Investigator(s)
| Submitter |
Ugyen Choden Gyalpozhing College of Information Technology, Royal University of Bhutan Ugyen Choden 
|
| Principal Investigator |
Chong Tsung Wen : Senior Consultant /Asst Professor Department: Urology Institution: Singapore General Hospital, Singapore Chong Tsung Wen
|
| Co-Investigator(s) |
Yonten Jamtsho Gyalpozhing College of Information Technology, Royal University of Bhutan Yonten Jamtsho
|
| Co-Investigator(s) |
Jigme Yoezer Medical Officer/Clinical Informatics Lead, Digital Health and Innovation Unit (DHIU) Department: CRRH, Gelephu Referral Hospital Institution: Ministry of Health Jigme Yoezer
|
| Co-Investigator(s) |
Charlene Liew Jin Yee Senior Consultant Radiologist/Assistant Professor Department: Radiology Institution: Changi General Hospital, Singapore Charlene Liew Jin Yee
|
| Co-Investigator(s) |
Lim Lee Jean Consultant Department: Urology Institution: Singapore General Hospital, Singapore Lim Lee Jean
|
Title(s) and abstract
| Scientific title | AI-powered Classification of Bhutanese CXR images |
| Public title | AI powered classification of Bhutanese-specific chest X-ray images (CXR): a collaboration between SingHealth Duke-NUS, the Digital Health and Innovation Unit and Gyalpozhing College of Information Technology, Bhutan |
| Background | Bhutan is a small Himalayan country with a population of about 800,000 and a geographically dispersed healthcare system structured into Basic Health Units, district/cluster hospitals, and three Regional Referral Hospitals (RRHs), including the national referral hospital, Jigme Dorji Wangchuk National Referral Hospital (JDWNRH) in Thimphu. A major challenge in the health system is the severe shortage of specialist doctors, particularly radiologists, with only nine in the entire country. This shortage places a significant diagnostic burden on junior doctors working in district hospitals, who frequently interpret chest X-rays (CXRs) without specialist support. CXRs are the most commonly used imaging modality in Bhutan, with over 77,000 performed annually. They are essential for diagnosing respiratory conditions such as pneumonia, pulmonary tuberculosis, and lung cancer, all of which remain significant and increasing public health concerns. Tuberculosis and acute respiratory infections are among the leading causes of morbidity in the country. The reliance on junior doctors for CXR interpretation contributes to diagnostic delays, variability in interpretation, and potential misdiagnosis. Although radiologists provide confirmation, their limited numbers mean that many CXRs are not reviewed promptly. This challenge is further compounded by increasing healthcare workload and workforce attrition. Artificial intelligence (AI), particularly deep learning applied to medical imaging, has demonstrated strong potential to improve diagnostic accuracy and efficiency. However, most existing AI models are trained on non-local datasets and may not generalise well to Bhutan’s population, imaging protocols, and disease patterns. In addition, commercial AI systems are often costly and difficult to sustain in low-resource settings. This project addresses these challenges by developing a Bhutan-specific AI-based CXR classification system to support clinical decision-making, improve diagnostic accuracy, and strengthen healthcare delivery. |
| Objectives | The overarching objective of the study is to develop an AI-powered chest X-ray (CXR) classification system tailored to the Bhutanese population to support the diagnosis of respiratory diseases and assist junior doctors in clinical decision-making. The study is based on the hypothesis that AI models trained on Bhutanese CXR data can improve diagnostic accuracy and reduce time to referral for specialist review. Specific Objectives CXR Annotation and Dataset Development To retrieve, de-identify, and curate Bhutanese CXR images from national repositories. Expert radiologists will classify images into “normal” and “abnormal” categories, including key respiratory conditions such as pneumonia, tuberculosis, pleural effusion, and lung cancer. This will establish the first structured Bhutan-specific CXR dataset. AI Model Development and Validation To develop deep learning models using Convolutional Neural Networks (CNNs) and transfer learning architectures such as ResNet, VGGNet, and DenseNet. Models will be trained for binary classification of normal versus abnormal CXRs and evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis. Pilot Clinical Implementation and Evaluation To deploy the AI system as a clinical decision-support tool in selected district hospitals in the Gelephu Mindfulness City (GMC) region. The study will assess diagnostic performance against standard practice and evaluate the impact on diagnostic accuracy, time to referral, and workflow efficiency. User Acceptance and System Evaluation To assess usability and acceptance among junior doctors through structured surveys and qualitative feedback, and identify barriers to adoption for system improvement. Policy and Scale-up Readiness To generate evidence to support national-scale deployment through the Ministry of Health and the Digital Health and Innovation Unit (DHIU). |
