AI-supported breast cancer screening that excludes low-risk mammograms from radiologist reading may be safe and effective, according to a prospective study published March 19 in Nature Medicine.
This AI strategy, where radiologists read about one-third of total exams, led to a 15% increase in the cancer detection rate — but also a near 15% increase in the recall rate — while maintaining positive predictive value (PPV). It also led to two-thirds of exams being excluded from radiologist reading.
“These results demonstrate the feasibility of a partially automated AI workflow in breast cancer screening, avoiding human reading of studies classified as low risk,” wrote a team of researchers led by Esperanza Elías-Cabot, MD, from the Reina Sofía University Hospital and the University of Córdoba in Spain.
Recent research shows that AI systems can improve the accuracy of screening mammograms. A 2026 study out of Sweden found that AI-supported mammography leads to more cancers being found during screening and a reduction in interval cancer diagnosis. It also led to workloads being lessened for radiologists.
Elías-Cabot and colleagues studied whether AI could safely reduce workload by excluding low-risk exams from radiologist reading. The prospective study included 31,301 women who underwent routine mammograms between 2022 and 2024.
The researchers employed two screening strategies: standard double-blind reading and partially autonomous AI-supported screening. In the latter strategy, cases that the AI tool (Transpara version 1.7, ScreenPoint Medical) classified as low risk were assessed as normal, while the rest underwent double reading with AI support.
The AI strategy led to radiologist workload decreasing by 63.6%. Cancer detection rate increased from 6.3 to 7.3 per 1,000 women (p < 0.001), while the recall rate increased from 4.8% to 5.5%, achieving noninferiority.
Subanalyses included how the AI strategy affected respective workloads for digital mammography and digital breast tomosynthesis (DBT). This approach led to workload reductions for both modalities, including a 62.1% reduction for mammography and 65.5% reduction for DBT.
However, the cancer detection rate and recall rate increased for digital mammography by 1.6 per 1,000 and by 1.3%, respectively. Both measures remained stable in DBT.
The study authors suggested that this approach could be used to safely implement a partially autonomous AI screening workflow.
The authors called for future studies on the legal and safety considerations involved in excluding low-risk studies from radiologist reading, as well as on the performance of AI in DBT.
“The ethical implications of relying on AI for initial reads versus radiologists must be carefully considered to ensure patient safety and trust in screening processes,” they added.
Read the full study here.