AI in mammography matches human double reading in Sweden: the value lies in better allocating radiologists' time

By El Economista · October 17, 2026.

A Swedish study of more than 105,000 women concludes that using artificial intelligence to support mammogram reading is not inferior to the traditional method of double reading by two radiologists. The trial is called MASAI (Mammography Screening with Artificial Intelligence). According to the article, its results were published in January 2026 in The Lancet and later analyzed in Radiology: Imaging Cancer.

The design was simple. One group followed the conventional procedure, with two radiologists per exam. In the other, the AI sorted mammograms by risk level: lower-risk exams were reviewed by a single specialist, while higher-risk ones received a double human reading.

The primary metric was the incidence of interval cancers, meaning those diagnosed after a screening mammogram and before the next scheduled check. With AI there were 1.55 cases per 1,000 women. In the conventional group there were 1.76 per 1,000. According to the published analysis, the AI-assisted method was not inferior on this measure. A lower incidence of certain interval cancers with features linked to a worse prognosis was also observed.

There is an important caveat. The researchers themselves stress that the study was designed to show non-inferiority, not superiority. That 1.55 is lower than 1.76 is encouraging, but it does not justify saying that AI is better than human double reading.

The WHO provides the context. According to figures the article attributes to its July 2026 update, about 2.4 million cases of breast cancer were diagnosed in women in 2024, and there were 694,000 deaths. In addition, it was the most common cancer among women in 164 of the 186 countries analyzed. Around 80% of cases occur in women with no risk factors other than sex and age, which is why mass screening remains the key tool.

Our take: the news is not that AI “detects tumors”, something we already knew it could do. What matters is that it has been measured in a trial with more than 105,000 women and against a clinical outcome that counts, interval cancers, rather than just lab metrics. That is the difference between a demo and evidence. The value here is also organizational: if AI orders the queue and frees up the second reading in low-risk cases, the scarce resource, radiologists' time, is concentrated where it adds the most.

That ties into a thesis we keep repeating: AI does not replace expert judgment, it changes where it is applied. The winner is the radiologist who supervises and decides on the difficult cases. The loser is repetitive, low-value reading. And the labor savings will only be good if accompanied by human verification, which remains the costly and critical part of the system.

There are reasonable caveats. Sweden has a specific healthcare system and population. It remains to be seen whether the results hold in other countries with different demographics, infrastructure and access. Long-term effects are not yet known either. And, as the article points out, detecting earlier is pointless if there is no timely access to diagnosis and treatment afterward.

In the short term, then, we expect gradual and cautious adoption, with regulation and auditing of each deployment. In the long term, this kind of evidence points in the direction we advocate: earlier and cheaper diagnoses, accessible even where specialists are lacking. It is a concrete step toward medicine that steadily reduces mortality from diseases like this one, provided it is measured rigorously, as it was here.