Clinical Application of AI in Mammography: Insights from a Prospective Study
Ebru Yilmaz, Mustafa Ege Seker, Nilgun Guldogan, Ebru Banu Turk, Servet Erdemli, Yilmaz Onat Koyluoglu, Sehla Nurefsan Sancak, Erkin Aribal
Academic Radiology, 2025
Abstract
Rationale and Objectives
This prospective study evaluated the performance of AI in a diagnostic clinic setting, comparing its effectiveness with radiologists of varying experience.
Materials and Methods
The study was conducted at a single center and included 1063 patients undergoing diagnostic or screening mammography. Five radiologists with different experience levels assessed the images using the fifth edition of the BI-RADS lexicon. Standalone AI software assigned risk scores (0−100), with scores above 30.44 considered positive. AI risk assessments were compared with radiologists’ BI-RADS scores. Radiologists also re-evaluated AI-positive mammograms as a second look. Ground truth was established through histopathology and two years of follow-up.
Results
Right and left breasts were analyzed separately, and 2126 mammography images were evaluated from 1063 women. A total of 29 cancers were diagnosed in 28 women. Among all examinations, 2.44% (52/2126) were positive, of which 46.15% (24/52) were true positive. Standalone AI detected 82.75% (24/29) of cancers, and the majority voting of radiologists scored positive (BI-RADS 0,4 and 5) in 8% (172/2126) where 89.65% (26/29) of cancers were detected. The AUC score of majority voting was 94.7% (95% CI: 91.1–98.3), and AI was 94.4% (95% CI: 88.5–100). AI was statistically not significantly different than (p=0.79) AUC of the majority voting. The re-evaluation assessment of AI-flagged images achieved an AUC of 94.8% (95% CI: 91.2–98.3), significantly different from the initial evaluation (p=0.015). However, it was not significantly different from AI (p=0.74).
Conclusion
AI algorithms in diagnostic settings can serve as effective CAD systems, aiding in breast cancer detection and reducing inter-reader variability.