Breast cancer screening programs in Europe face many challenges today. Double reading reduces recall rates and increases cancer detection, but many experts regard it as unsustainable, given the workforce crisis facing many countries. However, the use of AI promises to transform the status quo and open new possibilities.
Support is now growing for single-reader models and advanced imaging technology in some scenarios.
AuntMinnieEurope spoke with leading breast imaging researchers to get their perspective on what's working in AI presently in breast cancer screening in Europe, where screening programs generally rely on 2D mammography and adhere to the double-reader policy.
Breast cancer screening programs in Europe face many challenges today. Double reading reduces recall rates and increases cancer detection, but many experts regard it as unsustainable, given the workforce crisis facing many countries. However, the use of AI promises to transform the status quo and open new possibilities.
Support is now growing for single-reader models and advanced imaging technology in some scenarios.
AuntMinnieEurope spoke with leading breast imaging researchers to get their perspective on what's working in AI presently in breast cancer screening in Europe, where screening programs generally rely on 2D mammography and adhere to the double-reader policy.
Radiologists in Germany's national screening program face a heavily repetitive task, according to Prof. Dr. Christiane Kuhl, PhD, director of the Department of Radiology at Aachen University Hospital and president-elect of the European Society of Breast Imaging (EUSOBI).
"Not finding breast cancer reading hundreds of mammograms is an inhuman task," she told AuntMinnieEurope. "[However,] using AI for interpretation is not encouraged in the German mammography screening program and, in particular, is not accepted to replace radiologist reading."
Mainly based in private practices, mammographic screening in Germany is expected to increase with a recent policy change and a greater number of women now eligible -- approximately 11 million, according to the Germany Mammography Screening Program.
Challenges
Prof. Dr. Christiane Kuhl, PhD.
"Regulatory issues make modern technology difficult to implement in Germany," Kuhl said, adding that the obvious use is for AI to be used as a second reader, at least, where cases are predictably all negative.
"In mammographic screening, among 1,000 women who participate, typically only 5 to 7 women are found to have cancer," Kuhl explained.
"We already know that AI as second reader outperforms radiologists," she continued. "Thus, it is rather a matter of implementation and women’s acceptance to introduce autonomous AI for mammographic screening. Algorithms do not get tired, frustrated, or bored, unlike human beings.
"I believe, therefore, that autonomous AI can and should be used to replace radiologists to interpret screening mammograms altogether," Kuhl noted. "In that scenario, radiologists would only get to see the positive screening findings -- just as they do for the diagnostic assessment of radiologist-read screening mammograms. Radiologists should embrace this.”
At Aachen, advanced imaging techniques are used and used successfully with AI in some situations, she said. AI can not only be used to help read mammograms, but also to predict the risk of subsequent breast cancer.
Supplemental imaging -- tomosynthesis for screening, plus MRI for screening women with dense breasts and/or other risk factors -- is offered for women at increased risk but not available to women at large who receive an invitation for a screening exam through the population-based screening program.
PRAIM study
Published in 2025, one of the largest studies that yielded an AI model-based breast cancer detection rate and recall rate is PRAIM (PRospective multicenter observational study of an integrated AI system with live Monitoring) in Germany. PRAIM involved nearly 500,000 asymptomatic women ages 50 to 69.
Prof. Dr. Alexander Katalinic.
With 260,739 participants in the real-world AI implementation group, PRAIM compared the performance of AI-supported double reading to standard double reading (without AI), among those undergoing organized mammography screening at 12 sites between 2021 and 2023.
Using AI-supported double reading, radiologists spent 43% less time overall interpreting examinations tagged as normal, with a mean reading time of 39 seconds for normal examinations compared with 67 seconds for examinations not tagged as normal, reported principal investigator Prof. Dr. Alexander Katalinic of the University of Lübeck and University Medical Center Schleswig-Holstein, Germany, and colleagues as part of a January 2025 follow-up published in Nature Medicine.
"The decision referral approach used in our study allowed for improving the [breast cancer detection rate] without increasing the recall rate through a combination of a ‘safety net’ system and ‘normal triaging,’" wrote Katalinic and team.
"Radiologists using the AI-supported viewer were only alerted and shown suspicious computer-assisted diagnosis marks after they interpreted examinations deemed suspicious by the AI as normal," they explained. This approach limits automation bias and reduces false-positive recall rates while leaving the final recall decision to the radiologists, according to the group.
