UniFracNet: A Unified End-to-End Explainable AI Framework with Multi-Source Harmonized Training for Clinical Bone Fracture Detection

The reproducibility crisis in the bone fracture detection literature: het
erogeneity of input-label schemas, non-standardized preprocessing pipelines, and 
dataset-specific splits from most models trained on datasets render between-study 
comparisons meaningless. At the same time, this is the first system to combine 
detection and attention-based feature enhancement, as well as multi-method ex
plainability and clinically relevant decision support in a single feedback-iterative, 
reproducible pipeline suitable for immediate use following a detailed set of 
guidelines. We propose UniFracNet, a unified explainable AI framework that 
seamlessly handles both of these tasks. The first contribution is a systematic 
Multi-Source Harmonized Training (MSHT) pipeline, which entails four hetero
geneous datasets (Kaggle Bone Fracture X-ray (10,580), Stanford MURA 
(40,561), RSNA Pediatric Bone Challenge (~9,000), and the Radiopaedia clinical 
images) through role-specific dataset assignment, binary label harmonization, pa
tient-level stratified splitting, and augmentation after data amalgamation to yield 
a 51,141-image training corpus with zero label leakage. The second one includes 
the first reproducible benchmarking framework for bone fracture detection and 
its standardized evaluation using 5-fold cross-validation of six architectures 
(CNN, VGG16, ResNet-50, DenseNet-121, Vision Transformer, and UniFrac
Net) on the harmonized corpus. Compared to all baselines, UniFracNet with a 
recovery efficiency of 93.4%, sensitivity:91.9%, AUC:0.96, IoU:56.8% and 
Dice:72.4% provides heatmaps for clinical decision support validated by radiol
ogists.