MAD-Former: A Traceable Interpretability Model for Alzheimer's Disease Recognition Based on Multi-Patch Attention.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) diagnosis relies on integrating structural magnetic resonance imaging (sMRI) with deep learning.
• Existing CNN-based models lack multi-scale analysis and interpretability for AD detection.
• Identifying specific brain regions affected by AD is crucial for accurate diagnosis.

Purpose of the Study:

• To develop a traceable interpretability model for AD recognition using multi-patch attention (MAD-Former).
• To address limitations of single-scale analysis and lack of spatial localization in current deep learning models for AD.
• To enhance the accuracy and interpretability of automated AD diagnosis.

Main Methods:

• Proposed MAD-Former model with two parts: recognition and interpretability.
• Implemented a 3D brain feature extraction network for local features.
• Designed a dual-branch attention structure with varying patch sizes for multi-scale global feature extraction.
• Introduced an attention similarity position loss function.
• Developed a traceable method for 3D ROI space identification via attention selection and receptive field tracing.

Main Results:

• MAD-Former demonstrated outstanding performance on ADNI and OASIS datasets for AD recognition tasks.
• The model achieved reliable interpretability, identifying key brain tissues influencing diagnostic decisions.
• Highlighted the significant role of the Fusiform Gyrus (FuG) in Alzheimer's disease recognition.

Conclusions:

• MAD-Former offers a robust framework for multi-scale feature extraction and interpretable AD diagnosis using sMRI.
• The model's interpretability aids in understanding the neuroanatomical basis of AD detection.
• This approach advances automated AD diagnosis by combining high performance with transparent decision-making.