Q&A: AI-supported breast cancer screening increases detection rate, reduces workload

Key takeaways:

• AI-supported breast cancer screening led to increased early cancer detection and reduced workload for radiologists.
• Many women seem to feel positively about the use of AI in breast cancer screening.

AI-supported breast cancer screening increased early-stage cancer detection while reducing radiologists’ workload without increasing false positives, but more research is needed before clinical implementation, researchers reported.

“There has been a lot of hype around AI in health care for quite some time, and there has been a lot of papers evaluating the tools with retrospective data. When it comes to breast imaging and mammography, radiologists have been on the front line evaluating these tools due to the availability of mammograms and the capability of algorithms to analyze the images,” Kristina Lång, MD, PhD, a breast radiologist, clinical researcher and associate professor in the division of diagnostic radiology in the department of translational medicine at Lund University in Sweden, told Healio. “We also have an obvious, large possible application, and that is in the screening program.”

Healio previously reported that an AI algorithm may help to identify breast cancer on MRI 1 year earlier than current screening methods, according to a study published in Academic Radiology. In this study, researchers noted that radiologists could utilize this AI algorithm to identify those at low and high breast cancer risk.

In the Mammography Screening with Artificial Intelligence (MASAI) trial, a randomized controlled, parallel-group, noninferiority screening accuracy study published in The Lancet Digital Health, Lång and colleagues recruited 105,915 women (median age, 53.7 years) from four screening sites in southwest Sweden from April 2021 to December 2022. All women were eligible for mammography screening and randomly assigned to AI-supported screening (n = 53,043) or standard double reading (n = 52,872).

AI-supported breast cancer screening resulted in 338 detected cancers and 1,110 recalls while standard screening resulted in 262 detected cancers and 1,027 recalls. Cancer detection rates were higher for women in the AI-supported screening group vs. the standard group (6.4 vs. 5 per 1,000; proportion ratio = 1.29; 95% CI, 1.09-1.51; P = .0021). In addition, invasive cancer detection increased with AI-supported vs. standard screening (270 vs. 217 per 1,000; proportion ratio = 1.24; 95% CI, 1.04-1.48).

Healio spoke with Lång about the MASAI trial findings and the potential benefits of AI-supported breast cancer screening.

Healio: What were the goals of the MASAI trial?

Lång: What I wanted to achieve by using the AI support was to see if we can reduce the workload for radiologists. Many countries, such as those in Europe and Australia, use double reading. In doing this triage with the AI tool, we can then reduce the workload for radiologists, which is much needed since there is a workforce shortage in many countries. Most important for me was to see if we can improve the efficacy of the screening program to detect more cancers at an early stage and to be able to detect the relevant cancers early.

We have published two papers so far. The first goal was to see if it is safe to screen with AI, since this was one of the first randomized controlled trials assessing AI and screening mammography. We found that we detected more cancers even though we reduced the workload. In the second paper, we want to see how many cancers we detected, how many women were recalled and how many false positives we generated. We also wanted to characterize the types of cancers that we detected.

Healio: Compared with standard double reading, how did AI-supported screening perform?

Lång: We detected 29% more cancers using AI support, and we had to recall 8% more women, but most of the extra recalls resulted in cancers. We only had seven more false positives using AI, and the intervention arm included 53,000 women. We can increase the detection rate without increasing false positives and, at the same time, reduce work that we need to do as radiologists.

We also conducted analyses by breast cancer type. We found that this large increase in detection was mainly for small invasive cancers. There are many types of breast cancers, from low-risk to very aggressive types. We were able to detect breast cancers that are more aggressive during an early stage before they had metastasized. We also detected more in situ cancer before they become invasive.

Healio: How did the patients feel about AI integration into their mammography?

Lång: We were planning on conducting a survey for the women attending the screening program to gauge what their perception was, but we never did it because there were other studies coming out in different settings. The results from different survey studies shows that women are positive toward the use of AI. What women have said is they want to have a human in the loop. They do not want AI to be a stand-alone tool. That is how we designed our screen-reading protocol; there was always a radiologist in the loop. We had very few women who opted out of the trial. After we published the results, many women have asked, “Are you using AI now?” We have not yet implemented an AI screening program. We are working on that right now.

Healio: Do you believe it will be easy to integrate AI into mammography as a new standard?

Lång: When it comes to radiology, we are already very digitalized. Everything is computer-based. With the Picture Archiving and Communication System (PACS) that we use in radiology, you can easily adopt an AI tool. There are also several commercial algorithms that have been shown to work well.

First, we have two things we need to figure out. One, when you implement AI you need to monitor the algorithm in clinical practice and there are solutions for that. Then, you have to adapt when you perform a quality assurance of your screening program, measuring different indicators and how well AI works. When the algorithms are updated, we also need to address that.

How you implement AI in the clinic matters. We did one screen-reading protocol using AI, but you can do it in many different ways. There have been some results from a few prospective trials that all used different screen-reading protocols. Those have demonstrated a modest increase in cancer detection of about 4% to 6%.You have to pick the way you implement AI in a very careful way.

Healio: What other research related to AI and breast cancer screening is needed?

Lång: This is the first randomized trial. There are no results from others, and typically, if you want robust evidence, the findings must be reproduced in other settings and in other populations.

A limitation of our trial is that it was conducted at a single site, in a Swedish setting. Populations differ, and even if we have a large immigrant population, it is important to reproduce the results in other settings. Screening programs also differ, with different age and screening intervals, which reproducibility even more important. The development of AI is moving fast. We will likely see that AI is being implemented in screening programs soon. It likely won’t be long before AI-enabled screening becomes part of our guidelines.

In our trial, the primary endpoint was interval cancer rate. Hopefully we will see that we don’t have more interval cancers after AI screening. That would be surprising. This is that piece of evidence that we need to be able to really shout. It is important to see what happens in the next screening round and see if the incidence of advanced cancers is lower after screening with AI. We need some long-term follow-up studies to see the clinical impact.

References:

• Hernström V, et al. Lancet Digit Health. 2025;doi:10.1016/S2589-7500(24)00267-X.
• Lång K, et al. Eur Radiol. 2021;doi:10.1007/s00330-021-07686-3.

For more information:

Kristina Lång, MD, PhD, can be reached at kristina.lang@med.lu.se.