X-Ray-Based Tuberculosis Detection Using Portable Edge AI Device in Rural Regions

Tuberculosis (TB) continues to be a major global health challenge, especially in low-resource regions where diagnostic infrastructure is limited. Deep learning models using chest X-rays (CXR) have shown high accuracy for TB detection. However, their reliance on cloud computing restricts deployment in rural and offline settings. In this study, we evaluate three convolutional neural network (CNN) architectures—VGG16, InceptionNet, and ResNet—for automated TB detection and integrate Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance interpretability. Beyond accuracy, we assess the feasibility of deploying these models on a low-cost edge device (Raspberry Pi) and benchmark inference time against a cloud GPU environment. To test robustness, the models are further validated on two independent datasets representing different regions. Results demonstrate that VGG16 achieves the best diagnostic trade-off between TB and normal cases (AUC: 0.98, F1 Score: 0.93), with high accuracy and clinically interpretable heatmaps, while inference on Raspberry Pi remains comparable to cloud-based performance. This work highlights the potential of interpretable edge AI solutions as affordable and accessible tools for TB screening in underserved communities.