Performance of a Screening Mammography AI Algorithm Repurposed for Symptomatic Mammography in a Tertiary Outpatient Clinic
Diagnostics, 2026
Abstract
Introduction: The aim of the study was to evaluate the diagnostic accuracy of a commercial artificial intelligence (AI) algorithm originally developed for screening mammography when applied to symptomatic women presenting to a tertiary outpatient clinic.
Methods: This single-center, retrospective diagnostic accuracy study included women who presented with breast symptoms to a tertiary outpatient clinic between January and June 2013 and underwent digital mammography. An AI algorithm cleared by the U.S. Food and Drug Administration (FDA)-cleared AI algorithm was applied to all mammograms and generated continuous malignancy scores ranging from 1 to 100. Mammographic breast density was classified according to the American College of Radiology Breast Imaging Reporting and Data System (BI-RADS) by two experienced radiologists. Histopathology, when available, or otherwise a minimum of 2 years of clinical and imaging follow-up served as the reference standard. Diagnostic performance was assessed using receiver operating characteristic (ROC) analysis with calculation of the area under the curve (AUC) and 95% confidence intervals (CI) derived by patient level bootstrap resampling (n = 2000). Analyses were performed for the overall cohort and stratified by breast density (non-dense [BI-RADS A–B] vs. dense [BI-RADS C–D]).
Results: A total of 78 women (mean age, 55 ± 11 years) were included, of whom 16 had histopathological verification of suspicious lesions with proven breast cancer in 14 patients and 62 were classified based on follow-up alone. In the overall cohort (156 breasts, including 15 breasts with malignancies), the AI algorithm achieved an AUC of 0.96 (95% CI: 0.86–1.00). Performance remained high in non-dense breasts (AUC = 0.96; 95% CI: 0.88–1.00) and dense breasts (AUC = 0.99; 95% CI: 0.93–1.00), with no statistically significant difference observed between density subgroups (DeLong test, p = 0.36), although subgroup comparisons were underpowered. Decision curve analysis suggested a consistent positive net benefit across a wide range of threshold probabilities in both density groups.
Conclusion: In this preliminary, single-center retrospective cohort, a screening-trained AI algorithm showed promising diagnostic accuracy when applied to symptomatic mammograms. These findings require validation in larger, contemporary, multicenter cohorts before clinical implementation.
