XBFD-Net: A Unified Explainable Deep Learning Framework for Bone Fracture Detection with Dual Attention and Quantitative Localization

Existing deep learning approaches for bone fracture detection from 
plain radiographs address either detection or explainability in isolation, lacking a 
unified, clinically deployable framework. We present XBFD-Net, a novel end
to-end architecture that simultaneously addresses fracture detection, transparent 
localization, and quantitative explainability evaluation — three capabilities not 
previously combined in a single system. Firstly, we introduced a Dual Attention 
Fusion (DAF) module that utilizes the channel attention to suppress irrelevant 
areas of the image and spatial attention to focus on anatomical structures relevant 
to fracture recognition during the feature extraction stage. Unlike previous stud
ies that evaluate explainability only qualitatively, we propose a quantitative lo
calization benchmark based on Intersection over Union (IoU) and Dice coeffi
cient with respect to radiologist-annotated fracture masks. XBFD-Net outper
forms multiple popular methods, including Grad-CAM, Score-CAM, and atten
tion-only baselines by up to 15.2 IoU points in detection accuracy (93.4% recall, 
AUC: 0.96) and localization performance (IoU: 56.8%, Dice: 72.4%) across da
tasets against datasets. Qualitative analysis by the radiologists demonstrated that 
heatmaps delineated cortical disruptions (eg, small hairline fractures) accurately.