Abstract
Robust classification and localization of bone fractures are beneficial to avoid misdiagnosis or underdiagnosis. However, state-of-the-art classification methods aim to improve accuracy which lacks reliability, and tackled localization problems in a supervised manner with much-annotated data that leads to high costs. In this paper, we propose a multistage feature map (MSFM) learning network to predict the class of the image and the area of interest without annotated bounded box. MSFM consists of three stages to predict the representation with different objectives and aims to improve the accuracy and reliability of classification. The weakly supervised MSFM model localizes the region of interest (ROI) by taking representation from all the stages supervised by image-level labels only. We also introduced a feature augmentation technique to enforce the model to consider other discriminative regions. End-to-end training of MSFM is performed jointly at all stages. Based on the comprehensive experiments, our approach achieves state-of-the-art results on the standard MURA dataset, which includes the elbow, finger, forearm, humerus, shoulder, wrist, hand, and bone tumor dataset. Code: github.com/MAXNORM8650/MSFM.