AI

Generative Active Learning for Label-Efficient Pan-Organ Ultrasound Diagnosis

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 10/11/2026, 08:33:27 UTC.

Briefings

A new generative active learning framework named EvoDiff was developed to improve ultrasound AI by using diffusion models to generate synthetic probe images that guide the selection of real, informative samples for training. Evaluated across six benchmarks with 16,696 images from 15 datasets, EvoDiff matched or exceeded fully supervised performance while requiring significantly fewer labels. The system outperformed active learning baselines and improved foundation model generalization when tested on unseen external cohorts. A reader study involving six radiologists suggested the framework could reduce annotation workload by 24% to 84% at matched performance levels.

Key points

• EvoDiff uses diffusion-based optimization in feature space to generate synthetic probe images that guide the selection of real unlabeled samples for annotation.
• The framework was evaluated across six benchmarks covering 15 datasets and 16,696 images, achieving fully supervised performance with substantially fewer labels than baselines.

Context

Medical ultrasound AI faces challenges including data scarcity, high annotation costs, and limited cross-center generalizability. This research addresses these issues by proposing EvoDiff, a generative active learning approach that integrates diffusion modelling with information-theoretic objectives to optimize sample selection across heterogeneous datasets.

Limitations

Evaluation was conducted across six benchmarks covering 15 datasets and 16,696 images. The framework was tested on convolutional network and transformer-based foundation models. Cross-domain generalization was assessed via external validation on unseen cohorts. A reader study with six radiologists provided qualitative feedback on annotation workflows, but specific clinical outcomes beyond the stated workload reduction range remain unverified.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 10/11/2026, 08:41:58 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigeqlfc3savffnah7egtfscjw3gno4y737pk6zrnx7inbyf4irw4q