XMedFuse: Hierarchical Expert-Gated Fusion Network with Cross-Modality Attention for Robust Disease Diagnosis

The integration of multi-modal data in medical artificial intelligence (AI) holds transformative potential for enhancing diagnostic accuracy and clinical decision-making. This paper introduces XMedFuse, a hierarchical deep learning framework that fuses convolutional neural networks (CNNs) for medical imaging with multilayer perceptrons (MLPs) for clinical tabular data through a dynamic Mixture-of-Experts (MoE) architecture with cross-modality attention. Unlike conventional static fusion methods, XMedFuse enables patient-specific expert routing, adaptive modality weighting, and integrated interpretability-addressing critical limitations of single-modality AI systems and existing multimodal approaches. Evaluated on three benchmark datasets-ADNI (Alzheimer's Disease), APTOS 2019 (Diabetic Retinopathy), and HAM10000 (Skin Cancer)-XMedFuse achieved accuracies of 91.3%, 90.8%, and92.2%, respectively, consistently outperforming unimodal baselines ($\mathbf{p}