Robust Cross-Domain Generalization Using Unlabeled Target Data with Source-Domain Supervision
Researchers have developed novel methods for improving the generalization of AI models in medical imaging across different devices and clinical sites. One approach uses unlabeled target data with source-domain supervision, combining masked image modeling and contrastive learning to learn structural representations without labels, and adapting predictions with a confidence-aware infusion head. This method achieved over 6% Dice improvement on cross-device performance for pediatric wrist fracture assessment using point-of-care ultrasound. Another strategy focuses on domain-agnostic feature modulation for semi-supervised domain generalization, particularly in scenarios where domain labels are unavailable. This technique enhances class-discriminative features while suppressing domain-specific information, leading to more robust representations and improved pseudo-label accuracy. AI
IMPACT These methods offer more robust and label-efficient AI solutions for medical imaging, potentially improving diagnostic accuracy across diverse clinical settings and equipment.