Sherrier and Johnson (1987) used a region-based histogram equalization technique to improve the image quality of CXR locally and finally obtain an enhanced image. Zhang D. et al. (2021) used the dynamic histogram enhancement technique (Abin et al., 2022) used different image enhancement techniques such as Brightness Preserving Bi Histogram (BBHE) (Zadbuke, 2012), Equal Area Dualistic Sub-Image Histogram Equalization (DSIHE) (Yao et al., 2016), Recursive Mean Separate Histogram Equalization (RMSHE) (Chen and Ramli, 2003) followed by a Particle swarm optimization (PSO) (Settles, 2005) for further enhancing the CXRs for detecting pneumonia. Soleymanpour et al. (2011) used adaptive contrast equalization for enhancement, morphological operation-based region growing to find lung contour for lung segmentation followed by oriental spatial Gabor filter (Gabor, 1946) for rib suppression. Candemir et al. (2013) used graph cut optimization (Boykov and Funka-Lea, 2006) method to find the lung boundary. Van Ginneken et al. (2006) used three approaches, Active shape model (Cootes et al., 1994), active appearance models (Cootes et al., 2001), pixelwise classification to segment the lung fields in CXRs. Li et al. (2001) used an edge detection-based approach by calculating vertical and horizontal derivatives to find the RoI in CXR. Annangi et al. (2010) used edge detection with an active contour method-based approach for lung segmentation.

Abdullah-Al-Wadud et al. (2007) proposed enhancing the CXR images input for a CNN model for pneumonia detection. Hasegawa et al. (1994) used a shift-invariant CNN-based approach for lung segmentation. Hwang and Park (2017) proposed a Multi-stage training approach to perform segmentation using atrous convolutions. Hurt et al. (2020) used UNet-based (Ronneberger et al., 2015) semantic segmentation for extracting lung field and performed pneumonia classification. Li B. et al. (2019) used the UNet model to segment the lung part, followed by the attention-based CNN for pneumonia classification. Kusakunniran et al. (2021) and Blain et al. (2021) used UNet for lung segmentation for COVID-19 detection. Oh et al. (2020) used the extended fully convolution DenseNet (Jégou et al., 2017) to perform pixel-wise segmentation for lung fields in CXR to improve the classification performance for COVID-19 detection. Subramanian et al. (2019) used UNet based model to segment out the central venous catheters (CVCs) in CXRs. Cao and Zhao (2021) used a UNet-based semantic segmentation model with variational auto-encoder features in the encoder and decoder of UNet with an attention mechanism to perform automatic lung segmentation. Singh et al. (2021) propose an approach based on DeepLabV3+ (Chen et al., 2017b) with dilated convolution for lung field segmentation. Figures 5A, B showcase examples of the preprocessing tasks.

Hong et al. (2009a) proposed an approach to segment the diaphragm from the CXR using a rule-based method. Huo and Zhao (2014) proposed an approach to suppress the clavicle bone in CXR based on the edge detection algorithm. Chandalia and Gupta (2022) proposed a deep learning-based detection model to detect the inputted image as CT or CXR. Jiezhi et al. (2018) proposed a method to determine the quality of the inputted CXR image using deep learning.

From the above literature, pre-deep learning-based approaches require well-defined heuristics to either enhance or segment the lung region. A major focus has been laid on noise removal or contrast enhancement and lung segmentation. However, limited attention has been given to diseased ROIs segmentation. The common datasets used to perform lung segmentation are Montgomery and Shenzhen (Jaeger et al., 2014); however, the number of samples is limited. No dataset is available to focus on local findings.

Katsuragawa et al. (1988) developed an automated approach based on the two-dimensional Fourier transform for detecting and characterizing interstitial lung disorder. The approach uses the textural information for a given CXR as normal or abnormal. Ashizawa et al. (1999) used 16 radiological features from CXR and ten clinical parameters to classify a given CXR as one of the classes among 11 interstitial lung diseases using ANN. A statistically significant improvement was reported over the diagnostic results from the radiologists.

