AI as a Third Reader in Breast Cancer Screening

Key Takeaways

• Triple reading with AI was associated with higher cancer detection than double reading alone and standalone AI.
• Adding AI was associated with lower positive predictive value for consensus-conference referrals, along with higher consensus-conference workload and more recalled cases.

In a recent study, adding Transpara as an independent third reader to double reading in German mammography screening was associated with a cancer detection rate of 0.75%, versus 0.68% with double reading alone. The added detections also came with more cases moving into review and recall pathways, pairing higher detection with greater downstream assessment burden across the screening workflow.

This prospective diagnostic observational study enrolled women eligible for the German Mammography Screening program at six sites within one screening unit. Between August 2023 and February 2024, investigators included 15,356 women, whose mean age was 58.6 ± 5.6 years. Transpara was used independently as an AI-based third reader alongside standard double reading. Cases rated BI-RADS 4 or 5 by any reader, or given a software risk score of 10, were reviewed in consensus conference, and the primary endpoints were cancer detection rate and positive predictive value.

Across the cohort, 115 breast cancers were detected overall. The cancer detection rate was 0.75% with triple reading, with a 95% CI of 0.62% to 0.90%. Double reading yielded a cancer detection rate of 0.68%, with a 95% CI of 0.56% to 0.83%, while standalone AI yielded 0.66%, with a 95% CI of 0.54% to 0.81%. Compared with double reading, adding Transpara as a third reader was associated with a 9.5% increase in detection, with a 95% CI of 4.7% to 16.8% and p=0.002.

The positive predictive value for consensus-conference referrals was 5.1%, with a 95% CI of 4.2% to 6.1%, compared with 7.5% for double reading, with a 95% CI of 6.2% to 9.0% and p<0.001. For recalled cases, positive predictive value was 13.7%, with a 95% CI of 11.5% to 16.2%, versus 15.2% for double reading, with a 95% CI of 12.6% to 18.1% and p<0.001. These lower values were accompanied by greater workload at consensus conference and a higher number of recalled cases.

Subtype findings showed that all nine invasive cancers detected solely by AI were Luminal-A-like. Among 13 cancers missed by the software, four were triple-negative. The added detections were framed as complementary sensitivity for lower-risk lesions within this workflow. Human readers remain essential as AI may miss aggressive subtypes, such as triple-negative breast cancers.