Donnelly · Radiology 2024 · retrospective cohort study / algorithm development and validation · n=81,824

AsymMirai: Interpretable Mammography-based Deep Learning Model for 1-5-year Breast Cancer Risk Prediction.

Cited 53 times in the scientific literature.

Level 3 - non-randomized controlled study

Retrospective cohort study developing and validating a predictive algorithm

PubMed 38501952 · doi:10.1148/radiol.232780 · record verified 2026-08-31

What was done

Researchers developed AsymMirai, an interpretable deep learning model that bases short-term (1-5 year) breast cancer risk prediction on local bilateral dissimilarity between left and right breast tissue. The model was designed to approximate the black-box Mirai algorithm. The model was evaluated retrospectively using 210,067 screening mammograms from 81,824 patients (mean age 59.4 ± 11.4 years) in the EMory BrEast imaging Dataset (EMBED) from 2013 to 2020. Outcomes were compared against Mirai using Pearson correlation coefficients and the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals.

What was found

AsymMirai risk scores correlated moderately with Mirai (1-year risk: r = 0.6832; 4-5-year risk: r = 0.6988). AsymMirai showed slightly lower discrimination than Mirai across all timepoints: 1-year AUC was 0.79 (95% CI: 0.73, 0.85) versus Mirai's 0.84 (95% CI: 0.79, 0.89; P = .002), and 5-year AUC was 0.66 (95% CI: 0.63, 0.69) versus Mirai's 0.71 (95% CI: 0.68, 0.74; P < .001). In a subset of 183 patients where AsymMirai consistently highlighted the same tissue abnormality across successive years, the 3-year AUC was 0.92 (95% CI: 0.86, 0.97).

Why it matters

This study demonstrates that localized bilateral asymmetry accounts for a substantial portion of the predictive power in advanced black-box mammography AI, enabling clinically interpretable risk modeling with only a modest drop in discrimination.

Limits

The study is retrospective and relies on a single healthcare system dataset (EMBED). Although interpretable, AsymMirai underperformed the original black-box Mirai model at all timeframes by statistically significant margins. The high performance in the consistent longitudinal highlighting subset was derived from a very small subgroup (n = 183).

Cited by