Thian et al. (2021) combined two large publicly available datasets, ChestXray14 (Wang et al., 2017) and MIMICCXR (Johnson et al., 2019), to train a deep learning model for the detection of pneumothorax and assess its generalizability on six external validation CXR sets independent of the training set.

Homayounieh et al. (2021) proposed an approach to assess the ability of AI for nodule detection in CXR. The study included an in-house dataset trained on the deep model, which is pretrained on ChestXray14 (Wang et al., 2017) and ImageNet datasets for 14 class classifications. Lenga et al. (2020) used the existing continual learning approach for the medical domain for CXR-based analysis. Zech et al. (2018) assessed deep models for pneumonia using the training data from different institutions. Figure 5C showcases the classification pipeline using CXR for different lung diseases.

Lyman et al. (2019) proposed a model to differentiate CXR into normal or abnormal. The model is trained to find any abnormality like effusion, emphysema etc., to classify a given CXR as abnormal. Hong et al. (2009b) proposed a method for feature extraction to detect nodules in the CXR to reduce the false positives. Hong and Shen (2008) proposed an approach for automatically segmenting the heart region for nodule detection. Guendel et al. (2020) proposed a deep multitask learning approach to classify CXR for different findings present in it. The proposed approach also performs segmentation along with disease localization simultaneously. Clarke et al. (2022) proposed a computer-assisted diagnostic (CAD) method using wavelet transform-based feature extraction for automatically detecting nodules in the CXRs. Putha et al. (2022) proposed a deep learning-based method to predict the risk of lung cancer associated with the characteristics (size, calcification etc.) of nodules present in the CXR. Doi and Aoyama (2002) proposed a neural network-based approach to detect the presence of nodule and further classify them as benign or malignant. Lei et al. (2021) created a cloud-based platform for lung-based disease detection using CXR. Ting et al. (2021) proposed a transfer learning approach for detecting lung inflammation from a given CXR. Kang et al. (2019) proposed a transfer learning-based approach for predicting lung disease in the CXR image. Qiang et al. (2020) proposed a lung disease classification approach, which extracts the lung mask and enhances the segmented image and CNN-based model for feature extraction and classification. Luojie and Jinhua (2018) proposed a deep learning-based classification for lung disease for 14 different findings. Kai et al. (2019) proposed a deep learning system to classify the lung lesion in a given CXR. Harding et al. (2015) proposed an approach for lung segmentation and bone suppression in a given CXR to improve CAD results.

Researchers have generally developed algorithms for classification using supervised machine learning approaches. Both multilabel and multi-class classification tasks are studied. Due to availability of small sample size datasets with data imbalance, transfer learning is widely used in most research.

The current work in image generation for CXR has focused on alleviating the data deficiency for training deep models. It is observed that the synthetic data generated using GAN-based approaches improve model performance compared to the standard data augmentation methods such as rotation and flip.

Putha et al. (2021) proposed a deep learning-based system for detecting and localizing infectious diseases in CXR alongside using the information from the clinical sample for the same patient. Jinpeng et al. (2020) proposed a deep learning approach for automatic disease localization using CXRs based on weakly-supervised learning.

The current work in CXR-based analysis has focused on detecting the lung part in the given CXR or the disease area in the bounding box. Most of the work have used object detection algorithms such as YOLO, RCNN and its variants (Mask RCNN, Faster RCNN).

This recently explored area requires more attention. In CXR-based analysis, report generation allows a multi-modal learning using CNNs and sequential models. However, the task is challenging as the large text corpus is required with the CXR dataset and only a fewer datasets are available for this task.

Work done so far on model interpretability for CXR-based disease detection is based on post-hoc approaches such as saliency map or CAM analysis. Explainability in AI-based decisions is a must to rely on machine intelligence. Healthcare is a challenging domain, and the life of humans is at risk based on a false positive or false negative. There is a need to incorporate the inbuilt model explainability to handle noisy or adversarial samples, thus improving model robustness for CXR-based systems. Further, challenges occur due to the data imbalance and less model interoperability, as models are usually trained on data from a single hospital. This results in unfair decisions by learning sensitive information from the data. The existing work should encourage more pathways for robust and fair CXR-based systems, which will further increase the chances of deployment of such systems in places with poor healthcare settings.

