According to an analysis published on Wednesday, there is still a lack of sufficient quality studies to say whether artificial intelligence is effective for establishing a diagnosis from medical images (scanner, MRI, etc.). "Artificial intelligence (AI) seems to detect diseases from medical imaging with the same levels of relevance as health professionals" but "in view of the small number of good quality studies available, the true potential of the AI remains uncertain ", emphasize its authors.
>> READ ALSO - By the way, what is artificial intelligence?
Out of more than 20,000 publications reviewed, less than 1% were designed with a sufficiently robust methodology and results confirmed by independent experts, adds the article in The Lancet Digital Health , which presents itself as "the first systematic review" published studies on the subject.
"AI is no worse than humans"
In addition, only 25 had validated their artificial intelligence model by comparing it to the medical images of another sample of the population studied and only 14 compared the performance of machines and physicians on the same patient cases.
>> READ ALSO - Should we be afraid of artificial intelligence?
Artificial intelligence applied to medical diagnosis uses the deep learning method, which allows machines to perform complex tasks for which they have been trained, such as voice or visual recognition. "Among the few good quality studies, we found that deep learning can effectively detect diseases ranging from cancer to ophthalmic conditions as accurately as health professionals," says University consultant Alastair Denniston. Hospitals from Birmingham, UK, who led the study.
"It may be concluded that, based on the scant body of work comparing AI to physicians, AI is no worse than humans, but the data is limited and it is still too early for say, "said Tessa Cook, assistant professor of radiology at the University of Pennsylvania (USA), in an independent commentary on the study. Comparing the two is risky, she says, pointing out that doctors work in the real world, where medical data is "confusing, difficult to pin down and imperfect".