Differentiating atypical parkinsonian syndromes with hyperbolic few-shot contrastive learning.

Differences in iron accumulation patterns have been observed in susceptibility-weighted images across different classes of atypical parkinsonian syndromes (APS). Deep learning methods have shown great potential in automatically detecting these differences. However, the models typically require extensively labeled training datasets, which are costly and pose patient privacy risks. To address the issue of limited training datasets, we propose a novel few-shot learning framework for classifying multiple system atrophy parkinsonian (MSA-P) and progressive supranuclear palsy (PSP) within the APS category using fewer data items. Our method identifies feature areas where iron accumulation patterns occur in classes other than the target classification (MSA-P vs. PSP) and enhances stability by leveraging a superior hyperbolic space embedding technique. Experimental results demonstrate significantly improved performance over conventional methods, as validated by ablation studies and visualizations.