Deep learning approaches for preprocessing and analysis of neuroimaging data

In the last decade, human neuroscience has been profoundly impacted by the appearance of large, publicly available databases comprising neuroimaging data of hundreds to thousands of participants. Among other consequences, the advent of ‘big data' in neuroimaging has opened the way to deep learning approaches for both preprocessing and analysis. In this thesis, we focus on a specific neuroimaging modality, magnetic resonance imaging (MRI), and we assess the effectiveness of deep learning methods for preprocessing and the extraction of biomarkers from both structural and functional MRI data. Using several MRI datasets, including both healthy participants and patients suffering from different brain pathologies, we first systematically compare the preprocessing outcomes of a recently developed deep-learning preprocessing pipeline with those of a more traditional method. We then test the performance of deep-learning approaches to achieve robust MRI-based biomarkers predictive of demographics, behavior, and clinical status.