Highlights
- AI-supported mammography screening showed a non-inferior interval cancer rate compared to standard double reading (1.55 vs. 1.76 per 1000 participants).
- Sensitivity was significantly higher in the AI-supported group (80.5%) compared to the control group (73.8%), with no loss in specificity.
- The AI intervention was associated with fewer interval cancers characterized by unfavorable features, such as invasive growth and non-luminal A subtypes.
- Implementation of AI screening protocols can substantially reduce the screen-reading workload for radiologists without compromising diagnostic accuracy.
Background: The Evolution of Breast Cancer Screening
Breast cancer remains a leading cause of morbidity and mortality among women globally. Mammography screening programs have been the cornerstone of early detection, typically utilizing a double-reading protocol where two independent radiologists interpret each examination. While this approach increases sensitivity, it is resource-intensive and subject to human error, fatigue, and inter-observer variability.
Interval cancers—defined as primary breast cancers diagnosed between screening rounds or within two years of a negative screen—represent a significant clinical challenge. These cancers often have more aggressive biological profiles and poorer prognoses. Artificial intelligence (AI) has emerged as a promising tool to augment radiologist performance, yet until the publication of the Mammography Screening with Artificial Intelligence (MASAI) trial, its impact on interval cancer rates—a critical safety and efficacy metric—remained largely theoretical.
Study Design and Methodology
The MASAI study was a randomized, controlled, non-inferiority, single-blinded, population-based screening accuracy trial conducted in Sweden. The trial aimed to compare the interval cancer rate of AI-supported screening against the current gold standard of double reading without AI support.
Participant Allocation
Between April 2021 and December 2022, 105,934 women were randomized in a 1:1 ratio. After exclusions, the analysis included 52,946 participants in the intervention group and 52,969 in the control group. The median age of participants was approximately 53.7 years.
The AI Intervention
In the intervention group, AI was utilized in a dual capacity: as a triage tool and for detection support. Examinations were assigned an AI-derived risk score. Low-risk examinations were triaged to single reading by a radiologist, while high-risk examinations underwent double reading. Radiologists also had access to AI-generated marks highlighting suspicious findings. The control group followed the standard protocol of double reading by two radiologists without AI assistance.
Endpoints
The primary outcome was the interval cancer rate, with a non-inferiority margin of 20%. Secondary outcomes included sensitivity, specificity, and the clinical characteristics of interval cancers (e.g., tumor size, receptor status, and invasiveness).
Key Findings: Redefining Screening Performance
The results of the MASAI trial provide compelling evidence for the integration of AI into population-based screening programs.
Interval Cancer Rates
The primary analysis confirmed the non-inferiority of AI-supported screening. The interval cancer rate was 1.55 per 1000 participants in the AI group compared to 1.76 per 1000 in the control group. This yielded a non-inferior proportion ratio of 0.88 (95% CI 0.65-1.18; p=0.41), comfortably meeting the pre-defined safety threshold.
Enhanced Sensitivity and Stable Specificity
One of the most striking findings was the superior sensitivity of the AI-supported protocol. The intervention group achieved a sensitivity of 80.5% (95% CI 76.4-84.2), whereas the control group recorded 73.8% (95% CI 68.9-78.3). This improvement (p=0.031) was consistent regardless of participant age or breast density. Notably, this increase in sensitivity did not come at the cost of specificity, which remained identical at 98.5% for both groups.
Tumor Characteristics
Descriptive analysis suggested that AI might be particularly effective at identifying aggressive cancers earlier. The intervention group had fewer interval cancers that were invasive (75 vs. 89), larger than 2cm (T2+; 38 vs. 48), or of non-luminal A subtypes (43 vs. 59). This suggests that AI-supported screening may reduce the burden of clinically significant cancers that typically evade standard detection.
Expert Commentary: Clinical Implications and Workload Efficiency
The MASAI trial addresses a critical bottleneck in modern radiology: the sheer volume of screening examinations. By triaging low-risk cases to single reading, the AI-supported protocol significantly reduces the cumulative workload of radiologists. This efficiency does not appear to compromise patient safety; rather, it enhances the detection of invasive malignancies.
Biological Plausibility
The reduction in interval cancers with unfavorable characteristics in the AI group suggests that AI algorithms may be better at identifying subtle architectural distortions or micro-calcifications that human readers might overlook, particularly in dense breast tissue. This shift toward earlier detection of aggressive phenotypes could potentially translate into improved long-term survival rates, though further longitudinal data are required to confirm this.
Implementation Considerations
While the MASAI results are highly encouraging, clinicians and policymakers must consider the specific AI algorithm used and the technical infrastructure required for implementation. The trial utilized a high-performing system integrated into a well-organized national screening program. Generalizability to regions with different screening intervals or lower-volume centers remains a topic for further study.
Conclusion: A New Standard for Mammography
The MASAI trial provides robust, level-1 evidence that AI-supported mammography screening is not only safe but superior to traditional double reading in terms of sensitivity. By maintaining a non-inferior interval cancer rate and identical specificity while reducing the burden on the radiology workforce, AI-supported screening represents a significant advancement in evidence-based healthcare. These findings strongly support the transition toward AI-integrated screening protocols in clinical practice to improve early cancer detection and operational efficiency.
Funding and ClinicalTrials.gov
This study was funded by the Swedish Cancer Society, the Confederation of Regional Cancer Centres, and Swedish governmental funding for clinical research. The trial is registered with ClinicalTrials.gov, number NCT04838756.
References
Gommers J, Hernström V, Josefsson V, et al. Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial. Lancet. 2026 Jan 31;407(10527):505-514. doi: 10.1016/S0140-6736(25)02464-X. PMID: 41620232.
