Client-Adaptive Vision-Language Gatekeeper Cuts Annotation Time by 30% in Federated Breast Density AI

A client-adaptive vision-language gatekeeper reduced radiologist annotation time by ~30% (1.14 hours) in a breast-density federated learning task, retaining 84.4% of in-distribution images and achieving artifact-discrimination AUC 0.85.

• In a scanner-partitioned cohort of 222,700 training and 65,891 test mammographic images, PromptGate withheld 65.6% of artifact-containing images while retaining 84.4% of in-distribution (ID) images, resulting in a ~30% gross annotation time reduction (3.82 to 2.68 hours) in an enriched reader study.
• Queue purity improved from 88.5% without filtering to 92.5%, with similar downstream density accuracy; reader agreement with the manual artifact reference was 93.6% (κ=0.85), but exact agreement with clinical-record density labels was only 60.9% (κ=0.48).
• The single-reader, retrospective proof-of-concept used an artifact-enriched cohort; prospective multireader and multi-institutional evaluation is needed before deployment.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.