Use of AI in Germany's screening program is becoming increasingly common, Katalinic told AuntMinnieEurope.
"The national (statutory) cancer screening guidelines are over 20 years old, but they do not prohibit additional examinations as part of the screening program," he explained. "Many radiologists take advantage of this and already use AI in official mammography screening. This is remarkable because they do this without additional financing. However, work is underway to amend the guideline to include AI (this would also be the basis for reimbursement)."
Furthermore, “as our study has shown, the use of AI leads to improved detection rates," Katalinic said. "Although this is a nice result, it hardly changes the workload for radiologists. Here, there was a shift in the reading times. Normally tagged images were read faster. More time was spent on conspicuous ones."
"If we, at least for normal tagged images, dispense with the human double reading and replace it with a human single reading combined with AI, this would massively reduce the radiologists' workload," he continued. "Simulations reveal a potential of up to 57% reduction in workload, if normal images are not read by a human. This is an important point for Germany, because with the extension of the age limits, even more women will be eligible for screening, and the workload will increase."
Based on the available scientific evidence (PRAIM and [Mammography Screening with Artificial Intelligence] MASAI study), Katalinic said it is justifiable to dispense with human double reading when using AI.
The Netherlands
Kicky van Leeuwen, PhD.
In the Netherlands, about one million women are screened on an annual basis, according to a 2023 Radboud University Medical Center (UMC) estimate. The Netherlands is quite unique in the fact that it has a single, centralized screening program. Breast cancer screening takes place through mobile and several permanent mammography units.
Follow-up imaging and further analysis when needed -- using digital breast tomosynthesis (DBT), ultrasound, MRI, and biopsy -- are performed in a hospital setting where AI is currently being used to aid in follow-up assessment and breast cancer diagnosis. AI is a choice for hospitals, not a mandate.
"Hospitals individually can choose to have the AI in place to support follow-up," said Kicky van Leeuwen, PhD, who supports the responsible adoption of AI in healthcare through Romion Health and is involved in the Dutch breast cancer screening program.
Van Leeuwen has been monitoring AI product availability throughout Europe as founder of Health AI Register, a registry of commercially available AI solutions.
Currently, there are already 20 commercially available AI software applications in the European Union (CE-marked) for breast imaging. They span the modalities of 2D mammography, tomosynthesis, MRI, and ultrasound, van Leeuwen said.
European screening programs mostly use mammography, which is the modality with the majority of AI software available, currently 16.
Changing attitudes
Breast cancer screening policies in Europe have not been optimal for certain subgroups, and stakeholders have been asking if a uniform, one-size-fits-all program is best.
In March 2024, the Health Council of the Netherlands issued a position statement that AI could be applied for risk stratification and reading support to improve breast cancer screening. The council advised to begin preparing for AI implementation now.
A technical physician by training, van Leeuwen is steering the path forward.
"For triage and reading support, a lot of prospective evidence internationally, for example, from Sweden, Denmark, and Germany, has come out showing the potential to more effectively screen," she said. “When we talk about AI-based risk stratification, AI solutions may indicate, based on aspects such as breast density, characteristics of the mammogram, and age, that you would get more frequent or less frequent screening."
“So instead of having one every two years, one might go every year, and another person might only get one every three years, or someone may be referred for an additional MRI," van Leeuwen continued. "There's a lot of research going on around this topic, in the Netherlands and other countries, but to my knowledge, it is not being applied yet in Europe. There are also very few regulatory cleared AI solutions out there today that specifically support this.”
U.K. stance
Dr. Elisabetta Giannotti.
"Introducing any innovation into the NHS Breast Screening Program requires robust evidence, particularly from U.K. data, demonstrating both clinical benefit and cost-effectiveness," she explained. "This is because the screening program is nationally standardized, so any change needs to be applicable and scalable across the entire country."
This is one of the reasons traditional computer-aided detection (CAD) systems never entered routine clinical practice in the U.K., Giannotti noted.
"Most notably, the [Computer-Aided Detection Evaluation Trial] CADET trials did not show a significant increase in cancer detection but did result in more false positives," she said. "In contrast, CAD was adopted in the U.S., where the screening landscape is more fragmented and local decisions can differ."
