Overview

• Published in Radiology on August 19, the retrospective study from Radboud University Medical Center analyzed 41,469 Dutch screening exams from 15,522 women, including 332 screen-detected and 34 interval cancers.
• The workflow uses an entropy-based uncertainty on the AI’s probability of malignancy to decide: confident low-risk exams are finalized by AI, confident high-risk exams prompt recall, and uncertain cases go to radiologists for double reading.
• Under this strategy, radiologists read 61.9% of exams, while recall (23.6 per 1,000) and cancer detection (6.6 per 1,000) matched standard double reading (23.9 and 6.7 per 1,000).
• When the AI was confident, performance improved with AUC 0.96 versus 0.87 overall and sensitivity near double reading (85.4% versus 88.9%), though most exams were labeled uncertain.
• Researchers report younger women with dense breasts were more often classified as uncertain, estimate AI would make 19% of recall decisions, and urge prospective trials and commercial support for uncertainty outputs before clinical use.