Screening mammography performance according to breast density: a comparison between radiologists versus standalone...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• AI algorithms for mammogram assessment are not well-established in Asian populations.
• This study compares AI performance against radiologists in a large cohort of Korean women.

Purpose of the Study:

• To evaluate the performance of AI standalone detection versus radiologists in screening digital mammography.
• To assess the impact of breast density on AI and radiologist performance.

Main Methods:

• Retrospective analysis of 89,855 screening digital mammograms from Korean women (2009-2020).
• Lunit software used for AI-based malignancy probability scoring.
• Comparison of performance metrics (CDR, sensitivity, specificity, PPV, recall rate, AUC) between AI and radiologists across breast density categories (A-D).

Main Results:

• No significant difference in cancer detection rate (CDR) and sensitivity between AI and radiologists.
• AI showed significantly higher specificity, positive predictive value (PPV), and AUC, with a lower recall rate.
• AI performance advantages, especially in specificity and recall rate, were consistent across all breast density categories, including extremely dense tissue.

Conclusions:

• AI software exhibited comparable sensitivity to radiologists but outperformed them in specificity, PPV, recall rate, and AUC.
• These performance differences were most pronounced in women with extremely dense breast tissue.