Translate this page into:
Artificial intelligence (AI)-driven ensemble model for comprehensive chest X-ray abnormality detection and deployment
For correspondence: Dr Manjula Singh, Division of Delivery Research, Indian Council of Medical Research, Delhi 110 029, Professor, Academy of Scientific and Industrial Research, Ghaziabad, Uttar Pradesh, India e-mail: drmanjulasb@outlook.com; drmanjulasb@gmail.com
-
Received: ,
Accepted: ,
How to cite this article: Abhishek A, Chalga MS, Yadav RM, Agarwal K, Vohra V, Tayade A, et al. Artificial intelligence (AI)-driven ensemble model for comprehensive chest X-ray abnormality detection and deployment. Indian J Med Res. 2026;163:174-81. doi: 10.25259/IJMR_1854_2025.
Abstract
Background and objectives
Chest X-rays (CXR) are widely used for screening of thoracic abnormalities, particularly for tuberculosis (TB) in public health settings. However, the lack of trained radiologists in peripheral areas limits timely interpretation. This study presents the development and validation of DeepCXR v1.1, an artificial intelligence (AI)-powered tool designed to identify radiological chest abnormalities without relying on metadata or clinical inputs, making it ideal for large-scale screening programmes.
Methods
In present multicentric study, AI tool was trained on over 282,000 annotated data points from 54,000 CXR images (36,500 abnormal and 17,500 normal) collected from children and adults from 18 centres across 11 States in India. The tool employs a multi-model ensemble architecture including lung segmentation and lesion-specific models to classify images as normal or abnormal. The tool was validated on multiple datasets, and the final independent validation was done on 13927 CXR images collected prospectively from patients coming to the outpatient clinics of the departments of Medicine and Chest of participating centres.
Results
The tool demonstrated strong generalisability across training and validation datasets, achieving sensitivity of 92.2% [95% confidence interval (CI) 91.6, 92.7] and specificity of 77.4% (95% CI 76.1, 78.6) in a blind prospective validation. Its performance was independently validated by expert committees and health technology assessment panels. Advanced post-processing modules were integrated to enhance detection accuracy, particularly for complex anatomical regions such as the heart and diaphragm.
Interpretation and conclusions
DeepCXR v1.1, an indigenously developed AI tool for detecting abnormalities in chest X-rays offers a scalable, interpretable, and robust solution for augmenting radiological screening and improving early disease detection. The tool’s ability to function offline on basic hardware further supports its use in resource-limited settings.
Keywords
Abnormality detection
Artificial intelligence
Chest X-Ray
Lung segmentation
Tuberculosis
Tuberculosis (TB) remains one of the leading causes of death from a single infectious agent globally.1,2 Though microbiological confirmation remains the gold standard, in its absence diagnosis is often based on a combination of clinical evaluation and radiological findings on chest X-rays (CXR), which remain the most widely available diagnostic modality even in rural and hard-to-reach areas.3 However, while CXRs are available, their interpretation requires trained radiologists, whose availability is highly constrained in many parts of the world,4 leading to diagnostic delays.
Interpretation of CXR image is also time-intensive, susceptible to intra and inter-observer variability up-to 20-30%,5 and often inconsistent, especially under high volume public health programmes.6 Artificial intelligence (AI) tools offer a promising solution of rapid, automated chest abnormalities detection. Deep learning AI models especially those using convolutional neural networks (CNNs),7 when trained on sufficiently large and diverse datasets, annotated specifically for the purpose of training an AI algorithm can deliver high sensitivity and specificity in detecting some conditions such as tuberculosis, pneumonia, lung cancer, and interstitial lung diseases.8–10
Despite these advancements, challenges exist to detect early-stage subtle lesions and distinguishing between morphologically similar disease entities.11 Available models have limited scope restricted to detection of limited TB lesions or few conditions like pneumonia, often excluding paediatric cases, thus limiting real-world applicability.12 Our study aimed to develop a clinically deployable AI model as screening tool for differentiating abnormal and normal X-rays, as per requirements of screening programmes, specifically, India’s National TB Elimination Program (NTEP)13 by enabling scalable, reliable and cost-effective screening by categorising CXRs as normal or abnormal in high TB burden settings.