Govindarajan and Swaminathan (2021) used reaction-diffusion set method for lung segmentation followed by local feature descriptors such as Median Robust Extended Local Binary Patterns (Liu et al., 2016), local binary pattern (Liu et al., 2017) and Gradient Local Ternary Patterns (Ahmed and Hossain, 2013) with Extreme Learning Machine (ELM) and Online Sequential ELM (OSELM) (Liang et al., 2006) classifiers for detecting TB in CXR images using Montgomery dataset. Alfadhli et al. (2017) used speed-up robust features (SURF) (Bay et al., 2008) for feature detection and performed classification using SVM for TB diagnosis. Jaeger et al. (2014) collected different handcrafted features such as histogram of gradients (HOG) (Dalal and Triggs, 2005), the histogram of intensity, magnitude, shape and curvature descriptors, LBP Ojala et al., 1996) as set A for detection. They further used edge, color (fuzzy-color and color layout) based features as Set B for image retrieval. Chandra et al. (2020) used two-level hierarchical features (shape and texture) with SVM for TB classification. Santosh et al. (2016) used thoracic edge map encoding using PHOG (Opelt et al., 2006) for feature extraction followed by multilayer perceptron-based (MLP) based classification of CXR into TB or normal.

Duong et al. (2021) created a dataset of 28,672 images by merging different publicly available datasets (Jaeger et al., 2014; Wang et al., 2017; Chowdhury et al., 2020; Cohen et al., 2020) for three class classification; TB, pneumonia and normal. Authors performed a deep learning-based classification using a pretrained EfficientNet (Tan and Le, 2019) trained on ImageNet (Deng et al., 2009) dataset, pretrained Vision Transformer (ViT) (Dosovitskiy et al., 2020) and finally developed a hybrid between EfficientNet and Vision Transformer. For the proposed hybrid model, the CXR is given input to the pretrained EfficientNet to generate features which are later fed to the ViT and finally, the classification results are obtained. Ayaz et al. (2021) proposed a feature ensemble-based approach for TB detection using Shenzen and Montgomery datasets. The authors used Gabor filter-based handcrafted features and seven different deep learning architectures to generate the deep features. Dasanayaka and Dissanayake (2021) proposed a deep learning-algorithm comprising of data generation using DCGAN (Radford et al., 2015), lung segmentation using UNet (Ronneberger et al., 2015) and transfer learning approach based feature ensemble and classification. The authors used genetic algorithm-based hyperparameter tuning. Msonda et al. (2020) used the deep model-based approach with spatial pyramid pooling and analyzed its effect on TB detection using CXR allowing robustness to the combination of features, thus improving the performance. Sahlol et al. (2020) proposed an Artificial Ecosystem-based Optimization (AEO) (Zhao et al., 2020) on top of the features extracted from a pre-trained network, MobileNet, trained on ImageNet dataset as feature selector. The authors used two publicly available datasets, Shenzen and Pediatric Pneumonia CXR dataset (Kermany et al., 2018b). Rahman et al. (2020b) used a deep learning approach for CXR segmentation and classification into TB or normal. For segmentation, the authors used two deep models, UNet and modified UNet (Azad et al., 2019). Authors also used different existing visualizations techniques such as SmoothGrad (Smilkov et al., 2017), Grad-CAM (Selvaraju et al., 2017), Grad-CAM++ (Chattopadhay et al., 2018), and Score-CAM (Wang H. et al., 2020) for interpreting deep model for making classification decisions. The authors used nine different deep models for CNN-based classification of CXR into TB or normal. Rajaraman and Antani (2020) created three different models for three different lung diseases. First model was trained and tested on RSNA pneumonia (Stein et al., 2018), pediatric pneumonia (Kermany et al., 2018b), and Indiana (McDonald et al., 2005) datasets for pneumonia detection. The second model is trained and tested for TB detection using the Shenzhen dataset. Finally, the first model is finetuned for TB detection to improve model adaption for a new task and reported majority voting results for TB classification. Rajpurkar et al. (2020) collected CXRs from HIV-infected patients from two hospitals in South Africa and developed CheXaid, a deep learning algorithm for the detection of TB to assist clinicians with web-based diagnosis. The proposed model consists of DenseNet121 trained on CheXpert (Irvin et al., 2019) dataset, and outputs six findings (micronodular, nodularity, pleural effusion, cavitation, and ground-glass) with the presence or absence of TB in a given CXR. Zhang et al. (2020) proposed an attention-based CNN model, CBAM, and used channel and spatial attention to generate more focus on the manifestation present in the TB CXR. The authors used different deep models and analyzed the effect of the attention network on detecting TB. Table 3 summarizes the above work for TB detection using CXRs.

