Geoffrey Hinton, one of the fathers of modern artificial intelligence (AI), essentially claimed on October 27, 2016 in Toronto (Canada) that we must stop training radiologists because algorithms overtook them five years ago! Six years later, practicing radiology always, and the short-term craze around the machine learning (“machine learning”) has given way to more measurement.

“The promises are numerous and it is sometimes difficult to distinguish them from the facts”, Note, Tuesday, January 10, the National Advisory Ethics Committee (CCNE) and the National Pilot Commission on Digital Ethics (CNPEN), in their joint statement. Published under the title Medical diagnosis and artificial intelligence: ethical issues, it summarizes sixteen recommendations and seven points of vigilance. Where exactly is the use of AI, heralded as one of the medical revolutions of this century? Where is the use of this astronomical amount of digitized data (text, images and numbers) to train algorithms to help with diagnosis and therapy choice?

At the European Congress of Radiology, held in Vienna in July 2022, Emily Conant, Professor of Radiology at the University of Pennsylvania Hospital, said, ‘ called for cautionsays Isabelle Thomassin, Head of the Medical Imaging Department at Tenon Hospital (AP-HP, Sorbonne University). AI is an opportunity, we’re sure of it, but it has to be used properly. »

The analysis of medical images (conventional X-rays, mammograms, etc.) is currently the first concrete use of AI worldwide “a recent systematic review showed that in mammography, for example, methodological distortions existed in most scientific studies”, explains the French radiologist. Like that from google health in 2020, still published in the prestigious journal Natureaccording to which the AI-enhanced machine alone performs better than a radiologist. “The software was trained on a population of women whose breasts were much more susceptible to cancer than the general population.”explains the specialist. “However, it is statistically much easier for software to say ‘it’s a cancer’ when there are many cancers than when there are few. » The performance of the algorithms can also depend on the geographic origin of the images they generated. “Japanese women’s breasts are very dense, while northern Europeans, referred to as ‘light breasts,’ are fatter and easier to read.”this expert illustrates.

You still have 85.11% of this article to read. The following is for subscribers only.