Methods
This study was initiated under India TB Research Consortium (ITRC), Indian Council of Medical Research (ICMR) in collaboration with Institute of Plasma Research (IPR) and participating sites for data collection.
Study setting and sample size
This multicentric study was conducted between May 2022 and December 2024. The training and validation datasets were collected from 18 sites across 11 States in India after obtaining ethical clearance from participating sites (details of training strategies is provided as Supplementary Material). The sample size was independently determined for tuberculosis and non-tuberculous lung diseases in both paediatric and adult populations, considering the prevalence of tuberculosis and other lung diseases. Sample size for radiological evidence of one particular presentation of tuberculosis was calculated as 1900 X-rays, considering the prevalence of tuberculosis as 5% as per National TB prevalence survey 2019-2021.14 Considering 20 different presentations of tuberculosis, the total sample size was 38,000 X-rays, which was further subdivided into adult (90%) and paediatric (10%) groups as per the prevalence in these groups15 rounded to the nearest thousand. For non-TB diseases, sample size equivalent to that considered for the different TB presentations were considered for the training dataset.
Normal chest X-rays (reported normal by the radiologist) were collected from individuals without any chest complaint and coming for medical examination or pre-anaesthetic check-up. Abnormal chest X-rays were obtained from patients having confirmed disease (TB including various presentation as confirmed by gold standard tests and other lung diseases-based on respective gold/reference standard) having defined lesions as per the disease variant or the disease.
Some of the X-rays were available as archived data with site investigators or were collected from medical record department (MRD) of the respective Institute. These X-rays were obtained when patients visited the outpatients or when hospitalised for diagnosis. Some of the X-rays were collected prospectively from patients visiting the OPD for follow up or with fresh chest complaints with some carrying X-ray images done outside the participating Institutes. All CXRs were included only after the quality of the X-rays was satisfactory. One of the participating sites uploaded few CXRs from field study. These X-rays were taken from handheld X-rays machines.
Most of the X-rays were taken in the hospital from X-ray machines located in the Radiology department of the participating hospitals (Supplementary Table I). All X-ray machines used by sites had AERB (atomic energy regulatory board) approval and used the similar defined parameters for taking chest X-rays.
Training dataset composition
The dataset used to train the AI model comprised of total 54,000 chest X-ray (CXR) images, (36,500 abnormal and 17,500 normal), sourced from multiple Indian states (Supplementary Fig. 1).
The dataset included 9,930 images from children (including neonates) <18 yr and 44,070 from adults. From abnormal images, a total of 282,000 bounding box (BB) annotations were generated, representing the entire spectrum of pulmonary and extrapulmonary disease presentations - both tuberculous and non-tuberculous. With an average width of 131.96 and height of 171.96 pixels, BB size distribution, (Supplementary Fig. 1B), indicated a right-skewed distribution, suggesting a larger number of relatively small BBs, with fewer instances of extremely large BBs. These large BBs indicate, late-stage pleural effusion and structurally distorted cases. Normal images were used to balance the incorrect prediction.
Validation dataset
Comprised of 20972 CXR images (13583 normal and 7389 abnormal) (Supplementary Fig. 1E).
Data pre-processing and standardization
A tool Conv.4.1 was developed for pre-processing naïve CXR images to resize them to standard 850x700 pixels using contrast limited adaptive histogram equalisation (CLAHE) and enhance subtle features like early infiltrates and small nodules, reducing computational load while preserving spatial detail for accurate lesion annotation.
Data annotation
Pre-processed images were imported into label.me annotation tool.16 Text was entered for demarcation of regions of interest and suspected lesions by radiologists/pulmonary physicians as per reference taxonomy using polygon feature. The site radiologists and pulmonary physicians were trained on identification of lesions and annotations. All identified abnormal lesions were categorised under a single ‘abnormal’ category.
Data augmentation
To increase model robustness and reduce overfitting, augmentation techniques were applied ±2 degrees per image (resulting in 2 additional samples) inclusive of horizontal flipping; hue, contrast, and saturation adjustments and random square cropping (scale 0.2 to 1.5).
