Artificial intelligence can improve human detection of breast cancer through more efficient mammographic screening, while reducing the workload of radiologists, suggests a study published Friday.
This work, published in the medical journal “The Lancet”, focused on the analysis of nearly 106,000 mammograms carried out in Sweden as part of a national breast cancer screening program, from April 2021 to early December 2022.
These examinations, carried out on women with a median age of 53, were examined for half of them by two radiologists working successively, and, for the other half, by a radiologist assisted by AI.
It appears that AI-assisted mammographic screening made it possible to reduce the workload linked to reading the images, and above all, according to the study, gave “systematically better results”.
AI-assisted breast cancer screening made it possible to identify 81% of women with cancer during screening compared to 74% without AI, “without increasing the rate of false positives,” its authors conclude.
In addition, women who benefited from AI-assisted screening were diagnosed with fewer “aggressive and advanced” breast cancers (-12%) in the following two years.
This improvement was observed regardless of the age and breast density (one of the risk factors among others) of the women examined.
Correcting the AI diagnosis
Previous studies had notably shown similar effectiveness of AI-assisted screening with screening based on a double human look, then better effectiveness but targeted on small invasive cancers, reports the study.
The results of the work published in The Lancet suggest that AI-assisted mammographic screening can “effectively improve screening performance” and “reduce the workload” of radiologists, pleading for its implementation in clinical practice, believe its authors.
But according to Jean-Philippe Masson, president of the National Federation of Radiologists (FNMR), recent American studies have suggested that these tools “do not really save time and are expensive”.
“It would save time if the radiologist could trust the tool and say ‘Okay, the AI didn’t find anything, so I’m signing a report saying there’s nothing’. But this is not possible: the eye and the experience of the radiologist must correct the diagnosis of AI”, in particular to avoid “over-diagnoses”, he says.
Because sometimes “the AI tool will have seen a modification of the breast tissue which in reality is not a cancer, but for example, a crossing of vessels”, he adds.
Questioned by AFP, Stephen W. Duffy, professor emeritus of cancer screening at the Wolfson Institute of Public Health at Queen Mary University of London, considered the study “robust”.
If it confirms that AI-assisted screening is “as safe as reading by two human radiologists”, it remains to be confirmed by subsequent monitoring that women who have benefited from this type of screening develop less cancer, he warns.
Early detection has contributed to reducing mortality due to breast cancer, the study recalls, but “certain cancers are still not detected” by screening mammography, “despite the double reading of the images recommended by European directives”. Because certain rapidly progressing cancers occur between two screenings.
Coming into force in August 2024 and applicable in stages from August 2026, the European regulation on artificial intelligence (AI Act) has classified medical devices incorporating AI as “high risk” when they contribute to diagnosis or influence a therapeutic decision in particular. They will therefore have to be subject to reinforced security and performance requirements.
