Graph Reasoning Module for Alzheimer's Disease Diagnosis: A Plug-and-Play Method

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) causes widespread neuronal damage, affecting brain connectivity.
• Convolutional neural networks (CNNs) are increasingly used for AD detection via structural magnetic resonance imaging (sMRI).
• Existing CNN methods face challenges in effectively integrating spatially distant information crucial for AD diagnosis.

Purpose of the Study:

• To develop a novel Graph Reasoning Module (GRM) to enhance CNN-based AD detection.
• To improve the ability of deep learning models to capture relationships between distinct brain regions for more accurate AD diagnosis.
• To boost the performance of existing AD classification models.

Main Methods:

• Proposed a Graph Reasoning Module (GRM) for direct integration into CNN-based models.
• Designed an Adaptive Graph Transformer (AGT) block for adaptive graph construction from CNN feature maps.
• Utilized a Graph Convolutional Network (GCN) block for graph representation updates and a Feature Map Reconstruction (FMR) block for output generation.

Main Results:

• The GRM integration increased the balanced accuracy of AD classification models by over 4.3%.
• The GRM-embedded model achieved a state-of-the-art balanced accuracy of 86.2%.
• Demonstrated superior performance compared to existing deep learning-based AD diagnosis methods.

Conclusions:

• The proposed GRM effectively simulates relationships between brain regions, enhancing AD detection.
• GRM offers a significant improvement for CNN-based sMRI analysis in Alzheimer's disease diagnosis.
• This approach represents a promising advancement in deep learning for neurodegenerative disease detection.