Model architecture
The tool comprised an ensemble of six deep learning models tailored to specific tasks within CXR interpretation. Table I shows parameters for segmentation and detection model.
| Feature | Model 0 | Model 1 |
|---|---|---|
| Feature extractor | Inception-ResNet-v2 | ResNet-101 |
| Image resizer | Fixed 600x600 | Fixed 600x600 |
| Training steps | 500,000 | 1,800,000 |
| Learning rate | Cosine decay (0.005base) | Cosine decay (0.005 base, lower warmup) |
| Batch size | 4 | 8 |
| Data augmentation | Flip only | Flip, hue, contrast, saturation, crop |
| Instance segmentation | Yes (MaskR-CNN) | No (Faster R-CNN only) |
| Dropout | No | Yes |
Training strategy
Region proposal and classification
This was executed in two stages. In the first stage, the region proposal network (RPN) generated up to 300 region proposals per image. These proposals were filtered using an intersection over union (IoU) threshold of 0.7 to eliminate overlapping regions. The loss functions for this stage included a localisation loss weighted at 2.0 and an objectless loss weighted at 1.0. In second stage, a Mask R-CNN-based predictor performed BB classification and localisation. This stage used fully connected (FC) layers with L2 regularisation and applied a Softmax function for classification. To handle duplicate detections, batch non-maximum suppression (NMS) was applied with an IoU threshold of 0.6. The system allows a maximum of 100 detections per class and up to 300 detections per image.
Development flowchart
The training of AI model ( Fig. 1) was done on CXR with confirmed diagnosis. After scanning all images and annotation files for errors, a total of 36,500 abnormal and 17,500 normal images were identified, ensuring sufficient dataset for model training using preprocessing and augmentation techniques. The dataset was then split into training and testing sets, maintaining an 80-20% ratio.
- Flowchart depicting detailed steps of AI model training process.
The AI model was trained for minimum 500,000 epochs, with the test set used to identify the best-fit model and prevent overfitting ( Table I). Supplementary Figures 2, 3 and 4 depict the custom approach to identify lesions using AI prediction.
Finally, the model was validated using a separate set of true positive and true negative images. The training process was repeated with different hyperparameter configurations until the model achieved an acceptable accuracy.
Training and evaluation losses
Supplementary Figure 5 represents the training and evaluation loss trends over multiple epochs during model training indicating how well the model was learning.
The trend observed suggested that the model generalised well, as both losses decrease overtime. Additionally, applying regularisation techniques like dropout or increasing the dataset size helped enhance generalisation and prevent over-fitting, where the model memorises training data instead of learning general patterns or underfitting i.e model is not learning enough from the data.
Heat-map visualization
The generated heat-map overlay (Supplementary Fig. 6) on CXR highlights regions of interest where the current AI model has detected potential abnormalities. The colour intensity varies, with red and yellow areas indicating highly probable regions with potential pathological findings, while blue and green regions represent lower activity areas. The visualisation effectively demonstrates how the model localises critical areas in the radiograph, aiding in interpretability and clinical decision-making.
Operational consideration
This AI system is operational both online and offline, making it deployable on high-end GPUs as well as basic laptops. Its efficiency, rapid processing, and minimal infrastructure requirements make it particularly suitable for integration into public health screening programs in low-resource settings.
Analysis
The receiver operating characteristic (ROC) curve evaluated the model’s performance by plotting the true positive rate against the false positive rate across thresholds, with the area under the curve (AUC) quantifying overall class discrimination ability ( Fig. 2). The F1 score is the mean of precision and recall, providing a balanced measure of accuracy that emphasises correct identification of positive cases while accounting for both false positives and false negatives.
- Receiver operating characteristic (ROC) curve for determining optimal threshold for evaluating the model’s performance with the area under the curve (AUC) quantifying overall class discrimination ability.
Statistical analysis
The tool’s performance in terms of sensitivity, specificity, along with 95% confidence intervals was calculated by comparison with radiologist’s observation as normal or abnormal using SPSS software (IBM India Ltd., Bengaluru).
