FractureNet: A Region-Aware Ensemble with Adaptive Hybrid Augmentation for Multi-Anatomical Bone Fracture Detection

Two outstanding challenges for automated bone fracture detection 
from plain radiographs are (1) limited anatomical region representation in train
ing data, and (2) limited model generalization to clinical-grade image variability 
[4,5]. We propose a new framework, FractureNet, to tackle both challenges sim
ultaneously. Firstly, we propose Adaptive Hybrid Augmentation (AHA), a phys
ics-informed augmentation policy that dynamically applies region-specific trans
formations—beam-hardening simulation, trabecular texture synthesis, and sto
chastic occlusion masking—calibrated to the anatomy of each radiograph. The 
second contribution is a Region-Aware Ensemble (RAE) approach, which in
cludes DenseNet121, ResNet50, and one lightweight MobileNetV2 branch, inte
grated via an anatomy-conditioned, learnable gating network to incorporate local 
knowledge into the main CNN branches. FractureNet achieves 93.7% accuracy 
and 0.961 AUC evaluated on a combination of 51,141 radiographs (Kaggle Bone 
Fracture, Stanford MURA, and RSNA Bone Challenge datasets), outperforming 
single-architecture baselines by 3.2–6.8 percentage points, with the area under 
the ROC curve (AUC) as the headline discrimination metric. We show that it is 
possible to rapidly obtain visualizations of the bone fracture location from both 
Grad-CAM and LIME visualization maps, providing evidence of clinically co
herent localization choices that should aid in improving the trust and auditability 
of such emerging AI methods in emergency radiology workflows.