AI-Supported Mammography Met Noninferiority For Interval Cancers

Key Takeaways

• AI-supported screening had a non-inferior interval cancer rate versus standard double reading.
• Sensitivity was higher with AI support, while specificity was the same in both groups.
• Investigators also described fewer interval cancers with unfavorable characteristics, higher sensitivity across age and density, and reduced reading workload.

In the MASAI randomized trial, AI-supported mammography in Swedish population screening produced 1.55 interval cancers per 1000 participants, compared with 1.76 with standard double reading without AI. Sensitivity was higher with AI support, while specificity was unchanged. The trial compared AI-supported mammography with standard double reading without AI in routine screening within Sweden's population-based mammography program. It was designed to test whether AI support could achieve a non-inferior interval cancer rate.

The MASAI trial was randomized, controlled, non-inferiority, single-blinded, and population-based. A total of 105,934 women were assigned 1:1 to AI-supported mammography screening or standard double reading without AI, and 19 were excluded from analysis. Median ages were 53.8 years (IQR 46.5-63.3) in the intervention group and 53.7 years (46.5-63.2) in the control group. The AI system triaged examinations to single or double reading by radiologists and also provided detection support. Interval cancer rate was the prespecified primary outcome with a 20% non-inferiority margin; secondary measures included interval cancer characteristics, sensitivity, specificity, and subgroup sensitivity.

For the primary endpoint, the interval-cancer proportion ratio was 0.88, with a 95% CI of 0.65 to 1.18 and p=0.41, supporting non-inferiority. Sensitivity was 80.5% in the intervention group, with a 95% CI of 76.4% to 84.2%, versus 73.8% in controls, with a 95% CI of 68.9% to 78.3%. That difference was statistically significant, with a p value of 0.031. Specificity was 98.5% in both groups, with 95% CIs of 98.4% to 98.6% and a between-group p=0.88. Interval cancer rates were non-inferior, with higher sensitivity and unchanged specificity.

Descriptively, fewer interval cancers in the intervention group were invasive (75 versus 89), T2+ or non-luminal A, with counts of 38 versus 48 and 43 versus 59, respecitvely. Higher sensitivity was consistent across age and breast density, and for invasive cancer, but not for in-situ disease. The authors described the findings as favorable, linked the approach to reduced screen-reading workload, and wrote that AI-supported mammography may be considered for implementation in clinical practice.