CARNFusion: Context-Aware Multimodal Fusion with Meta-Learning for Robust and Interpretable Bone Fracture Diagnosis

Bone fracture detection from radiographic images remains a challenging task due to overlapping anatomical structures, heterogeneous fracture appearances, and imaging inconsistencies across acquisition settings. This study introduces CARNFusion, a context-aware multimodal meta-learning framework that integrates convolutional and transformer-based feature extractors with ensemble reasoning for accurate and interpretable fracture diagnosis. The framework unifies DenseNet201, Vision Transformer (ViT), and a custom Context-Aware Residual Network (CARN) through a cross-attention fusion mechanism (CAFM) and a correlation-aware normalization layer (CARNorm), while a LightGBM-based meta-learner refines decision boundaries and enhances probabilistic calibration. Experiments were conducted on two benchmark datasets: D1 (10-class Bone Break Classification) and D2 (binary Bone Fracture Detection). The proposed model achieved accuracies of 98.76% and 99.63% on D1 and D2, respectively, surpassing both individual and concatenated deep models. Ablation analyses verified the contribution of each module, and robustness evaluations showed less than 1% degradation under Gaussian noise, brightness, and contrast variations. The framework achieved an average inference latency of 1.3 ms per image, supporting real-time diagnostic deployment. Overall, the results demonstrate that context-aware fusion and meta-learning can substantially improve diagnostic precision, stability, and interpretability. CARNFusion provides a scalable and computationally efficient approach for radiographic screening and presents strong potential for implementation in intelligent, clinician-assisted fracture diagnosis systems.