An Effective Tuberculosis Detection System Based on Improved Faster R-CNN with RoI Align Method

Tuberculosis(TB) is a serious public health threat in the world. Detecting and treating TB in its early stages can significantly improve the survival rate of patients and serve as the most effective approach for TB prevention and treatment. Using deep learning models to diagnose TB is highly accurate and efficient, making it a competitive option for early diagnosis. We built an improved Faster R-CNN model, which can classify TB X-ray images and detect TB lesions with bounding boxes. Our model has been trained using the large TB dataset TBX11K, which contains 11,200 X-ray images and provides the bounding box annotation information in json files. Our model uses region proposal network to generate anchor boxes, and determines the features in each anchor belonging to the object or background. In the next step, we extract features from boxes of different sizes to ensure the length of output results is equal. Compared with the original Faster R-CNN, we replace region of interest(RoI) pooling with RoI align to avoid quantization problems. Our system can precisely capture and classify disease symptoms in X-ray images with an accuracy of over 90%, and this study contributes to the research of computer-aided TB diagnosis.