Kaijin (2019) proposed a deep learning-based approach for segmentation and pulmonary TB detection in CXR images. Venkata Hari (2022) proposed a deep learning model for detecting TB in chest X-ray images. Chang-soo (2021) proposed an automatic chest X-ray image reader which involves reading data from the imaging device, segments the lung part, followed by gray level co-occurrence matrix-based feature extraction and finally discriminates it as normal, abnormal or TB. Minhwa et al. (2017) proposed a CAD-based system for diagnosing and predicting TB in CXR using deep learning.

In most handcrafted approaches, the texture of CXR is used to define features, followed by any ML classifier. From the above, it is highlighted that the major focus for TB detection is on two datasets; Shenzhen and Montgomery. However, the two datasets contain below 1000 samples even when combined together. This results in poor generalization and needs a pretrained backbone network which is finetuned later. This is why pretrained models trained on the ImageNet dataset are widely used for TB classification from CXRs. Thus, there is a need for large-scale datasets for TB detection with segmentation masks and disease annotations to achieve model generalizability and interpretability.

Okumura et al. (2011) proposed a rule-based model for detecting the region of interests (ROIs) for nodule patterns based on the Fourier transform and an ANN-based approach for other ROIs which were not covered using the power spectrum analysis. The dataset is based on 11 normal and 12 abnormal cases of Pneumoconiosis, where normal cases were selected from an image database of the Japanese Society of Radiological Technology. Abnormal cases were selected randomly from the digital image database. Ledley et al. (1975) demonstrated the significance of textural information present in the CXR to detect the presence of coal work Pneumoconiosis (CWP). Hall et al. (1975) used the textural information present in CXRs and generated features based on spatial and histogram moments for six regions of a given segmented image. Authors performed classification based on maximum likelihood estimation and linear discriminant analysis (LDA). The authors further performed 4 class profusion classification for a given CXR in CWP workers. Yu et al. (2011) used the active shape modeling to segment out the lung from the CXR. The segmented image is divided into six non-overlapping zones as per the ILO guidance. On top of this, six separate SVM classifiers are built on the histogram and co-occurrence-based features generated from each zone. The authors also generated a chest-level classification by integrating the prediction results of the six regions. The experiments are carried out on a dataset of 850 PA CXRs with 600 normal and 250 abnormal cases collected from Shanghai Pulmonary Hospital, China. Murray et al. (2009) proposed based on an amplitude-modulation frequency-modulation (AM-FM) approach to extract the features and used partial least squares for classification. The authors extracted AM-FM features for multiple scales and used a classifier for each scale, later combining results from the individual classifiers. The authors performed the experiments on the CXRs collected from the Miners' Colfax Medical Center and the Grant's Uranium Miners, Raton, New Mexico, for CWP detection. Xu et al. (2010) collected a private dataset of 427 CXR images, consisting of 252 and 175 images for normal and Pneumoconiosis, respectively. The authors performed segmentation using an active shape model followed by dividing the image into six sub-regions. For each subregion, five co-occurrence-based features are extracted. A separate SVM is trained for each subregion, followed by the staging of Pneumoconiosis using a separate SVM.