Results
The AI model’s performance was evaluated using three distinct test datasets summarised in Table II.
| Model DeepCXR Version 1.1 | Number of images | Sensitivity % (95% CI) | Specificity % (95% CI) | Accuracy (%) | F1 score | Precision (%) |
|---|---|---|---|---|---|---|
| Dataset Case 1:(Controlled internal Dataset) | 3048 | 96.07 | 95.80 | 95.94 | 95.81 | 95.55 |
| Dataset Case 2:(Validation dataset 1) | 20971 | 94.2 (93.7, 94.8) | 88.3 (87.8, 88.9) | 90.4 | 89.80 | 91 |
| Dataset Case 3:(Validation dataset 2) | 13927 | 92.2 (91.6, 92.7) | 77.4 (76.1, 78.6) | 87.4 | 88.46 | 85 |
Case1: Controlled internal data set and curated test set for threshold selection
The first dataset was used to determine the ROC curve ( Fig. 2) and establish an optimal threshold for the base model. The CXR images used included true normal from patients screened for physical fitness (pre-employment fitness, fitness for surgery) and the abnormal from patients with confirmed diseases which were verified by a panel of three experts. A cut-off value of 0.3 was found to be most suitable for distinguishing abnormal images. Additionally, for grouped classification, any value above 0.75 was considered significant. The confusion matrix for this dataset is detailed in Supplementary Table II A.
Case 2: Large-scale validation dataset
The validation dataset consisted of 20,971 images, including 13,583 normal images and 7388 abnormal images verified by a panel of three experts. The model demonstrated strong performance with a sensitivity of 94.2 % and specificity of 88.3% (Supplementary Table II B).
Case 3: Independent validation dataset on prospectively collected data
The AI tool was validated on a larger dataset collected prospectively after obtaining consent from patients coming to the OPD of Medicine Department/Chest clinic of participating centres. It included 13927 CXR images and the analysis of AI results compared to radiologists’ reading was done by an independent statistician which yielded a sensitivity and specificity of 92.2% (95% CI 91.6, 92.7) and 77.4% (95% CI 76.1, 78.6), respectively (Supplementary Table II C). The developing and validation team were blinded to this dataset.
These results highlight the robustness of the AI model in detecting abnormalities across different datasets.
Discussion
Despite growing number of AI tools developed for healthcare, their performance in terms of specificity and diagnostic accuracy has restricted their uptake for public health use including the AI tools developed for screening of chest X-rays for differentiating normal /abnormal and identification of TB. Our AI tool (DeepCXR v1.1) consistently demonstrated sensitivity and specificity of 92.2% and 77.4%; and 94.2% and 88.3% on two independent larger validation datasets of 13,927 and 20,971 CXRs, respectively, as per the WHO Target Product Profile17 for screening better, than any of the currently available AI tools globally.
Our tool’s better performance is due to its training on an extensive diverse dataset of digital/scanned radiographs, in terms of disease spectrum and geographical variations, age related changes, anatomical variants18 and across all age groups ensuring its generalisability for use in real world settings which added to the strength of our tool in comparison to other commercially available tools.8,19 Limitations of other AI tools reported in various studies are comparatively lower specificity, ability to detect fewer common lesions only20,21 possibly due to training on smaller number20 and limited spectrum of presentations, reliance on publicly available datasets,22 or lack of diversity owing to data from single centre.23
In direct comparison with other AI systems as reported in studies, DeepCXR v1.1 not only outperformed them but also addressed limitations of other AI tools as reported by Mosquera et al19 (86.0% sensitivity; 88.0% specificity on small dataset of 1,064 images), and Jin et al8 (88.5% sensitivity; 72.3% specificity on 6,006 images; AUC 0.87) A.A.M. Jasmi et al24 in their work evaluated qXRver 2.1 with 1.3M CXRs, the tool achieved high sensitivity (96.2%) and NPV (99.9%) but low PPV (5.1%) due to low prevalence. In another smaller study by Blake et al25 (993 CXRs), qXRver 4.0 showed near-perfect sensitivity (99.7%) and NPV (99.3%) but modest specificity (67.4%). AI model by Anderson et al,22 reported comparable sensitivity of 90.8% and specificity of 88.7% on a publicly available dataset of 20,000 images only from adults thereby limiting its use only to adult population. Also, it does not classify lesions across tuberculous, non-tuberculous and healed lesions, especially post-TB sequelae which have been included in our study. In addition, our tool also generates a fully automated report like other AI tools26 and can be integrated into clinical workflows to assist radiologists, thus reducing reporting delays.
