, and the sensitivity is the measure of true negatives identified by the model. FCN shows low validation dice_coefficient and mean_iou, which is shown in Figs. 22A and 23A , and Fig. 24B shows low accuracy as well. Low values of parameters show that the FCN is not suitable for organ segmentation which is a valuable research finding.

As seen and discussed above, segmented lungs generated by FCN are not satisfactory, which shows that FCN is unsuitable for medical application, which can be convinced from Fig. 11 .

Two standard publicly available datasets such as Montgomery and Shenzhen are used to generate training and testing results. Best training and validation results are stated in Table 3 The response of the SegNet Model is reported in Figs. 12, 13, 14, and 15 . It can be seen from Figs. 12A and 13A that the values of validation mean_iou and dice_coefficient are higher than training mean_iou and dice_coefficient. Still, their values are 0.7914 and 0.6558. These values show that the model is not preferable for medical segmentation, and Fig. 14A and B shows that there is a big difference in values of validation loss, accuracy, and corresponding training values, which shows that the SegNet model is not suitable for a dataset with small size. As presented and discussed above, segmented lungs generated by SegNet are not satisfactory, showing that SegNet is not suitable for medical application, as shown in Fig. 16. 

Originally, U-Net was trained on training data, validated, and evaluated on test datasets. As a result, the U-Net's best performing training and validation parameters are stated in Table 4 .

Dice coefficient is a measure of overlaps between two sets here; these two sets are ground truth masks and predicted masks. Sensitivity is the measure of ground truth. It can be seen from Fig. 17A , B that dice coefficient is increased with every epoch; maximum training dice coefficient and sensitivity are 0.8776, 0.8756. Validation dice coefficient and sensitivity are 0.9217 and 0.8904. It can be seen from the sensitivity results that U-Net does image localization by predicting the image pixel by pixel.

Loss predicts error in the model and accurately predicts how well the model is performed; as can be seen in Fig. 18A , B, loss decreases and accuracy increases with every epoch, and Fig. 19A , B shows that both mean_iou and specificity increase with every epoch and maximum 

To train the U-Net ? ?, training dataset was utilized, and performance of the validation set is enhanced by using hyperparameters tuning, and the test dataset is used to evaluate the overall performance. Best performing training and validation results are stated in Table 5 . Figures 22A and 23A show that the validation dice coefficient and validation mean_iou are more significant than the training dice coefficient and training mean_iou. Validation dice coefficient and mean_iou are 0.9796, 0.9598. It shows that lung segmentation generated by U-Net ? ? and ground truth are almost similar. Figures 22B and 23B show that specificity increases steadily with every epoch. Validation values of specificity and sensitivity are greater than training values. Value received during validation are 0.9932 and 0.9753, which shows that the redesigned skip pathway between the encoder and decoder Subpaths bridges the semantic path between the encoder and decoder which increased the optimization of the U-Net??. Figures 24 and 25 represented that validation accuracy and precision are near one and loss declines with every epoch, which indicates that U-Net ? ? is pretty accurate in generating segmented lungs. Figure 26 represents the segmented lung segments generated by U-Net ? ? , ground truth, and the difference between ground truth and lung generated by U-Net ? ? . It can be noticed from Fig. 24a that U-Net ? ? generated pretty accurate lung segments.

In the section, the best training results of the four models are presented with various performance measures. Tables 6  and 7 show the training and validation results, respectively, and gives a fair idea about the superiority of U-Net ? ? implementation results.

The performance of the models used in this study is compared in this section. The authors trained four segmentation models on Shenzhen and Montgomery datasets and generated the results for all models: accuracy, precision, sensitivity, specificity, recall, precision, mean_iou, and dice_coefficient. Tables 6 and 7 compare the deep learning models trained and evaluated on the datasets. As segmented images play a vital role in the perfect diagnosis of the disease, Table 6 presents the performance for image segmentation. The algorithm that scores the best results in this study is U-Net ? ? . It can create segmented images with dice_coefficient 0.9796, mean_iou 0.9598, and accuracy 0.9874. The U-Net also scores acceptable results in this study and can segment lungs of dice_coefficient 0.9217, mean_iou 0.8572, and accuracy 0.9555, but SegNet and score are not satisfactory. All models in this study are based on encoder followed by decoder-type architecture, but U-Net ? ? performance has shown its best results because of redesigned skip pathways, dense skip connection, and deep supervision. So U-Net ? ? is the best performing model for chest X-ray images.

