2023-08-03 11:52:19
Artificial intelligence is ‘good’ like two radiologists at reading a mammogram and understanding if it deserves further study; he risks no more than 4 human eyes to see a suspicious lesion where there is none, and manages to almost halve the workload of flesh and blood specialists. Promoting AI for the contribution it can make to breast cancer screening are the preliminary results of the first study that directly compared the performance of artificial intelligence in the face of a mammography exam with the standard reading done by a couple of radiologists. Positive data published in ‘The Lancet Oncology’, which however are not enough to conclude whether the use of AI in breast screening is justified or not. To tell, the search will have to continue.
Between April 2021 and July 2022, more than 80,000 women aged 40-80 who had mammography done in 4 centers in southwest Sweden were randomly assigned into 2 groups: half of the exams were analyzed by intelligence artificial and subsequently also read by one or two radiologists, depending on the level of cancer risk detected by the AI on a scale of 1 to 10; for the other half, the traditional procedure was followed without the help of artificial intelligence, which involves analysis by two radiologists. The trial thus designed, called Masai (Mammography Screening with Artificial Intelligence), is still ongoing and it is expected that it will still take several years to arrive at the final result. But an interim analysis shows ‘promising’, if ‘tentative’ results, says lead author Kristina Lång of Lund University in Sweden.
The AI failed to express a risk score on the mammograms examined in only 0.8% of cases. Recall rates (women who were recontacted for further testing) averaged 2.2% for AI-supported screening and 2% for standard double reading without AI. Overall, 244 women (28%) recalled following AI screening found out they had cancer, compared to 203 women (25%) recalled after standard screening, resulting in 41 more cancers being detected with the help of AI. ‘Ai (19 invasive and 22 in situ). The false positive rate was 1.5% in both arms of the trial. AI-supported screening showed a cancer detection rate of 6 in 1000 women, compared to 5 in 1000 with the standard double reading, which is equivalent to one additional cancer detected for every 1,000 women undergoing mammography. One finding the researchers find particularly important is that radiologists in the Ai group avoided 36,886 reads compared to their control group colleagues (46,345 reads vs 83,231), a 44 percent reduction in workload for human specialists.
“AI-supported mammography screening – the study authors summarize – led to a similar cancer detection rate compared to the standard ‘human’ double reading, with a substantially lower ‘workload’ for radiologists, indicating that the use of artificial intelligence in breast cancer screening is safe.”
The data emerging from this interim analysis of the Masai trial “should be used for new studies and evaluations”, explains Lång, also with the prospect of plugging “the marked shortage of radiologists in many countries. But they are not sufficient – he specifies – to confirm that AI is ready to be used in mammography screening” within real prevention programs, i.e. outside the experimental setting. “We still have to understand – the expert points out – the implications” of the routine use of artificial intelligence “on the final result for the patient. In particular”, it should be clarified “whether the combination of radiologists’ experience with AI can help to detect ‘interval cancers’ that often escape traditional screening”, i.e. tumors that appear after a negative test result and before the next, “as well as the cost-effectiveness of the technology”.
“At the moment”, according to Lång “the main opportunity offered by artificial intelligence is that it could relieve radiologists of excessive reading volume”. In fact, “although our AI-supported screening system requires the supervision of at least one radiologist, it could potentially eliminate the need for double-reading of most mammograms, easing the workload” for the ‘white coats’ who ” could focus on more advanced diagnostics, shortening waiting times for patients.”
On Lancet Oncology, the article is accompanied by a comment by the Italian Nereo Segnan, former head of the Cancer Epidemiology Unit and former director of the Screening Department at Cpo Piemonte (Reference Center for oncological epidemiology and prevention in the region) , not involved in the Masai study.
The specialist notes that the cancer risk score provided by artificial intelligence seems “very accurate in being able to distinguish high-risk women from low-risk ones”, and therefore highlights the “significant potential” of AI “in screening protocols stratified by risk, to adequately modulate the recall criteria” of patients who deserve further diagnostic investigations.
However, “in the group where screening was supported by artificial intelligence,” Segnan notes “the possible presence of overdiagnosis or over-detection of indolent lesions, such as a relevant portion of ductal carcinomas in situ.” An element that “should lead to caution in the interpretation of results that would otherwise seem immediately favorable to the use of AI” for breast cancer screening programs. For the expert “it is therefore important to acquire biological information on the lesions detected”, and “the final results of the Masai study are expected to indicate them”. In conclusion, according to Segnan “an important question remains to which research must answer: artificial intelligence, if suitably trained”, in the face of a suspicious lesion “is able to detect biological characteristics important for the natural history of the disease, such as the ability of tumors to grow and spread?”.
#passed #breast #screening #shes #good #radiologists