Sunday, November 29 | 10:30 a.m. to 11:30 a.m. | S2-SSBR01-7 | Room S404
In this presentation, researchers will share real-world results from a tertiary care cancer center that compared screening interpretation with and without AI.
Radiology practices are increasingly adopting AI tools for screening mammography, but third-party data on how they perform after deployment remain scarce.
This prospective analysis led by researchers from the Memorial Sloan Kettering Cancer Center in New York included all batch-screened mammograms from 2025 and 2026. Exams acquired during the first half of each month were interpreted by a breast radiologist using AI (Transpara, ScreenPoint, version 1.7.4). Exams from the second half of each month were read without AI by one of the same group of radiologists. The team compared recall rate, cancer detection rate, and interpretation time, defined as the time from opening to signing the report. Reads shorter than one minute or longer than 20 minutes were excluded from the time analysis as outliers.
Of 37,485 screening exams, 20,885 (56%) were read without AI and 16,600 (44%) with AI. Recall rates rose significantly with AI, from 8.9% to 9.7% (odds ratio 1.11, p = 0.005). Cancer detection rate trended upward, from 6.7 per 1,000 without AI to 7.7 per 1,000 with AI (odds ratio 1.14), but the difference did not achieve statistical significance on univariate analysis (p = 0.3). Mean interpretation time also trended up slightly, from 4.2 minutes without AI to 4.4 minutes with AI (p = 0.4).
The authors conclude that AI use was associated with increased recall rates in this real-world setting. They noted that larger sample sizes will be needed to determine whether the upward trends in cancer detection and reading time are statistically significant.
Learn more by attending this session
