Using AI to identify very likely normal cases that may not need a second reader assessment in a French breast cancer screening program (BCSP): a retrospective evaluation

Jarraya H, Damiens Y, Amrane Y, Vendel C, Legros M, Brochart C, Rais J, Medjahdi M, Petit T, Rodríguez-Ruiz A
ECR2021

Aim and Objective
To retrospectively investigate an AI system as aid to first reader to identify which cases do not need double reading because very likely normal in BCSP with digital mammography (DM).

Materials and Method
A cohort of 998 DM was collected and enriched with additional 43 detected cancers (52 cancers in total). Five radiologists read screening exams, all DM (and their prior) were processed by an AI system (Transpara, ScreenPoint Medical), which categorizes them on a scale 1-10 representing the risk of abnormalities. Breast density was also recorded.
Different cutoff points of Transpara were investigated to define the largest group of screening exams with the lowest likelihood of cancer. Binomial method was used to compute 95% confidence intervals (CI) of the distributions.

Results
Detection wise, 49/52 detected cancers had an AI risk score of 8 or higher. Using AI to identify likely normal exams, cases with risk score 4 or lower included 0% (CI: 0-7.7) of cancers and 29% (CI: 26-32) of total screening volume. When taking breast density into account, cases with risk scores 7 or lower and with low breast density (A or B) included 0% (CI: 0-7.7) of cancers and 47% (CI: 44-50) of total screening volume. When looking at prior exams 2 years earlier, 43% (CI: 20-70) of detected cancers already had the highest AI risk score of 10.

Conclusion
Transpara could be used in screening programs as aid to the first reader to identify 29% of very likely normal mammograms where double reading may not be necessary. Combined with breast density, this volume could be increased to 43%.