Accurate early diagnosis of Parkinson's disease (PD) is challenging due to symptom overlap with atypical parkinsonian syndromes, such as multiple system atrophy (MSA), underscoring the importance of developing methods for precise differentiation to guide timely therapeutic interventions and enhance patient survival. In this study, a deep learning approach was employed to differentiate between PD and MSA based on MR images and PET/CT images. A total of 153 patients (119 with PD and 34 with MSA) were included and five-fold cross-validation was employed for evaluation. Image analysis in the experiment included co-registration of MRI and PET images, automatically segmenting the putamen and caudate nucleus regions in MRI images using the nnU-Net network. Subsequently, the regions of interest were extracted from the PET images based on the segmented MRI images, and then trained with Convolutional Neural Networks (CNNs) for classification. In the experiment, the nnU-Net achieved a Dice score of 84.92% in the task of putamen and caudate nucleus region segmentation. Furthermore, the CNNs could perfectly distinguish PD from MSA (accuracy = 89.71%, AUC = 0.87). In conclusion, a method for classifying PD and MSA based on MR and PET images was successfully developed, and significant potential was demonstrated in clinical applications.