A Hybrid CNN-Transformer Network for fMRI-Based Feature Encoding in Alzheimer's Disease Classification.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Functional magnetic resonance imaging (fMRI) is crucial for studying brain activity but faces challenges due to high dimensionality and temporal complexity.
• Effective feature representation is essential for accurate analysis and classification tasks in neuroimaging.

Purpose of the Study:

• To develop an end-to-end deep learning network for advanced fMRI feature encoding.
• To validate the network's effectiveness in classifying Alzheimer's disease (AD) using fMRI data.

Main Methods:

• A 3D Convolutional Neural Network (CNN) was used for spatial encoding of fMRI data at each time point.
• A specialized 3D transformer attention block with 3D positional encoding was designed to model spatial features and long-range dependencies.
• A cascaded transformer module integrated spatial features across time points to capture dynamic brain activity changes.

Main Results:

• The proposed deep learning network demonstrated improved feature representation for fMRI data.
• The method significantly enhanced Alzheimer's disease classification performance on two ADNI datasets.
• The approach effectively captured both spatial and temporal characteristics of fMRI data.

Conclusions:

• The developed deep learning network offers a robust solution for automated fMRI feature encoding.
• This method provides a powerful tool for analyzing complex neuroimaging data and improving disease classification.
• The findings highlight the potential of advanced deep learning techniques in neuroscientific research.