| Study Methods | This study uses a multi-phase, mixed-methods design combining retrospective dataset development, deep learning model training, and prospective clinical evaluation. Phase 1: Data Acquisition and Annotation Chest X-ray (CXR) images will be retrieved from Bhutan’s national imaging repository and fully de-identified by a trusted third party under the Digital Health and Innovation Unit (DHIU). Images will be distributed to radiologists across Regional Referral Hospitals for independent review. Each image will be classified as “normal” or “abnormal,” with abnormal cases including pneumonia, tuberculosis, pleural effusion, and lung cancer. Discrepancies will be resolved by senior radiologist consensus to ensure high-quality ground truth labels. Phase 2: Data Preprocessing Images will be standardised through resizing, normalisation, noise reduction, and data augmentation (rotation, flipping, and scaling). The dataset will be split into training, validation, and test sets to support robust model development and evaluation. Phase 3: AI Model Development Deep learning models will be developed using frameworks such as Keras, PyTorch, and OpenCV. Convolutional Neural Networks (CNNs) and transfer learning architectures (ResNet, VGGNet, DenseNet) will be trained to classify CXRs into binary outcomes (normal vs abnormal). Phase 4: Model Evaluation Model performance will be assessed using accuracy, precision, recall, F1-score, and confusion matrix analysis. Hyperparameter tuning and regularisation techniques will be applied to improve generalisability and reduce overfitting. Phase 5: Pilot Implementation The AI system will be deployed as a decision-support tool in selected hospitals in Gelephu Mindfulness City (GMC). A silent phase will first evaluate performance in parallel with clinical practice, followed by integration into workflow. Outcomes will be compared against standard radiologist interpretation. Phase 6: Statistical Analysis Diagnostic performance will be analysed using McNemar tests and proportion comparisons, while time-to-diagnosis differences will be assessed using |
| Expected outcomes and use of results | The project is expected to deliver significant clinical, technological, and policy-level outcomes for Bhutan’s healthcare system. Clinical Outcomes The primary expected outcome is improved diagnostic accuracy of chest X-ray interpretation in district hospitals. AI-assisted interpretation is expected to reduce misdiagnosis and improve early detection of respiratory diseases such as pneumonia, tuberculosis, pleural effusion, and lung cancer. An earlier and more accurate diagnosis is likely to improve patient outcomes, reduce complications, and decrease hospital admissions. The system is also expected to reduce time-to-diagnosis by supporting junior doctors in real-time interpretation and enabling faster referral to radiologists when required, particularly in remote settings with limited specialist access. Health System Efficiency The AI tool is expected to reduce the workload of the limited number of radiologists in Bhutan by triaging normal cases and prioritising abnormal findings. This will improve efficiency in radiology services, allow specialists to focus on complex cases, and reduce unnecessary referrals, thereby streamlining patient care pathways. Economic and Social Impact Earlier detection and treatment of respiratory diseases may reduce healthcare costs associated with prolonged hospitalisation and advanced disease management. Patients may also benefit from reduced loss of productivity and earlier return to work, contributing to broader socioeconomic benefits. Research and Capacity Building The study will generate the first Bhutan-specific chest X-ray dataset and establish foundational capacity in artificial intelligence for medical imaging in the country. This will support future research, training, and development of advanced AI applications, including disease-specific detection, segmentation, and explainable AI systems. Policy and National Scale-up Findings will inform the Ministry of Health on the safe and effective integration of AI into clinical workflows. Results will guide national policy on AI governance, deployment, and scale-up across all distri |
| Keywords | Artificial Intelligence, Chest X-ray, Deep Learning, CNN, Medical Imaging, Bhutan, Radiology, Diagnostic Support, Healthcare AI, Dataset Development |
Research Details
| Student research | No |
| Start Date | 01-Apr-2026 |
| End Date | 31-Mar-2029 |
| Key Implementing Institution | Central Regional Referral Hospital |
| Multi-country research | No |
| Nationwide research | Yes, with randomly selected geographical areas |
| Bhutan | |
| Research Domain(s) | Communicable Disease Research |
| Research field(s) | Tuberculosis |
| Involves human subjects | Yes |
| Intervention Evaluation Research | |
| Data Collection | Secondary data |
| Proposal reviewed by other Committee | Approved |