Clinical applications of breast AI have some leading radiologists in Europe rethinking the double-reader requirement, possibly sooner rather than later, for some digital mammography screening.
With an eye on precision medicine, traditional European breast cancer screening models are being challenged, and changes can be expected to influence women's healthcare in the near future, experts say.
In addition, AI-based image analysis is expected to find an important role not only in screening mammography but also in supplemental screening modalities such as breast ultrasound, digital breast tomosynthesis (DBT), breast MRI, and, in the future, contrast-enhanced mammography (CEM).
Clinical applications of breast AI have some leading radiologists in Europe rethinking the double-reader requirement, possibly sooner rather than later, for some digital mammography screening.
With an eye on precision medicine, traditional European breast cancer screening models are being challenged, and changes can be expected to influence women's healthcare in the near future, experts say.
In addition, AI-based image analysis is expected to find an important role not only in screening mammography but also in supplemental screening modalities such as breast ultrasound, digital breast tomosynthesis (DBT), breast MRI, and, in the future, contrast-enhanced mammography (CEM).
Paradigm shift
The challenges to the existing breast cancer screening programs are partly rooted in the call for individualized, personalized screening. Research at the European Institute of Oncology (IEO) in Milan, Italy, bears this out.
Dr. Filippo Pesapane, PhD.
Radiologists there are challenging the inertia of the current breast cancer screening paradigm, which is still largely based on the age-driven, "one-size-fits-all" model introduced in the 1970s, IEO radiologist Dr. Filippo Pesapane, PhD, told AuntMinnieEurope.
In a paper published in the European Journal of Radiology, Pesapane et al highlighted that evolving paradigms of breast cancer screening increasingly involve AI-assisted image analysis across various supplemental imaging modalities -- including breast ultrasound, DBT, breast MRI, and, potentially in the future, CEM.
While Italy’s national screening program still relies on 2D mammography, Pesapane noted that DBT has already been implemented in certain regions and clinical settings.
“For instance, at our institute -- the European Institute of Oncology -- DBT is routinely performed in all patients undergoing mammography,” he explained. Conversely, CEM remains largely limited to preoperative staging and has yet to be adequately evaluated for use in population-based screening programs. However, as a first exam, mammography is enough, according to Pesapane.
"While mammography has a very high false-positive rate, in my opinion introducing CEM imaging for screening could increase overdiagnosis," he said.
Ultrasound, rather than dedicated breast MRI, is used for supplemental screening, primarily due to the cost and economics of MRI compared with ultrasound, Pesapane noted.
"Having CEM would be a very good alternative to breast MRI, for increasing quality of the management with a suspicious or actual breast cancer," Pesapane continued. “We can imagine a near future where before having the surgery, all patients can have a staging exam with the contrast.”
Transparency
Pesapane's research at the IEO began raising awareness about the introduction of AI five years ago, with surveys of women in the population.
"Perceptions of AI were very positive for those with a lower scholar degree ... people with a higher degree said I don't trust in it," Pesapane said. "A crucial point is to understand the perceptions and worries of women, because if we want to introduce AI into the diagnostic process, we have to explain it to them in the right way."
Transparency is essential, said Pesapane, who consults with the Italian government about the best way to introduce and experiment with different AI tools in screening mammography.
Pesapane anticipated that AI support will be introduced into 2D screening mammography (as a third reader, not as a substitute for radiologists) in some regions of Italy within one year for evaluation and, depending on the results, possibly as a second reader or substitute for one reader.
"There's a willingness to introduce the technology into regional and national mammography screening,” he said. "AI is gaining momentum through isolated pilot projects, but there's still no unified national roadmap for its integration. Nevertheless, the shortage of radiologists, particularly in high-volume centers, is creating pressure for technological solutions -- making AI a likely driver of near-term transformation, especially if supported by European-level standard-setting." — Dr. Filippo Pesapane, PhD, European Institute of Oncology, Milan, Italy
“The different regions will decide different strategies to adopt AI," he said.
Pesapane emphasized that future advances in AI -- particularly those enabling risk-based, personalized screening -- and the development of novel imaging technologies will address the inherent limitations of mammography-based screening, especially for women with dense breast tissue or elevated breast cancer risk.
