An Efficient Approach For Tuberculosis Diagnosis On Chest X-Ray

Deep learning computer-aided detection (CAD) system for diagnosing abnormal findings related to TB disease on chest X-ray images (CXRs) is growing in research. However, the deep learning model's reliability depends on model quality, the generalization, and the large dataset. To deal with this problem, we provide a large TB dataset, namely VB-TB Dataset, and propose an efficient approach for TB chest X-ray diagnosis, including TB classification and area segmentation models. In our dataset, we obtained a total of 246,216 clean labeled images, including 228,827 TB negative images and 17,389 TB positive images with 2,982 TB area images annotated in pixel-level for the segmentation task. This study presents an efficient approach that significantly improves TB classification using Knowledge Distillation (KD) from ensemble models of different state-of-the-art (SOTA) convolution neural networks. Our TB classification model achieved outstanding performance compared with previous models on our VB-TB test set and TBX11K dataset. We also firstly applied an efficient semantic segmentation model based on the light-weight U-Net to localize precisely TB areas on CXRs, and the model also obtained impressive results scored by Dice Similarity Coefficient (DSC) on our TB area segmentation testset.