Related Experiment Video
Updated: Jul 14, 2026
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
Measuring Breast Density at Scale With AI: Insights From BreastScreen NSW
Richard Walton1, Douglas Dunn2, Matthew Warner-Smith1,3
1Cancer Institute NSW, St Leonards, New South Wales, Australia.
Introduction:
Breast density is a marker of risk both due to the increased burden of fibroglandular tissue, which may harbour malignancy, and to obscuration of lesions in mammography. This study aims to assess the reliability of Lunit INSIGHT MMG as a tool for estimating breast density compared to breast radiologists, evaluate density on a population level of > 600,000 clients, and compare the density in clients with/without implants.
Methods:
15,518 mammograms were abstracted to explore the intra- and inter-reader reliability between AI and breast radiologists. Variation in density was then evaluated by AI from 624,133 mammograms and stratified according to birth country and age. Finally, density was assessed in 4047 clients with implants and compared to 396,301 clients without implants.
Results:
AI vs. breast radiologist inter-reader variability weighted kappa coefficient was 0.72 (95% CI 0.71-0.73). Clients with a North-East Asia background had the highest density, whilst clients of Aboriginal background had the lowest density. Density reduced with age across all backgrounds, though at different rates. Clients with implants had higher density than age-matched no-implant strata.
Conclusion:
AI performed adequately in both intra- and inter-reader reliability. Notable gradients in breast density were observed with increased age but this effect was moderated by ancestry. This study population is significantly larger than any other in the published literature and has considerable diversity in age and ancestry.