“This is particularly relevant in Italy, where breast cancer screening remains fragmented, with substantial regional disparities in program implementation and access,” he noted.
Echoing the stance of many breast imaging specialists, Pesapane advocates for a transition toward personalized, risk-stratified screening models, with AI potentially acting as a key enabler of this paradigm shift.
Training and education
However, Pesapane also raised a note of caution moving forward, saying that radiologists involved in breast cancer screening must have a comprehensive understanding of the inherent limitations tied to diverse imaging methodologies, along with a profound grasp of the intricacies underpinning positive and predictive values.
From radiologists to technologists and patients, education will become a critical focus as AI plays a larger role in screening mammography.
"It seems important to increase AI CAD-specific training for radiologists," stated Dr. Karen Dembrower, PhD, of the Karolinska Institute, and colleagues as part of a secondary analysis of the large-scale, prospective AI ScreenTrustCAD trial in Sweden.
After ScreenTrustCAD met its primary end point -- demonstrating that AI computer-aided detection (CAD) plus one reader decreased workload, increased cancer detection, and recalled fewer women without cancer -- Dembrower's team focused on understanding the recall decisions made in the two-radiologist consensus discussion in an AI CAD-integrated screening workflow.
Importantly, consensus discussion recalled a smaller proportion of participants if their digital mammograms were initially flagged by AI CAD versus by a radiologist (4.6% vs. 14.2%; p < 0.001). Of note, among the participants whom the consensus discussion decided to recall, the cancer yield was several times higher when the examinations had been initially flagged by AI CAD compared with one radiologist (22% vs. 3.4%; p < 0.001), according to results.
Overall, the secondary analysis of ScreenTrustCAD raised concerns that, in an AI CAD-integrated screening workflow, radiologists tended to agree with the AI software too much on cases that were erroneously flagged, or too little when the software had correctly flagged the study.
For radiologists, "a potential action might be to use continuous monitoring and periodic review sessions where radiologists revisit AI CAD-flagged cases to form an intuitive understanding of AI CAD decisions," Dembrower noted for the team's report in Radiology.
Looking forward, ScreenTrustCAD will address interval cancer with an analysis, after the planned 24-month follow-up period for all included individuals.
Integrating AI predictable
Prof. Dr. Christiane Kuhl, PhD.
Those with strong opinions about integrating AI into radiological practice include German radiologist Prof. Dr. Christiane Kuhl, PhD, director of the Department of Radiology at Aachen University Hospital and vice president-elect of the European Society of Breast Imaging (EUSOBI). With different scenarios coming into focus, is there a movement toward replacing radiologists reading low-risk screening mammograms, completely by AI?
"To the best of my knowledge, no country so far even considers it," Kuhl told AuntMinnieEurope. However, integration of AI into mammographic screening programs is predictable, she said.
"In mammographic screening, among 1,000 women who participate, typically only 5 to 7 women are found to have cancer," Kuhl explained. "Reviewing hundreds of negative mammograms, i.e., not to find a cancer when one is in a search mode, is literally an inhuman task that will and should better be done by an algorithm."
"We already know that AI as second reader outperforms radiologists," Kuhl continued. "Thus, it is rather a matter of implementation and women’s acceptance to introduce autonomous AI for mammographic screening."
Risk-stratified screening
A much more interesting application of AI is not to interpret mammograms but instead to predict the risk of subsequent breast cancer, Kuhl added, predicting that AI will make a significant impact on risk-stratified screening.
"AI-based analysis of mammographic features -- like breast tissue architecture and texture -- can predict the risk that that individual will develop mammography-occult breast cancer within the next two or five years, with an accuracy that is higher than the accuracy so far attainable, for instance, by comprehensive history taking, like with the Tyrer-Cuzick model, and even more accurate than genomic testing,” Kuhl said.
"And, unlike comprehensive history taking where you need one or two office visits to really get the entire pedigree and ask for information that many women, in my experience, aren't even able to provide, AI-based texture analysis is indeed scalable to population-wide screening programs," Kuhl added
On the one hand, risk-stratified screening will determine screening intervals. On the other, AI will determine when screening beyond mammography -- using MRI -- is warranted.
