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

Functional magnetic resonance imaging (fMRI) is widely used to capture brain functional activities, but its high-dimensional nature and temporal complexity pose significant challenges for feature representation. To address these issues, this study proposes an end-to-end deep learning network for fMRI feature encoding and validates its effectiveness through Alzheimer's disease (AD) classification task. The proposed network first employs a 3D CNN to perform spatial encoding of fMRI at each time point, extracting local spatial patterns of brain images. Then, a transformer attention block specifically designed for 3D MRI is designed, which enhances spatial feature modeling through 3D positional encoding and captures long-range dependencies between brain regions. Finally, a cascaded transformer module is constructed to integrate spatial features across different time points, modeling the dynamic changes in brain activity. Experimental results on two fMRI datasets from ADNI database demonstrate that the proposed method improves feature representation and AD classification performance. By effectively capturing both the spatial and temporal characteristics of fMRI data, this approach provides a robust solution for automated fMRI feature encoding. The code is available at https://github.com/Yanteng32/CTF_FMRI.