BoneScanAI: A Clinically-Explainable Multiclass Fracture Classification Framework Using Hybrid Artificial Intelligence and Uncertainty-Aware Fusion

Bone fractures represent one of the most prevalent musculoskeletal injuries, requiring prompt and accurate diagnosis to reduce complications and improve patient outcomes. Existing automated diagnostic methods are limited by inconsistent radiographic image quality, insufficient fracture subtype granularity, and poor model interpretability—factors that hinder clinical adoption. This paper presents BoneScanAI, a computer-aided diagnosis framework that addresses these limitations through uncertainty-weighted ensemble fusion of classical and deep learning feature streams, enabling fine-grained six-class fracture classification from plain radiographic images. A curated dataset of 1,200 X-ray images was assembled from the MURA, MedPix, and Radiopaedia repositories. Preprocessing involved noise suppression via the Crimmins speckle filter and adaptive contrast enhancement using CLAHE. Handcrafted local descriptors extracted via SIFT and SURF were combined with deep hierarchical features from InceptionV3 and ResNeXt101 through an uncertainty-weighted ensemble mechanism that modulates sub-model contributions based on prediction confidence, preserving classification fidelity under variable imaging conditions. The framework classifies fractures across six distinct morphological subtypes, a granularity that informs treatment pathway selection and is seldom addressed in prior automated fracture detection literature. On 240 test images, the BoneScanAI framework achieved an accuracy of 93.75%, outperforming both traditional classifiers and existing reported fracture classification models. Model interpretability was established through Grad-CAM thermal activation mapping, localizing the fracture regions driving each classification decision. These findings demonstrate that uncertainty-aware ensemble fusion combined with structured multi-class output represents a distinct and practically deployable advancement in radiographic fracture diagnosis.

Hali tarjima qilinmagan