AI performs well in breast screening, but density complicates the picture

European breast screening is gradually moving away from uniform, age-based programs toward pathways that also consider individual risk. Screening ages vary across Europe. BreastScreen Norway offers biennial mammography to women ages 50–69, while programs in countries including Austria begin routine invitations at 45.

In risk-stratified screening, a woman’s risk factors influence how frequently she is screened and whether she receives mammography alone or supplemental imaging. The European Society of Breast Imaging (EUSOBI), for example, recommends MRI for women with extremely dense breasts.

Breast density is central to this transition. Dense tissue increases breast cancer risk and can hide tumors on mammograms, known as the masking effect. Automated density assessment could help determine who needs MRI, contrast-enhanced mammography (CEM), ultrasound, or more frequent screening. The catch: The measurement must be reliable.

200,000 screening examinations

In the study, published online on 13 August in European Radiology, statistician Marthe Larsen of the Department of Breast Cancer Screening at the Cancer Registry of Norway, Norwegian Institute of Public Health, in Oslo, Norway, and colleagues retrospectively analyzed 200,000 examinations from BreastScreen Norway.

The examinations were performed at eight breast centers between 2010 and 2021. The researchers selected 100,000 examinations acquired using Siemens Healthineers equipment and 100,000 obtained using Hologic systems.

The images were analyzed using Transpara version 2.1, a commercially available, CE-marked AI model (ScreenPoint Medical) that produces malignancy risk scores and automated volumetric breast density measurements. The dataset included 1,235 screen-detected cancers and 368 interval cancers.

AI performance fell as density rose

The AI performed well overall, but cancer-detection performance declined with increasing breast density when screen-detected and interval cancers were considered together. The area under the curve fell from 0.955 in the lowest-density category to 0.857 in the highest.

The interval cancer rate also increased from 0.5 per 1,000 examinations in the lowest-density group to 3.1 per 1,000 in the highest. Interval cancers are diagnosed after a negative screening examination but before the next scheduled round.

When only screen-detected cancers were considered, however, performance did not differ significantly between density groups, suggesting that interval cancers strongly influenced the decline.

Density label was not always consistent

Among examinations performed using Siemens equipment, 3.3% were placed in the highest volumetric density group, compared with 6.7% of examinations obtained using Hologic systems.

This does not prove that the equipment caused the difference. The systems were used at different centers and in different populations. Women screened using Hologic systems were more likely to live in urban areas, which may partly explain the variation, according to the researchers.

In 18.5% of examinations, the AI also placed the left and right breasts in different density categories, raising the question of which measurement should determine the screening pathway.

AI-derived mammographic risk scores

Major European trials have investigated supplemental screening for women with dense breasts. The Dutch DENSE trial examined MRI, while the U.K. BRAID trial compared abbreviated MRI, CEM, and automated breast ultrasound.

The Norwegian findings identify the work still required. Before automated density assessment can guide personalized screening, European programs may need cross-vendor calibration, standardized thresholds, and a consistent method for converting breast-level measurements into one clinical decision.

The full article can be found here.