Current diagnostic and surgical treatments for Parkinson’s disease (PD) rely on precise localization of the subthalamic nucleus (STN), a process that is time-consuming for specialists and could benefit from artificial intelligence-based tools. In this study, we compare the performance of three convolutional neural network (CNN) architectures—VGG16, ResNet50, and U-Net—in classifying T2-weighted magnetic resonance images (MRI) from patients with and without a PD diagnosis. The classification results showed that U-Net achieved the highest performance with an F1-score of 0.91, compared to VGG16 and ResNet50, which reached F1-scores of 0.85 and 0.84, respectively. Despite this, U-Net struggled with precision and recall for the control class, highlighting a challenge in handling class imbalance. Furthermore, we applied the K-means clustering to studies classified as PD-positive to localize the STN. Using the elbow method to determine the optimal number of clusters provided a reasonable approximation for the anatomical location of the STN. However, the clustering process faced limitations due to the sensitivity of the algorithm to image characteristics and the number of clusters. Though a quantitative comparison was not performed, the segmentation results were evaluated qualitatively through visual inspection and confirmed by expert manual annotations. Our results suggest that while CNNs like U-Net can achieve high classification accuracy, handling class imbalance and refining the clustering approach are necessary for further improving the automation of STN localization in PD patients.