Adapting HFMCA to Graph Data: Self-Supervised Learning for Generalizable fMRI Representations

Jakub Frąc, Qiang Li, Guido Van Wingen, Shujian Yu

Keywords: HFMCA, self-supervised learning, graph transformer, fMRI, representation learning

Abstract:

Functional magnetic resonance imaging (fMRI) analysis faces major challenges due to limited data and variability across studies. Existing self-supervised methods from computer vision often rely on positive–negative pairs, which are difficult to define for neuroimaging data. We adapt the Hierarchical Functional Maximal Correlation Algorithm (HFMCA) to graph-structured fMRI, providing a principled framework that measures statistical dependence and enables robust self-supervised pretraining. Across five neuroimaging datasets, our method yields competitive embeddings for multiple classification tasks and transfers effectively to unseen domains.
Code and supplementary material: https://github.com/fr30/mri-eigenencoder

Submission Number: 45