https://atlas.rsna.org/schemas/2025-11/model.json

Evaluation of Mirai on ChiMEC high-risk cohort

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10698602/

1.0

Supported by the University of Chicago Comprehensive Cancer Care Center Spotlight grant (6-9398-9660), Susan G. Komen for the Cure (SAC210203, TREND21675016), NIH/NCI (P20CA233307), and Breast Cancer Research Foundation (BCRF-21-071).

Retrospective case-control study; HIPAA compliant; approved by the University of Chicago Institutional Review Board.

Deep learning convolutional neural network–based model; each mammogram view encoded into a 512-dimensional vector and four views concatenated into a 2048-dimensional representation (Mirai).

Publicly available Mirai model with pretrained weights was used; validation code available by request at https://github.com/olopade-lab/mirai_validation

Predicts 1- to 5-year breast cancer risk from screening mammograms to support risk stratification and potential screening/management decisions.

Clinical decision support systems; risk stratification for screening and follow-up planning.

No fixed clinical threshold established; examples used include BI-RADS category 3 as cutoff (year-1 comparison) and the 75th percentile of Mirai risk scores (year-5 sensitivity/specificity illustration).

Automates generation of risk scores from standard screening mammography; supports, but does not replace, clinical decision-making.

Prediction of near- and long-term (1–5 year) breast cancer risk in women undergoing screening mammography; evaluated in a high-risk, racially diverse cohort.

Four standard screening mammographic views (bilateral CC and MLO) in DICOM For Presentation mode, plus time-to-event labels for evaluation.

Provide all four standard views without implants, foreign devices, or burned-in annotations; ensure adequate follow-up for outcome labeling in evaluation settings.

External performance lower for dense breasts in near-term prediction (not statistically significant); tested in a high-risk case-control cohort enriched for African American women, benign breast disease, and BRCA mutation carriers; calibration not assessed; potential dependence on ipsilateral premalignant patterns indicated by mirroring experiments; sample size limited for some subgroups; BI-RADS subcategory granularity unavailable.

Consider combining Mirai risk scores with BI-RADS assessment to potentially improve near-term discrimination; further validation and calibration in larger prospective cohorts recommended.

Study code for applying models is available by request (GitHub link provided); publicly available Mirai pretrained weights used; detailed AUC and subgroup analyses reported with CIs and statistical tests.