Deep Learning Meets Explainable AI: A Robust Framework for X-Ray Fracture Detection

Abstract

Bone fractures are commonly diagnosed using X-ray imaging and often pose challenges for accurate detection due to variability in appearance and image quality. This study proposes a three-stage deep learning framework to address these challenges, comprising preprocessing, binary classification, and localization. The preprocessing stage enhances image quality and removes extraneous parts of the X-ray images to isolate the primary regions of interest. The Swin Transformer V2 model achieves a classification accuracy of 95.41 %, while YOLOv11 attains a mAP50 of 0.606 for fracture localization. To ensure explainability, Gradient-weighted Class Activation Mapping (Grad-CAM) is utilized, visualizing the outputs from the last layer before the classifier and highlighting the regions most relevant to the model's decisions. This robust framework not only achieves high accuracy but also offers transparency, making it suitable for clinical applications in fracture detection.

Ähnliche Arbeiten