Deep Multi-Modal Discriminative and Interpretability Network for Alzheimer's Disease Diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Alzheimer's disease (AD) diagnosis benefits from multi-modal data fusion.
• Existing fusion methods often overlook sample structure and biomarker identification.
• High-order feature extraction in deep learning hinders biomarker discovery.

Purpose of the Study:

• To propose a novel deep learning method, DMDIN, for Alzheimer's disease diagnosis.
• To align samples in a discriminative common space for enhanced diagnosis.
• To identify significant brain regions (ROIs) and improve model interpretability.

Main Methods:

• Reconstructing modalities using multilayer perceptron (MLP) for hierarchical representation.
• Embedding structural information via shared self-expression coefficients with diagonal block constraints.
• Utilizing generalized canonical correlation analysis (GCCA) for a discriminative common space.
• Employing knowledge distillation for interpretability and capturing brain region influence.

Main Results:

• The proposed DMDIN method outperforms several state-of-the-art methods in Alzheimer's disease diagnosis.
• The network successfully aligns samples, grouping similar categories and separating dissimilar ones.
• Interpretability is enhanced, allowing for the identification of influential brain regions.

Conclusions:

• DMDIN offers a novel approach to multi-modal fusion for Alzheimer's disease diagnosis.
• The method effectively extracts discriminative information and enhances model interpretability.
• DMDIN shows significant potential for identifying biomarkers and improving diagnostic accuracy in AD.