A multi-architecture deep learning ensemble approach for Parkinson’s disease classification from structural brain MRI

Parkinson’s disease (PD) is a progressive neurodegenerative disorder in which early diagnosis remains challenging because structural changes observed on magnetic resonance imaging (MRI) are often subtle. This study presents a deep learning framework for PD classification from structural brain MRI using six complementary architectures: EfficientNetV2-S, DenseNet-169, ConvNeXt-Small, ViT-B/16, Swin-Small, and MaxViT-Tiny. Patient-level StratifiedGroupKFold cross-validation was adopted to prevent data leakage, and data augmentation techniques, including MixUp, CutMix, and Random Erasing, were applied to improve model generalization. Focal loss with label smoothing was used to address class imbalance, while Automatic Mixed Precision (AMP) accelerated model training. Predictions from individual models were combined using simple averaging, weighted averaging, and logistic regression stacking. The proposed stacking ensemble achieved a patient-level AUC of 0.973 (95% CI 0.958−0.986), with an accuracy of 95.8%, sensitivity of 96.1%, and specificity of 95.4%. Statistical analysis demonstrated a significant improvement over the best-performing individual model. Post-hoc temperature scaling reduced the Expected Calibration Error by 66%, improving the agreement between predicted probabilities and observed outcomes. Grad-CAM visualizations identified image regions consistent with previous neuroimaging studies. These observations should be interpreted as qualitative rather than quantitative evidence. Although the proposed framework demonstrated strong performance under patient-level internal validation, further evaluation on independent multi-center datasets is required to assess its generalizability.