A Nice Win for Radiology AI

AI actually saving lives *and* reducing work!

A couple years ago, The Lancet reported the interim findings of the MASAI trial – an AI-augmented workflow for mammography reading in Sweden. In Europe, the standard of care for mammography reading involves “double reading”, in which two breast radiologists review all images for cancer. The MASAI trial replaces one of those radiologists with AI.

In the initial report covering 80,000 mammography examinations, the AI workflow passed its initial test regarding the accuracy of cancer detection – “outperforming” the double-human standard in cancer detection at 6 per 1,000 versus 5 per 1,000. False positives were equal in each group, meaning the “positive predictive value” of the AI workflow was better than the baseline practice.

Now, this follow-up publication describes the completion of the study up to the full 100,000 examinations, as well further details on the types of cancer identified.

Consistent with the prior publication, a greater number of cancers were identified with the AI workflow. This portion of Table 3 is best illustrative of whether that was a Good Thing:

There is “more diagnosis”, and the question of whether this represents “over diagnosis” is a reasonable debate. However, the number of invasive cancers detected far exceeds the number of carcinoma in situ – and the authors further describe the subtypes of invasive cancer to be those where substantial downstream morbidity and cost was likely avoided by this improved workflow.

So, all good news on the cancer front – and with a 44% reduction in human-reviewed images, compared to baseline clinical practice. Chalk this one up as a nice win for AI.