Estimating explainable Alzheimer's disease likelihood map via clinically-guided prototype learning.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Alzheimer's disease (AD) diagnosis is challenging due to its gradual progression and irreversibility.
• Identifying AD biomarkers from structural MRI (sMRI) is complex.
• Current methods lack explainability in AD progression modeling.

Purpose of the Study:

• To propose a novel deep learning approach, XADLiME, for Alzheimer's disease progression modeling using 3D sMRIs.
• To enhance the explainability of AD likelihood maps derived from sMRI scans.
• To develop a method for biomarker identification and downstream task execution with explainable states.

Main Methods:

• Clinically-guided prototype learning to establish topologically-aware prototypes on latent clinical features.
• Development of an AD spectrum manifold and an estimating network to approximate AD likelihood maps.
• Integration of clinical and morphological perspectives for explainable likelihood map generation.

Main Results:

• Successfully modeled Alzheimer's disease progression using a novel deep learning framework (XADLiME).
• Generated explainable AD likelihood maps from 3D sMRI scans.
• Demonstrated the utility of likelihood maps as substitutes for unseen sMRI data in downstream tasks.

Conclusions:

• XADLiME offers a promising approach for explainable Alzheimer's disease progression modeling.
• The method improves biomarker identification and provides clinical/morphological insights.
• This deep learning framework enhances the interpretability of AD diagnosis using neuroimaging data.