Adaptive structured sparse multiview canonical correlation analysis for multimodal brain imaging association identification

Multimodal brain imaging data can be obtained conveniently through rapidly advancing neuroimaging techniques. These multimodal data, which characterize the brain from distinct perspectives, offer a rare opportunity to comprehensively understand the neuropathology of complex brain disorders. Thus, identifying hidden relationships among multimodal brain imaging data is essential and meaningful. The pairwise correlation between two imaging modalities has been extensively studied. However, the multi-way association among more than three modalities remains unclear and is highly challenging. The difficulty and indeterminacy are largely due to the loss imbalances caused by multiple modalities fusion and the lack of reasonable consideration of the relationship implicated in different brain areas. To address both issues, we propose a structured sparse multiview canonical correlation analysis (SMCCA) with adaptive loss balancing and a novel graph-group penalty. The adaptive loss balancing technique encourages SMCCA to fairly optimize each sub-objective. The graph-group constraint penalizes the brain’s regions of interest (ROIs) hierarchically with different regularizations at different levels. We derive an efficient algorithm and present its convergence. Experimental results on synthetic and real neuroimaging data confirm that, compared with state-of-the-art methods, our method is a better alternative as it identifies higher or comparable correlation coefficients and better canonical weights. Importantly, delivered by the canonical weights, the identified ROIs of each modality show a high correlation to each other and brain disorders, which demonstrates the potential of our method for untangling the intricate relationship among multimodal brain imaging data.