Rhamba: Region-Aware Self-Supervised Learning for fMRI Analysis

Rhamba, a novel framework, combines anatomically guided masking with hybrid Attention-Mamba architectures to enhance self-supervised learning in resting-state fMRI analysis. The models were initially trained on the ABIDE dataset, employing region-aligned patch embeddings alongside three masking techniques (Any, Majority, Pure) that vary in spatial specificity. Four different architectural configurations were tested: Mamba-only, Alternate (which interleaves Mamba and Attention), and two hybrid encoder-decoder setups (Attention-Mamba and Mamba-Attention). Subsequently, the pretrained models were fine-tuned for classification tasks using the COBRE and ADHD-200 datasets, focusing on detecting schizophrenia and ADHD. This research highlights the often-overlooked effects of region-aware masking and hybrid sequence modeling in neuroimaging.