Published July 31, 2026 | Version v1

Bone Fracture Detection in X-Ray Images Using Faster R-CNN with ResNet-50 Backbone

Description

The most frequent musculoskeletal injuries that we encounter in the emergency and orthopedics are bone fractures. Typically, they are visualized by sorting through X-rays in a hand-processed manner by a trained radiologist and that may result in small and large fractures going undetected due to exhaustion or uneven discerning, particularly when the fractures are small or hair-line. Herein this paper, we implement an automated bone-fracture detection system developed following a two-stage deep-learning application with a Faster-RCNN based on a ResNet-50 backbone. The first stage Faster-RCNN is the Region Proposal Network which initially identifies potential fractured sites and the predictions are later refined by the regressor classifier and bounding-box. Behind it, we trained Mill on ImageNet and soft deterrented it on a varied annotated dataset of X-rays which accounted with numerous varieties of fracture. We even had the model behind a barebones Flask application and thus the clinicians can simply drop pictures and obtain immediate outcomes, with bounding-box overlays and confidence figures. The results of our experiments demonstrate that the system achieves a solid mean Average Precision (mAP) of 87.3per cent and can operate at a fast pace, allowing its use in the clinic in near-real-time. Comparing it to in single to stage detectors such as YOLOv3 and EfficientDet, the two stage architecture produced a better localization of those small, low contraction fractures. This is aimed at assisting radiologists and not to displace them.

 

Files

Copy of Bone Fracture Detection in X-Ray Images Using Faster R-CNN with ResNet-50 Backbone.pdf