The randomized clinical trial ScreenTrustMRI showed that the group of women who underwent MRI had a cancer detection rate of almost 64 per thousand, Kuhl said.
"This is a degree of enrichment of a cohort with a mammography-negative finding that has never been seen before, about four times higher than the additional cancer detection rate that was reported in the [Dense Tissue and Early Breast Neoplasm Screening] DENSE trial, when you select women based on their breast density, and concentrate on the 6% to 8% of women with extremely dense breasts," she said.
Breast density has so far been considered the most important aspect of risk stratification for mammographic screening because it increases the individual risk for the development of breast cancer in women, and it reduces the likelihood that a given cancer can be seen by mammography due to the masking effect of dense breast tissue, according to Kuhl.
Adaptive imaging
Swedish radiologist Dr. Fredrik Strand, PhD, senior specialist in radiology at Karolinska University Hospital in Stockholm, Sweden, suggests that adaptive imaging -- an AI-directed protocol adaptation with breast MRI -- will one day apply to a select subset of women at high risk for breast cancer.
Strand, who led ScreenTrustMRI, called results from research in adaptive breast MRI scanning using AI “an important step toward adaptive, efficient, and patient-centered imaging workflows.”
In this scenario, an AI tool would triage MRI exams to determine in real-time whether to continue scanning beyond the initial MRI sequences. Such a tool would reduce MRI examination time and resource use.
In an editorial, Strand described the AI-assisted imaging strategy as dynamic, personalized protocol adaptation in a feedback-driven workflow.
"There is also an opportunity to explore how AI could be used to personalize not just protocol duration, but other acquisition parameters, such as spatial resolution, temporal dynamics, and even k-space trajectory selection," Strand explained. "Feedback-driven workflows would embody the full promise of AI in radiology -- not just automating existing tasks but reimagining them entirely."
The research and commentary published in June add to previous research into abbreviated MRI (vs. conventional MRI) in breast cancer screening.
Kuhl, whose own research on abbreviated MRI for breast cancer screening was first published over a decade ago, agrees with this approach. However, to AuntMinnieEurope she underscored that the utility of AI tools to further shorten acquisition protocols in screening settings remains to be seen given the substantial evidence that exists on short protocols that are limited to 120 seconds acquisition time already.
Questions remaining
The key is figuring out how to implement the AI, said Dr. Katja Pinker, PhD, division chief of breast imaging in the department of radiology at Columbia University Medical Center in New York City. With years of practicing radiology in Austria, Pinker still serves as an executive board member of the European Society of Breast Imaging (EUSOBI).
Europe's two-reader interpretation requirement makes AI implementation difficult in breast cancer screening, Pinker said. Should AI become a third reader? Should AI replace one reader? Should both humans be aided by AI? How will this impact recall rates?
With so many unresolved questions, the ESR AI Working Group has outlined key recommendations to help hospitals and developers navigate the new and uncertain AI landscape, saying that AI education must be integrated at all levels, from radiology and radiographer curricula to hospital-level AI training programs.
Likewise, how to monitor and assure quality over time will take shape as the EU has begun rolling out its European quality assurance (QA) scheme. QA in mammography has varied widely between countries. This finding, based on a 2022 survey conducted by EUSOBI, is expected to be revisited by 2027. The program has laid out quality requirements for participating breast cancer services such as mobile screening units, hospitals, cancer centers, and clinics.
Other key questions have also been raised. Will radiologists ultimately act as supervisors of AI-driven workflows, or will they retain primary control over image acquisition and interpretation? These are not merely technical questions but also professional and ethical ones, Strand pointed out.
While the optimal mode of AI implementation in the workflow is still unknown, it is likely that radiologist-to-AI interactions in breast cancer screening programs will continue profoundly throughout Europe.
As evidence comes in from more prospective studies involving large cohorts, European radiologists will gain clarity both on the most suitable ways to integrate AI CAD into everyday clinical workflows and on how to monitor the AI for safety and accuracy as their practices evolve.