Yang et al. (2021) proposed a deep learning-based approach for Pneumoconiosis detection. The proposed approach consists of a two-stage pipeline, UNet (Ronneberger et al., 2015) for lung segmentation and pre-trained ResNet34 for feature extraction on the segmented image. The dataset is collected in-house and includes 1,760 CXR images for two classes; normal and Pneumoconiosis. Zhang L. et al. (2021) proposed a deep model for screening and staging Pneumoconiosis by dividing the given CXR into six subregions. This was followed by a CNN-based approach to detect the level of opacity in each subregion, and finally, a 4-class classification was performed to determine normal I, II, and III stages of Pneumoconiosis for a UNet-based segmented image. The results are obtained on the in-house data of 805 and 411 subjects for training and testing, respectively. Devnath et al. (2021) applied a deep transfer learning CheXNet (Rajpurkar et al., 2017) model on a private dataset. The approach is based on the multilevel features extracted from the CheXNet and fed to a different configuration of SVMs. Wang X. et al. (2020) collected a dataset of 1881, including the 923 and 958 samples for Pneumoconiosis and normal, respectively. They used InceptionV3, a deep learning architecture to detect Pneumoconiosis in the given CXR to determine the potential of deep learning for assessing Pneumoconiosis. Wang D. et al. (2020) generated synthetic data for both normal and Pneumoconiosis using CycleGAN (Zhu et al., 2017), followed by a CNN-based classifier. The author proposed a cascaded framework of pixel classifier for lung segmentation, CycleGAN, for generating training images and a CNN-based classifier. Wang et al. (2021) collected a set of in-house 5,424 CXRs, including normal and Pneumoconiosis cases, belonging to 4 different stages (0–3). Authors used ResNet101 (He et al., 2016) for detecting Pneumoconiosis on segmented images from the UNet segmentation model and showed improved results compared to radiologists. Sydney, and Wesley Medical Imaging, Queensland, Australia. Hao et al. (2021) collected data consisting of 706 images from Chongqing CDC, China, with 142 images positive for Pneumoconiosis. Authors trained two deep learning architectures, ResNet34 and DenseNet53 (Huang et al., 2017) for the classification of CXRs into normal or Pneumoconiosis. Table 4 summarizes the above work based on the method of feature extraction.

Sahadevan (2002) proposed an approach to use high-resolution digital CXR images to detect early-stage lung cancer, Pneumoconiosis and pulmonary diseases. Wanli et al. (2021) proposed a deep learning-based approach for Pneumoconoisis detection using lung CXR image.

From the above-cited work, it is clear that there is no publicly available dataset. The current work is done on the in-house datasets with fewer samples. This draws our attention to the fact that the automatic detection of Pneumoconiosis from CXRs requires publicly available datasets for developing robust, generalizable and efficient algorithms.

Shaoliang et al. (2020) proposed a system for pneumonia detection from CXR using deep learning based on transfer learning.

Most of the work is done around (Stein et al., 2018) dataset in multi-class settings. However, there are challenges which need to be addressed other than the dataset challenge, which includes lung segmentation and model interpretability. Transfer learning is widely used to improve generalization for Pneumonia detection on CXRs. Pneumonia is a common manifestation of many lung disorders and is thus required to be detected in multilabel settings.

Rajagopal (2021) used both transfer learning (pre-trained VGG16) and ML (SVM, XGBoost) trained on a deep features-based approach for three-class classification for COVID-19 detection. Jin et al. (2021) used a pre-trained AlexNet to generate the features on CXR images followed by feature selection and classification using SVM.