Gaining momentum
Several promising AI systems have been developed and validated using U.K. data, according to Giannotti. A key resource is the OPTIMAM Mammography Image Database (OMI‑DB), a national archive containing almost 7 million screening images from across the U.K., widely used to train and evaluate AI algorithms, she said. Research is ongoing.
"For adoption into the national screening program, we still need large-scale U.K. evidence proving added value and cost-effectiveness within the NHS context," Giannotti said. "This is now being addressed through the [Early Detection using Information Technology in Health] EDITH trial, led by Prof. Fiona Gilbert, which has been set up and is starting, comparing three different screening arms to evaluate AI’s role in a real-world, population-based setting."
U.K. officials project a 40% shortfall in consultant radiologists by 2028. The radiologist-led EDITH trial aims to reduce the number of radiologists required to read a mammographic screen from two to one.
In the meantime, modernizing the NHS’s digital services to better engage with women is one of two key steps the country is currently taking -- using automation -- to improve breast cancer screening and detection. The other is offering women the opportunity to streamline access to care, either by booking their screening appointment directly or initiating a self-referral for symptomatic findings through an AI-assisted interface.
“The automation of screening invitations and the ability to book appointments online represent a key step toward improving access to care and efficiency, reducing no-shows, and possibly increasing screening attendance,” Giannotti said.
Radiology institutes
In Austria, population-based screening performed in private radiology institutes by dedicated breast radiologists means greater adoption of AI and DBT in breast imaging services, according to Dr. Katja Pinker, PhD, who now conducts research as division chief of breast imaging in the department of radiology at Columbia University Medical Center in New York City. Pinker serves as an executive board member of EUSOBI, and she worked previously at the Medical University of Vienna.
Dr. Katja Pinker, PhD.
Austria maintains a population-based breast cancer screening program for women ages 45 to 75, with examinations every two years. There is also still the possibility of "opt-in" for women at age 40 and 75-plus. The approach is a change from earlier days of opportunistic screening with annual screenings starting at age 40, which included a clinical breast exam, mammogram, and ultrasound with same-day results, Pinker explained.
Austria's breast cancer screening program is special in that ultrasound is incorporated and mandatory for women with ACR breast density categories C and D, according to Pinker.
"This is extremely unique within Europe," she said. "We recognized also that density is important in the sense of how it impacts finding cancer on the mammogram, on top of the increased risk, so ultrasound is formalized in the program."
Importantly, radiologists do not batch-read screening mammograms in this decentralized setting.
Personalized services
Screening may be performed in private radiology practices/institutes or in academic centers, with small private practices having been migrating into larger radiology institutions over the last decades, and many of these already use some form of AI in breast care, Pinker added.
"As an interpretation aid, AI is particularly helpful because the radiologist often reads while the person is there and is waiting for their result," Pinker explained. "But when you read the exam, you have the AI output -- that gives you diagnostic confidence. Work-list prioritization is less relevant when you are not batch reading."
While her research work concentrates on functional breast imaging with high-resolution multiparametric MRI and contrast-enhanced mammography, Pinker said AI in clinical practice is currently mainly implemented for mammography, DBT, and to some extent, ultrasound.
Silent revolution
"When we initially were talking about the implementation of AI, everybody was a little bit weary in the beginning from the sobering experience we had with the old CAD," Pinker explained. "When that came out, the idea was that reading with CAD will improve our performance, but long-term data have shown that this was not the case, but sometimes even the opposite."
"The old CAD systems were based on small amounts of data using hand-crafted features and machine learning without continuous algorithm improvements that we now know are necessary," she said. "They were reflective of the computational technology and power that we had at that time.
"Then, with the advent of deep learning, that field [AI CAD] exponentially accelerated," she said. "We were able to overcome the disappointing experiences we had with the old CAD, and since data is rapidly amounting that we can harness AI to our benefit. AI-enhanced mammography has become quickly a reality."
AI in breast cancer detection can involve so many medical imaging techniques, combinations of modalities, and algorithm-based models that it can be challenging for European radiologists to adjust to new information.
While differences in culture, imaging modality preferences, and national approaches create variability and may hinder the broader adoption of AI in Europe, both motivated and skeptical radiologists are asking: "How long will it take for AI to change the paradigm of human double-reading in breast cancer screening programs?" Time will tell.