Park · Diagnostics (Basel, Switzerland) 2024 · retrospective cohort and model validation study · n=21438

Artificial Intelligence-Powered Imaging Biomarker Based on Mammography for Breast Cancer Risk Prediction.

Cited 12 times in the scientific literature.

Level 3 - non-randomized controlled study

Retrospective cohort study for prognostic/diagnostic model development and external validation.

PubMed 38928628 · doi:10.3390/diagnostics14121212 · record verified 2026-08-28

What was done

Authors developed a deep learning risk prediction model using 36,995 serial mammograms from 21,438 women (17.5% cancer-enriched). The model was externally evaluated on an independent validation set of 16,894 mammograms (including 4,002 followed by a cancer diagnosis within 5 years) to predict 1- to 5-year breast cancer risk. Performance was assessed using C-indices and receiver operating characteristic AUCs, and compared against the Mirai deep learning algorithm as well as standard clinical risk tools (Tyrer-Cuzick and Gail models) using DeLong's test.

What was found

In the external validation set, the AI model achieved an overall C-index of 0.76 and 1- to 5-year AUCs of 0.90 (1-year), 0.84 (2-year), 0.81 (3-year), 0.78 (4-year), and 0.81 (5-year). The model demonstrated significantly higher AUCs than the Tyrer-Cuzick model (AUC: 0.57, p < 0.001) and the Gail model (AUC: 0.52, p < 0.001), while achieving performance comparable to Mirai (numerical metrics for Mirai were not reported in the abstract).

Why it matters

Deep learning models analyzing raw mammographic images provide substantially higher discriminatory accuracy for 1- to 5-year breast cancer risk than traditional questionnaire-based clinical risk tools.

Limits

The datasets used for both training and validation were heavily enriched with cancer cases (~23.7% in validation), which deviates substantially from standard screening population prevalence and may distort predictive performance. The abstract does not report confidence intervals, demographic characteristics, scanner variations, or prospective clinical outcomes.

Cited by