MoRE-XAI: Moment-Based Rigid Registration and Multi-Perspective Explainability for Fracture Localization in X-Ray Imaging

Accurate fracture detection and localization in radiographic imaging remain challenging due to anatomical variability, image misalignment, and the limited interpretability of conventional deep learning models. Existing explainable AI (XAI) methods provide useful insights but often lack the robustness required for clinical deployment. This work introduces MoRE-XAI (Moment-based Rigid-registration Enhanced eXplainable AI), a clinically oriented deep learning framework that integrates adaptive alignment, attention-driven feature extraction, and multi-perspective interpretability. The framework incorporates an adaptive moment-based rigid registration stage to ensure consistent anatomical alignment, an EfficientNetB0 backbone enhanced with a Convolutional Block Attention Module (CBAM) to increase sensitivity to fracture-relevant features, and a multi-task learning strategy that simultaneously optimizes classification and weakly supervised localization. Interpretability is achieved by combining Grad-CAM, SHAP, LIME, and with Monte Carlo Dropout–based confidence weighting, thereby producing reliability-aware explanations. Experimental evaluation on the MURA musculoskeletal radiograph dataset demonstrated an accuracy of 81.49%, outperforming ResNet50 (76.5%) and DenseNet121 (78.9%), with further improvement to 83.99% through ensemble learning. Cross-domain validation on NIH Chest X-rays and BraTS MRI scans confirmed strong generalization across imaging modalities. The resulting explanations consistently highlighted clinically meaningful regions such as fracture zones, cortical edges, and joint boundaries, while uncertainty-weighted heatmaps attenuated spurious activations. These findings establish MoRE-XAI as a reliable and interpretable diagnostic support system, with potential for broader applications in medical imaging where alignment, robustness, and transparent decision-making are essential.