Incomplete learning of multi-modal connectome for brain disorder diagnosis via modal-mixup and deep supervision

Recently, the study of multi-modal brain networks has dramatically facilitated the efficiency in brain disorder diagnosis by characterizing multiple types of connectivity of brain networks and their intrinsic complementary information. Despite the promising performance achieved by multi-modal technologies, most existing multi-modal approaches can only learn from samples with complete modalities, which wastes a considerable amount of mono-modal data. Otherwise, most existing data imputation approaches still rely on a large number of samples with complete modalities. In this study, we propose a modal-mixup data imputation method by randomly sampling incomplete samples and synthesizing them into complete data for auxiliary training. Moreover, to mitigate the noise in the complementary information between unpaired modalities in the synthesized data, we introduce a bilateral network with deep supervision for improving and regularizing mono-modal representations with disease-specific information. Experiments on the ADNI dataset demonstrate the superiority of our proposed method for disease classification in terms of different rates of samples with complete modalities.