Our AI model offers several notable strengths. Its high-resolution segmentation of lung zones enhances precision in localising abnormalities, allowing for more accurate detection. By employing model specialisation, the system confidently identifies a diverse range of pathologies, including subtle lesions, frequently overlooked by radiologists. Robust post-processing layers further refine accuracy by filtering-out false positives using anatomical knowledge, such as recognising costochondral junctions, rotated films, and diaphragm variations. Additionally, an image quality assessment module filters out poor-quality scans, ensuring reliable performance even in peripheral settings with varied imaging equipment. Finally, the system incorporates explainable AI through heatmap overlays, providing clinicians with transparent and interpretable evidence of model predictions, which facilitates trust and supports clinical adoption. A key strength of the system lies in its ensemble-based architecture, which integrates six specialised AI models, each optimised to detect specific abnormalities such as pleural effusion, infiltrates, small nodules, cardiomegaly, and TB-related changes. The ensemble design allows the system to overcome inherent limitations of single-model approaches by leveraging complementary strengths of different networks. This enhanced sensitivity to detect abnormalities, reduced false negatives increasing specificity. Additionally, a dedicated segmentation model isolates the lung region, enabling more focused analysis reducing interference from anatomical structures.
The AI tool is designed for flexible deployment across varied settings. In online mode, using an NVIDIA RTX 4090 GPU, the model processes images in ⁓30 sec each. In offline mode, on a standard Intel i3 laptop, inference takes about 50 sec per image. This dual-mode capability ensures that the tool can be effectively used in both urban hospitals and rural, resource-constrained environments without relying on cloud connectivity.
Despite being unique in view of its training on a vast diverse dataset, the tool has some limitations, as performance of tool for detecting abnormality in X-rays with antero-posterior view may be limited. Although trained, AI may not be able to differentiate between active and healed lesions; hence, needs clinical correlation. Sensitivity may be low for very early parenchymal lesions, retrocardiac and supraclavicular lesions. The tool has low confidence in detecting early-stage disease. All these has been added as disclaimer in the report by AI tool.
In addition, we would like to highlight here that the performance of the tool was validated in comparison to Radiologists reporting which itself varied across 18 sites due to known inter-radiologists variations and might have resulted in lower specificity during independent validation as compared to a dataset (case 1 and 2) wherein the radiologists annotations were checked and verified by a group of experts. Therefore, validations against reporting by one radiologist would have been better for comparison.
To conclude, this indigenous AI-based chest abnormality detection tool (DeepCXR v1.1) has demonstrated high accuracy for detecting abnormalities in Chest X-rays and can be used as screening tool for large-scale screening specially in resource constrained settings. Its use for mass screening specially for TB will help in faster identification of abnormal CXRs which can be then subjected to confirmatory tests leading to early diagnosis thus reducing need for confirmatory tests in entire population in high TB endemic areas and reducing transmission by early initiation of treatment, helping National programme to achieve Goal of TB Elimination.
