Measuring short and long-term breast cancer risk by combining mammographic texture models, an AI-based CAD system, and established risk factors

Lauritzen A D, Rodriguez-Ruiz A, von Euler-Chelpin M C, Lynge E, Vejborg I, Nielsen M, Karssemeijer N, Lillholm M
ECR2021

Aim and Objective
To investigate the combined effect of mammographic texture and the exam score of an AI-based CAD system in terms of quantifying short- and long-term breast cancer risk.

Materials and Method
This retrospective study comprised a cohort of 52637 double-read screens of four FFDMs from the Danish Capital Region breast cancer screening program, including 154 interval cancers (IC, diagnosed 6-24 months after screening) and 808 long-term cancers (LTC, diagnosed 2-5 years after screening). For each exam, three metrics were computed. Percent mammographic density (PMD) was measured using a deep-learning-based tool. Mammographic texture-based risk (TBR) was computed using the deep-learning architecture, ResNet34, trained to detect images of women with high probability of developing cancer in the future. An AI-CAD
system, (Transpara, ScreenPoint Medical), analyzed all studies providing an exam score from 1 to 10, where a high score indicates a high probability of visible malignancy.
Using baseline screening mammograms, we examined risk segregation performance of the models for IC (short-term) and LTC using area under the ROC curve (AUC-ROC, 95%
CI). Logistic regression was used to combine co-variates and validated using 5-fold cross-validation.

Results
TBR yielded an AUC-ROC of 0.69 (0.64-0.73) for ICs and 0.66 (0.64-0.68) for LTCs. The AI-CAD yielded an AUC-ROCs of 0.67 (0.62-0.71) for ICs and 0.64 (0.62-0.66) for LTCs.
Combining TBR, AI-CAD, PMD, and age yielded an AUC-ROCs of 0.72 (0.68-0.76) for ICs and 0.68 (0.66-0.69) for LTCs.

Conclusion
Results indicate that combining texture-based risk, Transpara Exam score, PMD, and age improved risk segregation for both ICs and LTCs compared to texture and CAD alone suggesting that both systemic and localized findings can contribute to risk modelling.