Neural networks (NNs) possess the capability to learn complex data relationships, recognize inherent patterns by emulating human brain functions, and generate predictions based on novel data. We conducted deep learning utilizing an NN to differentiate between Parkinson's disease (PD) and the parkinsonian variant (MSA-P) of multiple system atrophy (MSA). The distinction between PD and MSA-P in the early stages presents significant challenges. Considering the recently reported heterogeneity and random distribution of lesions in MSA, we performed an analysis employing an NN with voxel-based morphometry data from the entire brain as input variables. The NN's accuracy in distinguishing MSA-P from PD demonstrates sufficient practicality for clinical application.