Using an AI system and breast density to quantify the short-term risk of interval cancer in screening: a large retrospective evaluation

Wanders A J T, Mees W, Janssen N, Dalmis M U, Rodriguez-Ruiz A, Sechopoulos I, van Gils C H, Karssemeijer N, Mann R M, van Rooden J-K
ECR2021

Aim and Objective
To create a short-term breast cancer risk model combining the assessments of an artificial intelligence (AI) detection system and breast density (BD) to assess the risk of interval cancers (IC) in screening mammograms.

Materials and Method
This nested case-control study included DM exams of 2332 IC cases of women that participated in the Dutch breast cancer screening program, and a randomly selected group of 584 normal screening controls verified by at least 2 years of follow-up. For each mammogram, a score 1-10 indicating the risk of presence of visible abnormalities was computed by an AI detection system (Transpara, ScreenPoint Medical), and BD was computed using automated software.
A neural network-based model was trained combining the AI score and BD to predict the short-term risk of IC in screening mammograms using 10-folds cross validation. The area under the receiver operating characteristic curve (AUC) of Transpara, breast density, and the risk model for detection of IC at screening mammograms was estimated with 95% confidence intervals and compared using T-test for paired samples.

Results
The AUC of the neural-network risk model to detect IC in screening was 0.801 (0.774-0.827), higher (P<0.01) than the AI system or BD alone, 0.741 (0.713-0.769) and 0.695 (0.660-0.730). At 90% specificity, the risk model achieved a sensitivity of 49.7% (43.2%- 56.2%), higher (P<0.01) than the AI system or BD alone, 42.2% (37.1%-47.4%) and 23.2% (16.1-30.3%).

Conclusion
Combining breast density and assessments by Transpara improves the selection of women at risk for developing IC over either method alone, allowing for the correct prediction of about half of the IC at 90% specificity