This paper proposes a Prototype-Guided Deformable Memory Transformer (Proto-MemFormer) model for Parkinson's Disease (PD) MRI classification. In the encoding stage, the model integrates a prototype-guided memory mechanism with a deformable attention structure to dynamically aggregate local morphological features and global semantic information. In the decoding stage, a position-calibrated retrieval module is introduced to enhance cross-sample feature alignment and discriminative representation. Experiments conducted on two public datasets, NTUA-Parkinson and PPMI, demonstrate that the proposed model achieves Accuracy, Precision, Recall, F1-Score, and AUC of 93.45%, 93.72%, 93.21%, 93.46%, and 94.81%, respectively, on the NTUA-Parkinson dataset, outperforming current state-of-the-art deep learning methods. Moreover, in the hyperparameter and training set scaling experiments, the model exhibits performance fluctuations of less than 3%, verifying its stability and robustness under different data and environmental conditions.