Two-stage deep learning pipeline for tuberculosis detection in chest radiographs

Tuberculosis (TB) is a critical public health challenge, particularly in environmentally devastated regions such as Karakalpakstan, Uzbekistan. We propose a two-stage deep learning pipeline for automated TB detection and lesion localization using the TBX11K benchmark dataset. In Stage 1, three CNN backbones — ResNet-18, ResNet-50, and EfficientNet-B0 — classify chest X-ray images into healthy, sick-non-TB, and TB-positive categories. In Stage 2, YOLO11 localizes TB lesion regions via bounding box prediction on TB-positive images. All three backbones exceed 99% accuracy, while YOLO11 achieves a mAP@0.5 of 0.7227 at 1.3 ms inference speed. The lightweight ResNet-18 remains competitive with deeper models, supporting deployment in resource-limited clinical settings.