Stabilizing MRI-based Parkinson’s disease classification across clinical sites using ensemble deep learning

Parkinson’s disease (PD) is a progressive neurodegenerative disorder for which early and reliable diagnosis remains challenging, particularly when artificial intelligence models trained on single-center neuroimaging data fail to generalize across clinical sites. Although deep learning has shown considerable promise for magnetic resonance imaging (MRI)-based PD detection, model instability and performance degradation across heterogeneous imaging sources continue to limit real-world deployment. In this study, we investigate the role of ensemble deep learning in stabilizing cross-site PD classification using T1/T2-weighted brain MRI. Three pre-trained convolutional neural networks (VGG16, MobileNetV2, and DenseNet121) are fine-tuned on a primary clinical dataset collected from Shrimann (Fortis) Hospital, India, and evaluated under a harmonized preprocessing pipeline. To assess generalizability, the trained models are externally validated on an independent multi-center dataset from the Parkinson’s Progression Markers Initiative (PPMI), without any retraining or fine-tuning. A stacked ensemble model is constructed by integrating the probabilistic outputs of the base learners through a meta-learning strategy. Its performance is compared with individual models and a simple averaging ensemble. Experimental results demonstrate that ensemble learning consistently reduces cross-domain performance degradation, achieving classification accuracies of 96% on the primary dataset and 94% on the external PPMI dataset. An ablation study further confirms that the observed performance improvements arise from ensemble integration rather than individual model dominance. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) is employed to provide neuroanatomically interpretable evidence supporting the model’s predictions. These findings highlight ensemble learning as an effective mechanism for improving robustness and cross-site reproducibility in MRI-based PD diagnosis, offering a practical pathway toward clinically reliable AI-assisted neuroimaging systems.