Can AI Enhance Breast Cancer Screening While Reducing False Positives?

A large observational study has demonstrated that artificial intelligence (AI)-supported mammography screening improves breast cancer detection rates without negatively impacting recall rates. Radiologists utilizing AI assistance achieved a 17.6% higher breast cancer detection rate (6.7 per 1,000) compared with standard double reading without AI (5.7 per 1,000), while the recall rate remained noninferior, decreasing slightly from 38.3 per 1,000 to 37.4 per 1,000.

Mammography screening programs significantly reduced breast cancer mortality; however, there is room for improvement in sensitivity and specificity. Enhanced sensitivity could lower interval cancer rates and lead to earlier, more effective treatment, while improved specificity could reduce false positives, minimizing patient anxiety, and unnecessary, costly medical procedures.

According to the study authors, current screening programs, such as those in Germany, require interpretation of large volumes of mammograms by two independent radiologists, often requiring consensus conferences to maintain accuracy. This process is time-intensive, and a growing shortage of radiologists compounds the burden. Expanded screening guidelines which include younger (40-45 to 49 years) and older (70 to 74 years) age groups further increases the workload.

Integrating AI into mammography workflows offers a potential solution. Retrospective studies have shown that AI can match or even exceed radiologists in accuracy, detecting 20% to 40% of interval cancers missed by human readers and improving identification of subtle abnormalities. AI could also reduce the repetitive workload by confidently categorizing normal or highly suspicious cases while referring ambiguous cases to radiologists. Although previous studies demonstrated increased cancer detection with AI, they were limited by small sample sizes. To address these limitations, this study assessed AI’s performance in a large, real-world multicenter setting within the German mammography screening program.

This multicenter, noninferiority study was conducted across 12 screening sites in Germany between July 2021 and February 2023. A total of 463,094 women aged 50 to 69 years participated, with 260,739 undergoing AI-supported double reading by 119 radiologists who voluntarily opted to use the AI system. Key outcomes included breast cancer detection rate (BCDR), recall rate, and positive predictive values (PPVs) of recalls and biopsies. Statistical analyses accounted for potential confounders and biases.

Radiologists using AI achieved a statistically superior BCDR of 6.7 per 1000 compared with 5.7 per 1000 in the control group, with a percentage difference of 17.6% (95% CI, +5.7% to +30.8%). The recall rate in the AI group (37.4 per 1000) was noninferior to that in the control group (38.3 per 1000), with a difference of −2.5% (95% CI, −6.5% to +1.7%). Additionally, the PPV of recalls improved from 14.9% in the control group to 17.9% in the AI group, and the PPV of biopsies rose from 59.2% to 64.5%.

However, the study’s observational design introduces limitations. Radiologists chose whether to adopt AI, potentially leading to selection bias. Some radiologists preferentially used AI for examinations flagged as normal, while more complex cases were reviewed using traditional methods. Viewer functionality differences, such as synchronized zoom capabilities, may have influenced radiologists’ preferences. Despite these biases, statistical adjustments and sensitivity analyses confirmed the robustness of the findings.

The study’s strengths include its large sample size, real-world setting, and prospective design, which mitigates biases often seen in retrospective analyses. Moreover, subgroup analyses demonstrated that AI-supported screening achieved noninferior or superior BCDRs across varying breast densities, ages, and screening rounds, indicating its suitability for diverse patient populations.

“Compared to standard double reading, AI-supported double reading was associated with a higher breast cancer detection rate without negatively affecting the recall rate, strongly indicating that AI can improve mammography screening metrics,” the study authors concluded.

Reference

Eisemann N, Bunk S, Mukama T, et al. Nationwide real-world implementation of AI for cancer detection in population-based mammography screening. Nat Med. Published online January 7, 2025. doi:10.1038/s41591-024-03408-6