Segmentation is an important step to reduce the chance of data leakage and forces the classification architecture to focus only on essential areas and helps improve classification accuracy. The application of the segmentation technique has proven to be very helpful in the real world. The existing literature review section papers implemented machine learning and deep learning techniques for lung segmentation and got encouraging results. However, to the best of our knowledge, not even a single paper discussed U-Net ? ? and compared other segmentation techniques for lung segmentation. In this study, we studied four benchmark neural network architectures: U-Net, FCN, SegNet, and U-Net ? ?, and the performance of these architectures is thoroughly explored and studied in this paper. The results generated by FCN, U-Net, SegNet, and U-Net ? ? are evaluated on Shenzhen and Montgomery datasets. Comparison between the results of four architectures shows that the U-Net ? ? surpasses other architectures by a considerable margin and achieved 98% accuracy because of its state-of-the-art architecture. FCN did not achieve satisfactory results with 78% and hence was not encouraged to explore further studies based on image segmentation. The future work could be implementing other respiratory problems [e.g., chronic obstructive pulmonary diseases (COPD), pneumonia, etc.] using chest X-rays. Advanced feature extraction techniques with machine learning algorithms and the ensemble model localization scheme can be used to further downstream analysis, detect lung abnormality, and visualize explainable artificial intelligence (XAI) Grad-CAM.

Data availability It is confirmed by the authors that data supporting this research finding are present within the article, and the publicly available datasets used in this study are Montgomery County X-ray Set and Shenzhen Hospital X-ray Set [36] . 

Global Tuberculosis Report (2019) WHO. World Health Organization

Chest radiography in tuberculosis detection: summary of current WHO recommendations and guidance on programmatic approaches. World Health Organization

Error and discrepancy in radiology: Inevitable or avoidable?

Perceptual and interpretive error in diagnostic radiologycauses and potential solutions

The role and performance of chest X-ray for the diagnosis of tuberculosis: a cost-effectiveness analysis in

Chest radiograph abnormalities associated with tuberculosis: reproducibility and yield of active cases

Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics, and transfer learning

Understanding the mechanisms of deep transfer learning for medical images

An empirical study on the determinants of health care expenses in emerging economies

The health of the healthcare workers

Imagenet classification with deep convolutional neural networks

Reliable tuberculosis detection using chest x-ray with deep learning, segmentation, and visualization

Coronavirus: comparing COVID-19, SARS and MERS in the eyes of AI

Automatic detection of coronavirus disease (COVID-19) using X-ray images and deep convolutional neural networks

Deep convolutional neural network with transfer learning for detecting pneumonia on chest X-rays

Classification of lung diseases using deep learning models

Automated TB (2019) classification using an ensemble of deep architectures

Multiple-instance learning for computer-aided detection of tuberculosis

Deeplearning: a potential method for tuberculosis detection using chest radiography

Tb detection in chest radiograph using deep learning architecture

Computer-aided tuberculosis detection from chest X-ray images with convolutional neural networks

Efficient deep network architectures for fast chest X-ray tuberculosis screening and visualization

Deep learning models for tuberculosis detection from chest X-ray images

Pre-trained convolutional neural networks as feature extractors for tuberculosis detection

Detection of pulmonary tuberculosis manifestation in chest X-rays using different convolutional neural network (CNN) models

Application of a convolutional neural network using transfer learning for tuberculosis detection

Using deep learning to classify X-ray images of potential tuberculosis patients

End-to-end lung cancer screening with deep three-dimensional learning on low-dose chest computed tomography

Automatic lung cancer prediction from chest X-ray images using the deep learning approach

Computer-aided classification of lung nodules on computed tomography images via deep learning technique

A combined deep CNN-LSTM network for the detection of novel coronavirus (COVID-19) using X-ray images

Labeled optical coherence tomography (oct) and chest x-ray images for classification

U-Net: Convolutional networks for biomedical image segmentation

UNet??: a nested U-Net architecture for medical image segmentation. In: Stoyanov D (ed) Deep learning in medical image analysis and multimodal learning for clinical decision support

SegNet: a deep convolutional encoder-decoder architecture for image segmentation

Two public chest X-ray datasets for computer-aided screening of pulmonary diseases

Fully convolutional networks for semantic segmentation

Tuberculosis diagnostics and localization in chest X-rays via deep learning models

Automatic lung segmentation based on texture and deep features of HRCT images with interstitial lung disease

Lung segmentation and automatic detection of COVID-19 using radiomic features from chest CT images

CovidXray-Net: optimizing data augmentation and CNN hyperparameters for improved COVID-19 detection from CXR

Segmentation and quantification of COVID-19 infections in CT using pulmonary vessels extraction and deep learning

Automatic lung segmentation algorithm on chest X-ray images based on fusion variational auto-encoder and three-terminal attention mechanism

Conflicts of interest The authors declare no conflict of interest.