AI software as a third reader in breast cancer screening-a prospective diagnostic observational study

Despite advances in mammography screening, some cancers remain undetected, prompting the evaluation of artificial intelligence (AI) as an independent third reader to reduce missed cancers.

In this prospective study, women eligible for the German Mammography Screening were enrolled at six sites belonging to one screening unit between August 2023 and February 2024. Each mammogram underwent double reading and was independently analyzed using Transpara, an AI-based detection software. Cases rated BI-RADS 4 or 5 by any reader or given a risk score of 10 by the software were reviewed in a consensus conference. Endpoints included: primary-cancer detection rate (CDR) and positive predictive values (PPV); secondary-analysis of cancers detected only by the software or missed by it.

15,356 female participants (mean age 58.6 ± 5.6 years) were included. Overall, 115 breast cancers were detected (CDR triple reading: 0.75%; 95% CI: 0.62%, 0.90%). CDR of double reading and standalone AI was 0.68% (95% CI: 0.56, 0.83%) and 0.66% (95% CI: 0.54, 0.81%). Using Transpara as a third reader increased the detection rate by 9.5% (95% CI: 4.7%, 16.8%) compared to double reading (p = 0.002). The PPV for consensus-conference referrals was 5.1% (95% CI: 4.2%, 6.1%), lower than double reading 7.5%(95% CI: 6.2%, 9.0%; p < 0.001). For recalled cases, the PPV was 13.7%(95% CI: 11.5%, 16.2%) versus 15.2% (95% CI: 12.6%, 18.1%; p < 0.001). All nine invasive cancers detected solely by AI were Luminal-A-like cancers. Among 13 cancers missed by the software, four were triple-negative.

Adding Transpara as an independent third reader improved detection rates, mainly by identifying additional Luminal-A-like cancers, and increased the workload to the consensus conference and the number of recalled cases.

Question Does the integration of AI software as an independent third reader improve cancer detection rates in mammography screening without increasing false-positive findings and recall rates? Findings AI as an independent third reader increased cancer detection by 9.5%, mainly identifying Luminal-A-like cancers, significantly decreasing the positive predictive values of cases referred to at the consensus conference and increasing the number of recalled cases. Clinical relevance Using AI as an independent third reader enhances mammographic cancer detection by offering radiologists complementary sensitivity, especially for low-risk lesions. However, maintaining human readers is essential, as AI may miss aggressive subtypes like triple-negative breast cancers.