Convolutional neural networks have shown great potential for the analysis of medical images. In the neuroimaging field, they have been used to classify a plethora of brain diseases with encouraging results. Nevertheless, to ease the adoption of this method in clinical practice, the networks’ behavior needs to be better understood. In the present study, we successfully detected abnormal patterns in real pathologic brain images via 3D convolutional neural networks trained with simulated data. We provided visualizations of the network activations to support predictions. To produce the simulated data used to train the networks, we altered the intensity of specific brain regions using mean diffusivity maps of healthy subjects, computed from diffusion weighted imaging. Adopting 10-fold cross-validation, the performance of the 3D convolutional neural network was evaluated over a range of intensity increases on a hold-out set of simulated data and real pathologic images of multiple system atrophy patients, a rare neurodegenerative parkinsonian syndrome. Indeed, the regions altered in the simulated dataset were compatible with those involved in multiple system atrophy, i.e. the cerebellum and putamen. We obtained performances competitive with the state of the art for multiple system atrophy classification (best accuracy = 0.88). Visual interpretations of networks’ outcome revealed the targeted regions in training were highlighted in the real pathologic data as well. Our approach allows to discern MSA patients from controls via convolutional neural networks trained with hypothesis-driven simulated images. This work paves the way for further applications, tailoring network architecture and regions of interest.