Acknowledgment
Authors acknowledge the following: ICMR-Expert Committee on TB Diagnostics and Monitoring Committee: Dr. V. M. Katoch (Former Secy, DHR & DG, ICMR; Chairman); Dr. D. Behera (Former Dean, PGIMER, Chandigarh; Co-Chair); Dr Chakravarthy Bhagvati (Professor, School of Computer and Information Sciences, University of Hyderabad), Dr. Rohit Sarin (Former Director, NITRD and Technical Advisor NTEP, Govt. of India); Dr. A.P. Dubey (Former Director Prof. and HOD MAMC); Dr. Sarman Singh (Former Director AIIMS Bhopal); Dr. R.M Pandey (Ex-HOD, Biostatistics, AIIMS and Dr. A.S. Paintal, Chair of ICMR); Dr. Camilla Rodriques (Sr. Consultant, Hinduja Hospital, Mumbai); Dr. R.P. Joshi (former DDG-TB, NTEP, Delhi); Dr. Urvashi B. Singh (DDG-TB, CTD, MoHFW); Dr. Ravi. P. Singh (Sr. Scientist E1, CSIR, Pune); Dr. Prabha Desikan (Former Director BMHRC, Bhopal); Dr. Ahilandeshwari Prasad (Prof. Radiology, Dr RML Hospital); Dr. Munish Guleria (Prof. Dept of Radiology, RMLH) and Dr.Tavpritesh Sethi (IIT, Delhi) for their constant support throughout the study. Dr R.M. Pandey (A.S. Paintal Chair, ICMR Hqrs.) for independent evaluation of the performance of the AI tool. Authors wish to thank Dr Rajiv Bahl (Secretary DHR and DG, ICMR) for his technical inputs and support and also wish to acknowledge Director, IPR and Dean R&D and administration for their inputs. Ms Meenakshi Bakshi for guidance on quality management.
Author contributions
AA, AS, MS: Development of the AI tool; MSC: Development of the portal for data collection, study design, data cleaning, writing of the report; RMY, AK, SG: Study design, data collection, data cleaning, study coordination; KA, VV, AT, MVB, PM, HM, PVB, SB, AD, DP, CP, SP, JT, PN, AA, IS, SSM, AG, RR, SSM, AM, CRC, MP, RM, BD, AS, RKM, KC, DC, PKS, JKP: Enrolment of participants in the study, data collection, data cleaning at their respective sites; AT, VV, KA, MP: Annotation of the study as central annotation team; MS: Conceptualised the study, study design, wrote the study protocol, contributed overall coordination, data curation and report writing, had full access to all the data in the study, manuscript writing, edited the article throughout all stages and had final responsibility for the decision to submit for publication. All authors edited the article throughout all stages and approved the final printed version of manuscript.
Financial support and sponsorship
This study was funded by India TB Research Consortium, Indian Council of Medical Research, New Delhi (Grant number 5/8/5/37/ITRC/AI tool-Chest X-rays/2021/ECD-I).
Conflicts of Interest
Indian Council of Medical Research is the sponsor of the study and was involved in conceptualization, study design, coordination, monitoring and convening meetings to ensure compliance to protocol and timely completion, however, had no role in participant enrolment and data acquisition. The authors declare that they have no competing interests; no financial relationships with any organizations that might have an interest in the submitted work; no other relationships or activities that could appear to have influenced the submitted work.
Use of Artificial Intelligence (AI)-Assisted Technology for manuscript preparation
The authors confirm that there was no use of AI-assisted technology for assisting in the writing of the manuscript and no images were manipulated using AI.
References
- World Health Organization. Global tuberculosis report. 2023. Available from: https://www.who.int/publications/i/item/9789240089695, accessed on September 25, 2025.
- Assessing and comparing chest radiograph interpretation in the Department of Internal Medicine at the University of the Witwatersrand Medical School, according to seniority. Afr J Thorac Crit Care Med.. 2023;29 10.7196/AJTCCM.2023.v29i1.265
- [CrossRef] [Google Scholar]
- Training for rural radiology and imaging in Sub-Saharan Africa: Addressing the mismatch between services and population. J Clin Imaging Sci.. 2012;2:37.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Frieden T, ed. Toman’s tuberculosis : case detection, treatment, and monitoring : Questions and answers (2nd ed). Geneva: World Health Organization; 2004.
- Deep learning at chest radiography: Automated classification of pulmonary tuberculosis by using convolutional neural networks. Radiology.. 2017;284:574-82.
- [CrossRef] [PubMed] [Google Scholar]
- Explainable artificial intelligence in deep learning–based detection of aortic elongation on chest X-ray images. Eur Heart J - Digital Health.. 2024;5:524-3.
