Robust Multi-Scale 3D Crossformer Model for Multimodal Medical Image Analysis using Volumetric MRI and Structured Data

Accurate and robust analysis of volumetric medical images in conjunction with structured clinical data is essential for the early detection of complex diseases. We introduce a novel 3D MRI CrossFormer framework that harnesses hierarchical cross-scale embeddings and dual-range self-attention to extract rich spatial features from volumetric MRI scans. These spatial embeddings are then fused with clinical feature embeddings, derived via a lightweight multilayer perceptron, through an intermediate fusion strategy. In an early-diagnosis case study on Alzheimer's disease using 679 subjects from the ADNI2/3 cohorts, our multimodal architecture achieves 99.3% accuracy and 99.7% AUC, demonstrating exceptional robustness and high precision. Ablation experiments confirm that both the cross-scale embedding layers and the clinical-fusion mechanism are critical for maximizing diagnostic performance. These findings highlight the potential of the 3D MRI CrossFormer with clinical fusion for wide-ranging applications in multimodal medical image analysis. The code is available at https://github.com/InfoLab-SKKU/MML-3DCrossFormer.