Bridging Accuracy and Interpretability: Explainable Deep Learning for Parkinson’s Disease Diagnosis from MRI

This paper presents a deep learning framework for automated Parkinson’s disease (PD) detection from T1-weighted MRI scans using the NTUA dataset. The proposed pipeline combines advanced preprocessing, robustness evaluation, and explainable AI techniques to improve both diagnostic performance and clinical interpretability. Contrast Limited Adaptive Histogram Equalization (CLAHE) was applied to enhance anatomical details, while SMOTE was applied to the deep feature vectors extracted from the training MRI images to address class imbalance. Several state-of-the-art convolutional neural networks were evaluated through six patient-wise train–test splits to assess robustness and generalization across subjects. In addition, a targeted experiment using only axial MRI slices from 50 subjects was conducted to reduce irrelevant anatomical information and emphasize brain regions potentially associated with PD. Among the evaluated models, performance varied substantially across subject-wise splits, highlighting a strong dependency on patient selection. While peak configurations reached high individual metrics, the aggregate subject-wise analysis demonstrated a more conservative baseline. To improve transparency, Grad-CAM visualizations were generated, showing that the models primarily focused on central brain structures relevant to Parkinsonian neurodegeneration, with minimal attention extending to non-diagnostic regions. The results demonstrate that combining adaptive preprocessing, patient-wise evaluation, and explainable deep learning holds promise for MRI-based Parkinson’s disease detection under a preliminary, dataset-specific evaluation, though substantial performance variability remains across different subject selections, rather than demonstrating immediate clinical readiness.