Diffusion Kernel Attention Network for Brain Disorder Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Functional brain network (FBN) analysis is key for brain disorder classification.
• Current methods lack integration, limiting performance due to sequential processing.
• Transformer models show promise in complex feature relationship modeling.

Purpose of the Study:

• To develop an integrated Transformer-based approach for FBN modeling, analysis, and brain disorder classification using resting-state fMRI (rs-fMRI) data.
• To address challenges of limited training samples and indirect connections in FBN analysis.
• To propose a novel Diffusion Kernel Attention Network.

Main Methods:

• Replaced standard Transformer dot-product attention with kernel attention to reduce parameters and introduce non-linearity.
• Incorporated a diffusion process over kernel attention to model indirect brain region interactions.
• Validated the method on ADHD-200 and ADNI datasets for ADHD and Alzheimer's disease classification.

Main Results:

• The proposed Diffusion Kernel Attention Network significantly reduces parameters, mitigating small-sample issues.
• The model effectively captures complex functional connections and indirect interactions.
• Demonstrated superior performance in classifying ADHD and Alzheimer's disease compared to existing methods.

Conclusions:

• The integrated approach using Diffusion Kernel Attention Network offers a powerful new tool for brain disorder classification.
• Kernel attention and diffusion processes enhance Transformer's capability for FBN analysis.
• This method shows significant potential for advancing neuroimaging-based diagnostics.