Interpretable multimodal MRI fusion in Alzheimer's disease classification using extended parallel multilink joint ICA and 3D ResNet.

Accurate identification of Alzheimer's disease (AD) stages is important for improving clinical understanding and supporting early-stage assessment. In this study, two multimodal MRI fusion strategies were developed to integrate structural MRI, resting-state functional MRI, and diffusion tensor imaging for pairwise binary classification among four groups: normal cognition (NC), subjective cognitive decline (SCD), mild cognitive impairment (MCI), and AD dementia (ADD). The extended parallel multilink joint independent component analysis (Epml-jICA) combined with a support vector machine (SVM) approach (machine learning) and the ensemble 3D ResNet model (deep learning) were evaluated on 664 participants with multimodal MRI. The results demonstrated that the optimal area under the receiver operating curve (AUROC) values for ADD vs. NC and SCD vs. NC were 95.68% and 81.25%, respectively. Furthermore, systematic interpretability analyses using SHAP, Grad-CAM, and anatomical localization of cross-modal important components were conducted, identifying model-associated imaging patterns that were consistent with prior AD-related findings and may provide insights into different disease stages.