Chowdhury et al. (2020) proposed a dataset by merging publicly available datasets (Wang et al., 2017; Mooney, 2018; Cohen et al., 2020; ISMIR, 2020; Rahman et al., 2020a; Wang L. et al., 2020) for COVID-19 and used eight pretrained CNN models [MobileNetv2, SqueezeNet, ResNet18, ResNet101, DenseNet201, Inceptionv3, ResNet101, CheXNet (Rajpurkar et al., 2017), and VGG19] for the three class classification; normal, viral pneumonia, and COVID-19 pneumonia. Khan et al. (2020) proposed CoroNet, a transfer learning-based approach using XceptionNet-based approach, trained end-to-end for classification of CXR images into normal, bacterial pneumonia, viral pneumonia, COVID-19 using publicly available datasets. Islam et al. (2020) proposed a CNN-LSTM based architecture for detecting COVID-19 from CXRs for a dataset of 4,575 images. Pham (2021) compared the fine-tuning approach with the recently developed deep architectures for 2-class and 3-class classification problems for COVID-19 detection in CXRs on three publicly available datasets. Al-Rakhami et al. (2021) extracted deep features from pre-trained models and performed classification using RNN. Duran-Lopez et al. (2020) proposed COVID-XNET, for detecting COVID-19 from CXR images based on CNN for binary classification. Gupta et al. (2021) proposed InstaCovNet-19, by stacking different fine-tuned deep models with variable depth as to increase model robustness for COVID-19 classification on CXRs. Punn and Agarwal (2021), Wang N. et al. (2020), Khasawneh et al. (2021), Jain et al. (2021), El Gannour et al. (2020), Panwar et al. (2020b), and Panwar et al. (2020a) used transfer learning based approach for differentiating COVID-19 from viral pneumonia and normal CXRs. Abbas (2021) proposed a CNN-based class decomposition approach, DeTraC, which aims to decompose classes into subclasses and assign new labels independent of each other within the datasets by adding a class decomposition layer and later adding back these subsets to generate final predictions. The authors used the COVID-19 Classification from CXR images on publicly available datasets. Gour and Jain (2020) proposed a stacked CNN-based approach using five different submodules from two different deep models; first fine-tuned VGG16 and second a 30-layered CNN, and the output is combined by logistic regression for three classifications for COVID-19 using CXRs. Malhotra et al. (2022) proposed COMiT-Net, a deep learning-based multitasking approach for COVID-19 detection from CXR, simultaneously performs semantic lung segmentation, and disease localization to improve model interpretability. Pereira et al. (2020) combined both handcrafted and deep learning-based features and performed two-level classification for COVID-19 detection. Rahman T. et al. (2021) compared the effect of different enhancement techniques and lung segmentation on classification tasks based on transfer learning for differentiating CXRs as COVID-19, normal, and Non-COVID. Li et al. (2020) developed COVID-MobileXpert, a knowledge distillation-based approach consisting of three models, one large teacher model, trained on a large CXR dataset and two student models; one finetuned on COVID-19 dataset to discriminate COVID-19 pneumonia from normal CXRs and another a small lightweight model to perform on-device screening for CXR snapshots. Ucar and Korkmaz (2020) proposed Bayes-Squeeznet, based on pretrained SqueezeNet and Bayesian optimization for COVID-19 detection in CXRs. Shi et al. (2021) proposed a knowledge distillation-based attention method with transfer learning for COVID-19 detection from CT and CXRs. Saha et al. (2021) proposed EMCNet, based on extracting deep features from CXRs and training different machine learning classifiers. Mahmud et al. (2020) proposed a CovXNet, based on training a deep model on different resolution CXR data, Stacked CovXNet, and later finetune it on COVID-19 and non-COVID-19 CXR data as a target task. Table 6 summarizes the above work for COVID-19 detection using CXRs.

Shankar et al. (2022) proposed a deep learning-based SVM approach for classifying chest X-rays affected with COVID-19 or normal.

The research is very recent, and papers produced on different datasets are generated either with fewer samples or combining more than one dataset. The CXR data released post-pandemic is collected from multiple centers across the globe. Further, only a fewer works have incorporated inherent model interpretability. To the best of our knowledge, no work has been established for segmentation, report generation, or disease localization and the primary focus is on the classification task.