Deep Clustering for FMRI Data Representation

Abstract

This study investigates the use of autoencoder-based dimensionality reduction and deep clustering techniques to analyze functional connectivity (FC) matrices derived from resting-state fMRI data. Avalanche analysis was applied to extract key time points for FC computation, followed by encoding the high-dimensional FC matrices into 16-dimensional latent representations using MLP and CNN autoencoders (AEs) and variational autoencoders (VAEs). Among these models, the MLP AE outperformed others in reconstruction fidelity and variance differentiation in latent variables, enabling effective K-means clustering of distinct brain states. In contrast, VAEs showed lower reconstruction accuracy and limited clustering performance. These results demonstrate the potential of autoencoder-based methods in FC analysis for identifying functional brain states. Future work will explore dynamic FC and hyperparameter optimization for VAEs to enhance clustering outcomes.