Prostate cancer: an AI that reads MRI scans as well as radiologists

Why is a prostate MRI hard to read?

Prostate cancer is very common in men, but not every case is dangerous. Some grow so slowly that they will never cause trouble. Doctors mainly look for clinically significant cancers, the ones that need treatment.

MRI, a scan that uses magnets rather than X-rays, helps decide whether a biopsy, a tissue sample, is needed. Radiologists score each suspicious area on a scale called PI-RADS. But reading these scans takes a lot of experience, trained radiologists are in short supply and the number of scans keeps rising.

A reliable AI could ease that workload, avoid unnecessary biopsies and make a good reading available where experts are scarce.

How was the PI-CAI study run?

An international consortium built an AI using 10,207 MRI scans from 9,129 patients, taken between 2012 and 2021. It learned from 9,207 scans from three Dutch centers, then was tested on 1,000 other scans from four centers in the Netherlands and Norway.

Alongside this, 62 radiologists from 45 centers in 20 countries, with a median of 7 years' experience reading prostate MRI, read 400 of those scans using PI-RADS. The AI was also compared with the reports written at the time in hospitals, with access to the patient's history and colleagues' views.

The reference was tissue analysis plus at least 3 years of follow-up. The target was grade group 2 or higher cancers, a measure of how aggressive a cancer looks under the microscope.

What did the study find?

Significant cancer was confirmed in 2,440 cases. On the 400 scans read by the 62 radiologists, the AI reached an area under the curve of 0.91, against 0.86 for the radiologists as a group. This measure sums up how well significant cancers are told apart from the rest, with 1 being perfect.

In practice, at the same specificity, the AI found 6.8% more significant cancers. Or, at the same sensitivity, it gave 50.4% fewer false alarms and flagged 20.0% fewer low-risk cancers, which could spare people unnecessary biopsies and treatment.

On the 1,000 scans compared with routine hospital reports, the AI's specificity was 68.9% against 69.0%, at the same sensitivity. The authors judge it comparable to standard care.

What could this mean for patients?

The authors see this AI as a support tool for radiologists at first diagnosis. Fewer false alarms could mean fewer unnecessary biopsies, with their discomfort and risks.

Flagging fewer low-risk cancers could also reduce overdiagnosis, meaning the discovery of cancers that would never have caused harm but lead to worry and treatment.

What are the limits of this study?

The AI was assessed on past scans, not used live in the care of new patients. The authors write that prospective validation is needed to test how useful it is in practice.

The data came from European centers, mostly Dutch, so the AI still has to prove itself on other scanners and populations. And against routine hospital reports, non-inferiority was not formally confirmed by the planned criterion, even though the results were very close.