Recent findings from Lund University in Sweden reveal that AI-supported mammography screening is a promising, safe alternative to the traditionally accepted practice of double reading by radiologists.
The timing of the research, which is published in the journal The Lancet Oncology, is particularly important given a current shortage of breast radiologists.
Mammography has been a staple in breast cancer detection for decades. Every year in Sweden, approximately one million women undergo mammography screening.
To ensure high sensitivity in detection, each mammography screening is meticulously reviewed by two breast radiologists in what’s termed as “double reading.”
However, this practice is under threat, owing to an alarming shortage of breast radiologists – not just in Sweden but globally. The implications of this deficit could be dire, jeopardizing the screening service and potentially risking lives.
With this in mind, the potential of integrating AI into mammography screening has lately sparked keen interest. However, the exact methodology for its optimal implementation and its subsequent clinical implications have remained shrouded in uncertainty.
The Mammography Screening with Artificial Intelligence (MASAI) trial is the first randomized controlled trial to assess the effects of AI-supported screening, it’s a pioneering effort to determine the efficacy and safety of AI in mammography.
Kristina Lång, associate professor in diagnostic radiology at Lund University, explained how the research was conducted.
“In our trial, we utilized AI to earmark those screening examinations that manifested a high risk of breast cancer. These examinations were subjected to the standard double reading by radiologists. Conversely, those deemed low-risk by the AI were assessed by a single radiologist,” said Professor Lång.
“Moreover, during the screen reading process, radiologists had the added advantage of AI which underscored suspicious findings on the images.”
Overall, 80,033 women were methodically divided into two distinct cohorts: 40,003 underwent AI-supported screening, while the remaining 40,030 were subjected to the standard double reading without the support of AI. The outcomes were remarkable.
“Through the integration of AI, there was a notable detection of 20% more cancers when juxtaposed with standard screening, with no uptick in false positives,” said Professor Lång.
This is significant since a false positive might cause unnecessary distress, recalling a woman for further tests only to eventually clear her of any cancer suspicion.
The benefits extended beyond just accuracy. AI introduction led to a 44% reduction in the screen-reading workload for radiologists. To offer a perspective on the time savings, consider this: an average radiologist pores over 50 screening examinations hourly.
The study suggests that, due to AI’s efficiency, roughly five months of a radiologist’s time was saved in reading about 40,000 screening examinations.
However, Lång exercises caution, noting: “Our study was confined to a single site in Sweden. The next logical step is to assess if these encouraging results retain their validity in varied settings, perhaps with different radiologists or AI algorithms.”
“While there could be other avenues to integrate AI in mammography screening, a prospective setting remains paramount for evaluation.”
The MASAI trial, which now boasts an enrollment of 100,000 women, is poised to venture into its next phase. The research team is gearing up to discern the types of cancers identified with and without AI’s assistance.
A significant metric under the lens will be the interval-cancer rate – cancers diagnosed between screenings, known to have a generally unfavorable prognosis when compared to screen-detected cancers. This rate will be meticulously examined after the trial’s women have undergone a minimum two-year follow-up.
“There’s a delicate balance between the advantages and potential pitfalls. A method that detects more cancers isn’t inherently superior. The gold standard is to pinpoint clinically significant cancers early on. Yet, this must be juxtaposed with the perils of false positives and overdiagnosis of dormant cancers,” said Professor Lång.
“Our initial findings underscore the safety of AI-supported screening, especially given the unchanged cancer detection rate amidst a marked dip in screen-reading workloads. The upcoming analysis of interval cancers will further elucidate if AI-driven screening indeed revolutionizes the domain, making it more precise and effective.”