In the field of psychoneurology, analysis of neuroimaging data aimed at extracting distinctive patterns of pathologies, such as epilepsy and depression, is well known to represent a challenging problem. As the resolution and acquisition rates of modern medical scanners rise, the need to automatically capture complex spatiotemporal patterns in large imaging arrays suggests using automated approaches to pattern recognition in volumetric images, such as training a classification models using deep learning. On the other hand, with typically scarce training data, the choice of a particular neural network architecture remains an unresolved issue. In this work, we evaluate off-the-shelf building blocks of deep voxelwise neural architectures with the goal of learning robust decision rules in computational psychiatry. To this end, we carry out a series of computational experiments, aiming at the recognition of epilepsy and depression on structural (3D) and functional (4D) MRI data. We discover that our investigated models perform on par with computational approaches known in literature, without the need for sophisticated preprocessing and feature extraction.