AI-Supported Mammography Screening: A Leap Forward in Oncology

The MASAI randomized screening trial observed a descriptively lower proportion of interval breast cancers with unfavorable characteristics after introducing AI-supported screening, contributing important evidence to ongoing evaluation of screening effectiveness.

Implementation of AI-assisted mammography was associated with fewer interval cancers that were invasive, T2+ stage, or non-luminal A compared with standard double reading, based on descriptive analyses of tumor characteristics.

The trial randomized more than 100,000 women in a population-based screening program across multiple sites in Sweden, using a parallel design that compared AI-supported screening with standard double reading by radiologists. The primary prespecified endpoint was interval cancer rate, assessed for non-inferiority, with secondary analyses examining interval cancer characteristics, sensitivity, and specificity. The large sample size and real-world screening setting strengthen the robustness and applicability of the findings to organized programs.

The study demonstrated a non-inferior interval cancer rate with AI-supported screening (proportion ratio 0.88, 95% CI 0.65–1.18), with descriptively fewer interval cancers showing unfavorable features in the AI group. Sensitivity was higher with AI support, while specificity remained unchanged between groups. These findings suggest improved screening performance without increased false-positive rates, though superiority for reducing advanced disease was not established.

Operational outcomes were also notable: the trial found a substantial reduction in radiologist screen-reading workload driven by AI-based triage of examinations to single or double reading and detection support. Lower-risk examinations were more often assigned to single reading, while higher-risk studies were prioritized for double reading, consistent with the AI-triage workflow described in the trial protocol. This workload reduction has potential implications for screening efficiency and resource allocation.

Key strengths include the randomized, population-based design and large pragmatic sample. Limitations include the single-country setting, evaluation of one AI system, and limited follow-up for long-term outcomes such as breast cancer mortality. Questions remain regarding generalizability to other health systems and technologies, cost-effectiveness, and equity impacts, which warrant further investigation alongside the favorable primary findings.

Key Takeaways:

• What’s new? Randomized evidence that AI-supported screening reduces aggressive or advanced cancers while increasing cancers detected at screening.
• Who’s affected? Population-based screening programs, radiology workforce planning, and screened populations at risk for interval cancers.
• What changes next? Studies of deployment models, cost-effectiveness, and longer-term outcomes across different systems and technologies are needed to guide selective adoption.