Bone Fracture Detection and Localization on MURA Database Using Faster-RCNN

Abstract

Using computer-aided diagnosis systems for helping radiologists and reducing the time of diagnosis is vital. In this paper, Faster-RCNN with three different backbone structures for feature extraction is applied for fracture zone prediction on bone X-rays of the MURA database. We used just three subsets of all seven subsets of the database. These subsets contain X-rays from the humerus, elbow, and forearm. The results of the experiments show that Faster-RCNN with Inception-ResNet-Version-2 as the feature extractor has the best performance. AP of this model on test samples in the best condition of parameters setting reaches 66.82 % for IOU=50%.