AI in mammography: how to read the MASAI trial results

Cancerforskning förklarad med lera och stop-motionanimering

Faculty of Medicine, Lund University · 2023-10-25 · Source

A short Swedish-language research explainer from Lund University. This 2023 animation introduces the study background; it does not present the 2026 follow-up results.

Research: a system working with radiologists

In the Swedish randomized MASAI trial, AI assigned mammograms to single or double radiologist reading and highlighted suspicious areas. The comparator was standard double reading without AI. The 2025 analysis reported cancer detection of 6.4 versus 5.0 per 1,000 participants and a 44.2% reduction in screen-reading workload. This does not mean all radiology work fell by that amount or that physician assessment was removed.

2026 follow-up: a lower rate does not establish superiority

Interval cancer rates were 1.55 versus 1.76 per 1,000, with a ratio of 0.88 (95% confidence interval 0.65–1.18). The prespecified 20% margin for non-inferiority was met; the interval includes 1, so a significant reduction was not demonstrated. Sensitivity was 80.5% versus 73.8%; specificity was 98.5% in both groups. Interval cancers were diagnosed after being undetected at screening.

Clinical assessment: screening has benefits and harms

NCI describes early detection alongside the possibility of false positives, false negatives and overdiagnosis. Overdiagnosis means finding a cancer that would not cause problems during a person’s lifetime; it is different from a false positive. When reading an AI story, we should consider how these outcomes are followed, as well as how many cancers are found.

Regulatory authorization: a separate record

The FDA’s AI device list links to devices authorized for marketing in the United States and their decision records; the agency explicitly states that the list is not comprehensive. A favorable research result does not establish authorization for every software version, intended use or country. This research summary does not replace verification of a particular product’s current regulatory status.

Editorial perspective: what should local implementation establish?

We read these publications as an evaluation of a specific AI workflow alongside radiologists. An institution should define the model and version, imaging equipment, patient population, reading protocol and follow-up plan together. The same outcome should not be assumed to transfer automatically to another center. This article provides neither personal screening advice nor evidence of clinical effectiveness for RAI products.