Evaluation of a Mammography-based Deep Learning Model for Breast Cancer Risk Prediction in a Triennial Screening Program
Menée à partir de données portant sur 134 217 femmes incluses dans le programme britannique de dépistage du cancer du sein (âge moyen : 59,1 ans), cette étude évalue la performance d'un algorithme d'apprentissage automatique, utilisant des mammogrammes et générant un score de risque, pour identifier les patientes nécessitant des dépistages plus fréquents ou des examens d'imagerie supplémentaires dans le cadre d'un programme de dépistage triennal et réduire ainsi l'incidence des cancers du sein de l'intervalle
Background : Deep learning risk algorithms for personalized breast cancer screening outperform traditional methods in retrospective evaluations, but triennial screening assessments are lacking.
Purpose : To evaluate the predictive ability of 3-year risk scores generated by a deep learning algorithm (Mirai) to identify women who developed interval cancers (ICs) in the UK breast screening program, which invites women aged 50–70 years for triennial mammography.
Materials and Methods : For this retrospective study, Mirai processed digital screening mammograms with negative results collected from a 3-year cohort (January 2014 to December 2016) across two sites and two primary mammography systems. Exclusions included screen-detected cancers (baseline and next round), implants, and nonstandard views. The reference standard was no cancer diagnosis within 40 months of negative screening, confirmed histopathologically. The primary objective was predicting ICs at 1-, 2-, and 3-year time points after baseline screening. Secondary objectives were assessing predictions across age quartiles and Breast Imaging Reporting and Data System (BI-RADS) breast densities. Areas under the receiver operating characteristic curve (AUCs) and true positives (ICs) were calculated across operating thresholds. Risk score distributions were compared with the Mann-Whitney U test, and AUCs were compared with the DeLong test.
Results : Analysis included 134 217 examinations from the same number of women (mean age, 59.1 years ± 7.9 [SD]), including 524 ICs. There was no evidence of performance differences among 1-, 2-, and 3-year IC predictions (P
≥
.63), age quartiles (P
≥
.73), or breast densities (P
≥
.99). Overall AUCs were 0.72 (95% CI: 0.65, 0.78), 0.67 (95% CI: 0.64, 0.70), and 0.67 (95% CI: 0.65, 0.70) for 1-, 2-, and 3-year IC predictions, respectively. C indexes for age quartiles were 0.67 (95% CI: 0.62, 0.71) for age younger than 52 years, 0.70 (95% CI: 0.65, 0.75) for age 52–58 years, 0.71 (95% CI: 0.67, 0.75) for age 59–65 years, and 0.71 (95% CI: 0.67, 0.75) for age of 66 years and older. C indexes for BI-RADS categories a, b, c, and d were 0.70 (95% CI: 0.62, 0.78), 0.69 (95% CI: 0.65, 0.73), 0.68 (95% CI: 0.64, 0.71), and 0.67 (95% CI: 0.62, 0.73), respectively. Three-year risk scores retrospectively predicted 3.6% (19 of 524), 14.5% (76 of 524), 26.1% (137 of 524), and 42.4% (222 of 524) of ICs for women assigned the highest 1%, 5%, 10%, and 20% of scores.
Conclusion : Mirai could identify women for more frequent screening or additional imaging, detecting ICs earlier.
Radiology , article en libre accès, 2025