- [Google Scholar]
- Diagnostic effect of artificial intelligence solution for referable thoracic abnormalities on chest radiography: A multicenter respiratory outpatient diagnostic cohort study. Eur Radiol.. 2022;32:3469-7.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Performance of a deep-learning algorithm for referable thoracic abnormalities on chest radiographs: A multicenter study of a health screening cohort. PLoS One.. 2021;16:e0246472.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Application of artificial intelligence in chest radiograph interpretation: Is the time ripe yet? J Clin Sci. 2024;13:157-8.
- [CrossRef] [Google Scholar]
- CheXclusion: Fairness gaps in deep chest X-ray classifiers. Pac Symp Biocomput.. 2021;26:232-43.
- [CrossRef] [PubMed] [Google Scholar]
- The application of artificial intelligence and radiomics in lung cancer. Precis Clin Med.. 2020;3:214-27.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Central Tuberculosis Division. Ministry of Health and Family Welfare, Government of India. National Strategic Plan for Tuberculosis Elimination 2017–2025. Available from: https://tbcindia.mohfw.gov.in/wp-content/uploads/2023/05/National-Strategic-Plan-2017-25.pdf, accessed on September 3, 2025.
- Central Tuberculosis Division. Ministry of Health and Family Welfare, Government of India. Indian Council of Medical Research, ICMR-National National Institute for Research in Tuberculosis. National TB Prevalence Survey in India (2019 – 2021): Summary report. Available from: https://tbcindia.mohfw.gov.in/wp-content/uploads/2023/05/25032022161020NATBPSReport.pdf, accessed on September 3, 2025.
- Profile of respiratory problems in patients presenting to a referral pulmonary clinic. Lung India.. 2008;25:4-7.
- [CrossRef] [PubMed] [Google Scholar]
- Wada, K. Image polygonal annotation with python (Polygon, Rectangle, Circle, Line, Point and Image-Level Flag Annotation). 2018. Available from: https://github.com/wkentaro/labelme, accessed on September 25, 2025.
- World Health Organization. WHO target product profiles. Available from: https://www.who.int/observatories/global-observatory-on-health-research-and-development/analyses-and-syntheses/target-product-profile/who-target-product-profiles, accessed on September 25, 2025.
- Complete spectrum of X-ray images from all chest conditions crucial for training to develop robust artificial intelligence model for TB triage. Indian J Tuberc 2026 doi: 10.1016/j.ijtb.2026.01.002
- [CrossRef] [Google Scholar]
- Chest x-ray automated triage: A semiologic approach designed for clinical implementation, exploiting different types of labels through a combination of four deep learning architectures. Comput Methods Programs Biomed.. 2021;206:106130.
- [CrossRef] [PubMed] [Google Scholar]
- Development and validation of a deep learning-based automatic detection algorithm for active pulmonary tuberculosis on chest radiographs. Clin Infect Dis.. 2019;69:739-47.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Development and validation of deep learning-based automatic detection algorithm for malignant pulmonary nodules on chest radiographs. Radiology.. 2019;290:218-2.
- [CrossRef] [PubMed] [Google Scholar]
- Deep learning improves physician accuracy in the comprehensive detection of abnormalities on chest X-rays. Sci Rep.. 2024;14:25151.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. Proc IEEE Conf Comput Vis Pattern Recognit. 2017:3462-71.
- [CrossRef] [Google Scholar]
- Post-deployment performance of a deep learning algorithm for normal and abnormal chest X-ray classification: A study at visa screening centers in the United Arab Emirates. Eur J Radiol Open.. 2024;13:100606.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Using artificial intelligence to stratify normal versus abnormal chest X-rays: External validation of a deep learning algorithm at East Kent Hospitals university NHS foundation trust. Diagnostics (Basel).. 2023;13:3408.
- [CrossRef] [PubMed] [PubMed Central] [Google Scholar]
- Automatic report generation for chest X-ray images via adversarial reinforcement learning. IEEE Access.. 2021;9:21236-50.
- [CrossRef] [Google Scholar]