A curriculum-guided unified framework for robust unsupervised domain adaptation on multi-cohort Parkinson's disease diagnosis.

BACKGROUND

Current artificial intelligence (AI) models for Parkinson's disease (PD) diagnosis via magnetic resonance imaging (MRI) are significantly impeded by domain shift, and performance often decreases due to heterogeneity in imaging protocols and scanners across hospitals. While existing unsupervised domain adaptation (UDA) methods combining self-training and adversarial training enhance model generalizability, their dependence on pseudo-labels often introduces confirmation bias from noisy predictions. This study aimed to mitigate pseudo-label noise, enhance domain-invariant feature learning and develop a novel UDA framework for robust cross-center PD diagnosis.

METHODS

We propose a curriculum-guided unified UDA (CGU-UDA) framework that integrates self-training and adversarial training. Its core innovation is a feedback loop between adaptive pseudo-label refinement and contextual feature regularization. First, a curriculum learning scheduler dynamically adjusts confidence thresholds per class based on real-time learning progress and progressively filters high-quality pseudo-labels. Second, a consistency constraint module enforces prediction agreement between original target images and their randomly masked image and leverages these refined labels to promote robust, context-aware feature learning. Finally, an adversarial domain discriminator conditioned on a randomized multilinear map is applied to align feature distributions across domains.

RESULTS

Evaluations on two independent multi-cohort PD MRI datasets show that CGU-UDA consistently surpasses leading UDA benchmarks. It achieves an average increase of over 2% in classification accuracy and over 3% in area under the curve (AUC) across varied cross-domain tasks. On Hospital→Parkinson's Progression Markers Initiative (PPMI), accuracy improved from 65.81% of baseline to 67.95%, AUC from 0.6474 to 0.7148; on PPMI→Hospital, accuracy improved from 63.83% to 65.96%, AUC from 0.6324 to 0.6687. Ablation studies confirm that both the dynamic thresholding mechanism and the masked consistency constraint are crucial to this performance gain.

CONCLUSIONS

This work advances robust medical AI model by directly tackling the pseudo-label noise problem in domain adaptation. The CGU-UDA framework demonstrates strong potential for deploying reliable diagnostic models across diverse clinical settings and leading to effective clinical application.