On self-supervised multi-modal representation learning: An application to Alzheimer's disease

Introspection of deep supervised predictive models trained on functional and\nstructural brain imaging may uncover novel markers of Alzheimer's disease (AD).\nHowever, supervised training is prone to learning from spurious features\n(shortcut learning) impairing its value in the discovery process. Deep\nunsupervised and, recently, contrastive self-supervised approaches, not biased\nto classification, are better candidates for the task. Their multimodal options\nspecifically offer additional regularization via modality interactions. In this\npaper, we introduce a way to exhaustively consider multimodal architectures for\ncontrastive self-supervised fusion of fMRI and MRI of AD patients and controls.\nWe show that this multimodal fusion results in representations that improve the\nresults of the downstream classification for both modalities. We investigate\nthe fused self-supervised features projected into the brain space and introduce\na numerically stable way to do so.\n