New ultrasound AI model generalizes to new devices without retraining labels
Self-supervised pretraining boosts cross-device fracture detection by 6% Dice.
A team led by Yuyue Zhou at the University of Alberta has proposed a label-efficient method for making ultrasound AI models robust to domain shifts between different scanners. Their work, published on arXiv, tackles the real-world problem of deploying fracture detection AI trained on one device (Philips Lumify) to another (TeleMED portable) without re-annotating target data. The solution combines two self-supervised techniques—masked image modeling (MIM) and contrastive learning—applied to unlabeled target-domain images to learn structural representations. A confidence-aware infusion head then adaptively integrates predictions from the source-trained model and the target-pretrained representations.
Tested on 318 images from 62 pediatric POCUS videos, the method achieved a 6% Dice improvement over baseline in the target domain. Crucially, the source and target datasets were kept strictly separate, preserving data privacy. The authors note this framework is easily extensible to multi-center studies or federated learning setups, offering a practical path to clinical deployment of ultrasound AI across heterogeneous hardware ecosystems.
- Combines masked image modeling & contrastive learning on unlabeled target data to bridge domain shift
- Achieves 6% Dice improvement on target device (TeleMED) vs baseline using only source labels (Philips Lumify)
- Privacy-preserving: source and target data kept strictly separate; no annotation needed for new devices
Why It Matters
Enables scalable deployment of medical AI across different ultrasound machines without costly relabeling or sharing patient data.