Weakly supervised learning based bone abnormality detection from musculoskeletal x-rays

Accurate localization of abnormalities within X-ray images is of the utmost importance for arriving at the correct diagnosis. Weakly supervised learning (WSL) aims to train deep learning models for object detection and localization using only image-level labels (without using localized annotation). Most existing WSL methods use a class activation map (CAM) to generate a localization map. However, CAM-based methods have been criticized for their lack of robustness. In this work, we present a novel weakly supervised multi-stage (WSMS) learning network for accurate and efficient classification and detection of abnormalities in X-ray images. WSMS trains to localize informative regions in the image with image-level supervision. In the first stage, the WSMS network encodes the image into feature representations and localizes activated regions that contain the detailed structure of the image. The second stage proposes informative regions based on attention maps at different scales, which are used for detecting abnormalities without requiring part annotations. The final stage uses a shared weight encoder to determine if the detected region contains an object of interest. WSMS combines the objective from all stages that potentially increase the robustness. WSMS method achieves an accuracy of 97.9%, Kappa scores of 92.8%, Matthew’s correlation coefficient (MCC) of 92.8%, and AUC of 96.7% for classification on the benchmark datasets and outperforms the state-of-the-art results. WSMS was tested on multiple different datasets to ensure the generalizability and reproducibility of the model. This shows the potential usability of WSMS to significantly advance medical image analysis and improve patient care in healthcare. This method achieve SOTA results without using any localized annotated data. The proposed method also removes the need of highly tedious target abnormality annotation.