Uma abordagem multitarefa para avaliação cérebro-comportamento a partir de ressonância magnética funcional

Abstract

Functional magnetic resonance imaging is used to verify the behavior of brain networks in cognitive disorders. To compose this assessment, behavioral data and standardized clinical tests can be associated. Linear algorithms are used in the literature to find a correlation between image data and a target behavioral variable, the final value of a cognitive test, for example. When there is more than one variable to be predicted, algorithms with multitasking learning can be used. Considering the premise that when tasks are related they can be predicted simultaneously, the present work aims to develop a multitasking approach using graph convolutional networks to answer the following research question: is it possible to predict multiple behavioral variables with multitasking learning and verify brain regions in common considering how to input data from rs-fMRI? Our approach demonstrated competitive performance in relation to the literature on single approaches. In addition, we validated which brain regions had their importance altered between the models, with 53 brain regions with a significant change in the matched population and scores submitted to the Wilcoxon test in the importance attributed by the single and multi models.