CRAD: Cognitive Aware Feature Refinement with Missing Modalities for Early Alzheimer's Progression Prediction.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate Alzheimer's disease (AD) diagnosis often relies on multiple neuroimaging modalities, but incomplete data is a common challenge.
• Existing multimodal knowledge distillation (KD) methods struggle with transferring reliable knowledge due to a focus on high-level features.

Purpose of the Study:

• To develop a novel framework (CRAD) for early Alzheimer's progression prediction that overcomes the limitations of missing neuroimaging modalities.
• To enable robust cross-modal knowledge transfer, particularly of shallow and intermediate features, for improved diagnostic accuracy.

Main Methods:

• Proposed a Consistency Refinement-driven Multi-level Self-Attention Distillation (CRAD) framework.
• Implemented a multi-level distillation module for progressive cross-modal knowledge transfer.
• Introduced a self-attention distillation module (PF-CMAD) leveraging feature self-similarity for interpretable and efficient distillation.
• Incorporated a consistency-evaluation-driven confidence regularization strategy with adaptive distillation controllers.

Main Results:

• CRAD achieved superior accuracy in AD diagnosis and early progression prediction using only MRI data.
• Demonstrated robust cross-dataset generalization performance.
• The proposed method effectively transfers shallow and intermediate knowledge, overcoming limitations of previous high-level feature distillation.

Conclusions:

• The CRAD framework offers a promising solution for Alzheimer's disease prediction with incomplete multimodal neuroimaging data.
• CRAD enables lightweight, reliable, and interpretable knowledge transfer, enhancing diagnostic capabilities in clinical settings.