AI-supported mammography improves early detection of breast cancer

Artificial intelligence (AI)-supported mammography identifies more cancers during screening and reduces the rate of breast cancer diagnosis by 12% in the years following, finds the first randomised controlled trial of its kind involving over 100,000 Swedish women published in The Lancet journal. 

The interim safety results of the MASAI trial, published in The Lancet Oncology in 2023 [1], found a 44% reduction in screen-reading workload for radiologists. Additionally, a different early analysis of the trial published in The Lancet Digital Health [2], found a 29% increase in cancer detection without an increase in false positives. 

The full results of the trial, published today, show AI-supported mammography also reduces cancer diagnoses in the years following a breast cancer screening appointment by 12% – a key test of screening programme effectiveness. 

Our study is the first randomized controlled trial investigating the use of AI in breast cancer screening and the largest to date looking at AI use in cancer screening in general. It finds that AI-supported screening improves the early detection of clinically relevant breast cancers which led to fewer aggressive or advanced cancers diagnosed in between screenings. 

Widely rolling out AI-supported mammography in breast cancer screening programmes could help reduce workload pressures amongst radiologists, as well as helping to detect more cancers at an early stage, including those with aggressive subtypes. 

However, introducing AI in healthcare must be done cautiously, using tested AI tools and with continuous monitoring in place to ensure we have good data on how AI influences different regional and national screening programmes and how that might vary over time." 

Dr. Kristina Lång, lead author from Lund University, Sweden

Mammography screening has been associated with a lower breast cancer death rate, largely due to the early detection and treatment of the cancer. However, despite European guidelines recommending two radiologists read mammograms, some cancers still go undetected in screening. 

Estimates suggest that 20-30% of breast cancers diagnosed after a negative screen and before the next scheduled screen (interval cancers) could have been spotted at the preceding mammogram. Interval cancers are often more aggressive or advanced than cancers detected during routine screening, making them harder to treat effectively. 

Previous observational studies and interim results of this trial have found AI-supported mammography increases breast cancer detection compared with standard screening, however a key question has been if this increase in breast cancer detection translates into a reduction in interval cancers. 

Between April 2021 and Dec 2022, over 100,000 women who were part of mammography screening at four sites in Sweden were randomly assigned to either AI-supported mammography screening (intervention arm) or to standard double reading by radiologists without AI (control arm). Double reading, where two radiologists read each mammogram, is standard practice in European screening programmes.

In the intervention arm, a specialist AI system analysed the mammograms and triaged low-risk cases to single reading and high-risk cases to double reading performed by radiologists. AI was also used as detection support to the radiologists, in which it highlighted suspicious findings in the image. 

The AI system was trained, validated, and tested with more than 200,000 examinations from multiple institutions across more than ten countries. 

During the two years follow up, there were 1.55 interval cancers per 1,000 women (82/53,043) in the AI-supported mammography group, compared to 1.76 interval cancers per 1,000 women (93/52,872) in the control group: a 12% reduction in interval cancer diagnosis for the AI arm. 

Additionally, there were 16% fewer invasive (75 v 89), 21% fewer large (38 v 48), and 27% fewer aggressive sub-type cancers (43 v 59) in the AI group compared to the control arm. 

In the AI-supported mammography group, 81% of cancer cases (338/420) were detected at screening, compared to 74% of cancer cases (262/355) in the control group: a 9% increase. The rate of false positives was similar for both groups, at 1.5% in the intervention group and 1.4% in the control group. 

First author Jessie Gommers, PhD student, Radboud University Medical Centre, Netherlands, says, "Our study does not support replacing healthcare professionals with AI as the AI-supported mammography screening still requires at least one human radiologist to perform the screen reading, but with support from AI. However, our results potentially justify using AI to ease the substantial pressure on radiologists' workloads, enabling these experts to focus on other clinical tasks, which might shorten the waiting times for patients." 

The authors note several limitations including that the analysis was conducted in one country (Sweden), was limited to one type of mammography device and one AI system which might limit the generalisability of the results. Additionally, in this trial, radiologists were moderately to highly experienced, which could limit the generalisability of the findings to less experienced radiologists. Lastly, information on race and ethnicity was not collected. 

Dr. Lång says, "Further studies on future screening rounds with this group of women and cost-effectiveness will help us understand the long-term benefits and risks of using AI-supported mammography screening. If they continue to suggest favourable outcomes for AI-supported mammography screening compared with standard screening, there could be a strong case for using AI in widespread mammography screening, especially as we face staff shortages."