AI chest X-ray tool for Tuberculosis screening in low-resource settings

Tuberculosis remains a major global health problem, especially in resource-constrained, high-burden settings where access to expert radiology is limited. To address this gap, a team led by John Onyango Agumba developed and evaluated a deployable deep learning system aimed at screening for Tuberculosis from chest radiographs. The project set out not only to build an algorithm that can distinguish Normal from Tuberculosis chest X-rays, but also to provide interpretable visual explanations of its decisions and to make the tool usable where internet and computing resources are scarce. The researchers focused on creating a model that could run offline on both desktop and mobile devices, so that clinicians and health workers in remote settings could use it as an assistive screening and decision support tool. A key feature of the work was explainability: the system integrates Gradient-weighted Class Activation Mapping (Grad-CAM) so users can see which parts of an image influenced the model’s prediction, helping clinicians evaluate and trust the automated output while maintaining practical deployment options for constrained environments.

The study used publicly available chest X-ray datasets labeled as Normal or Tuberculosis. The core classifier was a DenseNet121-based transfer learning model trained with stratified training, validation, and test splits to preserve class balance. Training included data augmentation and class weighting to address dataset limitations, and model performance was assessed using accuracy, precision, recall, F1 score, receiver operating characteristic (ROC) curve, and area under the ROC curve (AUC). For interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize the regions that most influenced model predictions; these visualizations showed that the model tended to focus on anatomically relevant lung regions, particularly the upper and mid-lung fields in Tuberculosis cases. To enable practical use, the trained model was converted to TensorFlow Lite and packaged into a Windows desktop application and a Flutter-based mobile application for offline inference and visualization. Deployment testing confirmed consistent prediction outputs and Grad-CAM visualizations across both platforms, and the model demonstrated strong classification performance on the independent test dataset with high accuracy and AUC values.

The combination of accurate classification and visual explanations makes this system promising as an artificial intelligence-assisted screening tool in radiology workflows. Grad-CAM visualizations offer a straightforward way for clinicians to see why the model made a particular call, which supports interpretability and could help integrate automated screening into clinical decision-making. Offline deployment via TensorFlow Lite in both a Windows desktop application and a Flutter-based mobile application demonstrates feasibility for use in remote clinics, mobile screening campaigns, and other resource-limited environments where internet access and powerful servers are unavailable. By packaging the model for practical, local use and preserving explainability, the system aims to augment existing diagnostic capacity rather than replace human judgment, offering a scalable option to help find cases earlier and direct limited resources more effectively in high-burden settings.