Robust cross-domain generalization using unlabeled target data with source-domain supervision.

Main Methods:

• A target-informed self-supervised pretraining strategy combining masked image modeling (MIM) and contrastive learning was employed to learn unlabeled target-domain representations.
• A confidence-aware infusion head was introduced to adaptively integrate predictions from the pretrained model.
• The approach utilized labeled source data for supervised training and unlabeled target data for self-supervised pretraining, keeping datasets strictly separate.

Main Results:

• The proposed method significantly improved cross-device performance on pediatric POCUS data, achieving over a 6% Dice improvement on the target domain compared to the baseline.
• The AI model demonstrated enhanced generalization capabilities on data acquired from a different ultrasound probe (TeleMED) than the one used for training (Philips Lumify).
• The strategy proved effective in learning structural representations from unlabeled target-domain data, overcoming domain shift challenges.

Conclusions:

• The developed label-efficient and privacy-preserving approach enables robust cross-device ultrasound AI for pediatric fracture assessment.
• This framework offers a practical solution for generalizing AI models without costly retraining or manual annotation on new datasets.
• The methodology can be extended to multi-center studies and federated learning setups for broader AI application in medical imaging.