Deep Generative Transfer Learning Predicts Conversion To Alzheimer{'}S Disease From Neuroimaging Genomics Data

Alzheimer's disease (AD) represents nowadays the most common form of dementia affecting 46 million people worldwide. Important insights into the AD continuum are currently available thanks to functional and structural magnetic resonance imaging (fMRI/sMRI) techniques which allow to assess brain activity and cortical/subcortical atrophy, respectively, being now established biomarkers of the AD. In this paper we propose an imaging genetics framework which integrates sMRI/fMRI and genetics data with the aim of discriminating not only AD from healthy controls, but also mild cognitive impairment (MCI) subjects that converted to AD (MCIc) from those that remained stable over time (MCInc). This is one of the main challenges in the state-of-the-art owing to the prognostic and treatment impact. To this end, the model was first trained and tested for CN versus AD classification, and then transfer learning was exploited for assessing its performance in differentiating MCIc vs MCInc subjects in an unseen cohort. Experimental results showed that the proposed method allowed reaching competitive performance with respect to methods relying on the complete set of modalities for each subject, while outperforming the state-of-the-art in case of inclusion of subjects with missing views.