Computer-aided detection (CAD) software for mammography has been available for decades but has traditionally been hampered by a high number of false-positive marks. AI software based on deep-learning algorithms is showing promise, however, for helping to improve specificity in screening mammography and other breast imaging modalities.
Just over 40.5 million mammograms were performed in the U.S. in 2023, according to Mammography Quality Standards Act (MQSA) national statistics. Although mammogram is the most widely used screening modality, a known problem is that 9.5% of the 10% of women contacted for further testing after an initial breast cancer screening are unduly burdened by a false-positive exam. These false-positive results cause anxiety for women and lead to unnecessary further testing and costs.
"In order to reduce false-positive BIRADS assessments, the most desired improvement for imaging analysis software in mammography today is reduction in [false positives per image]," wrote the authors of a study of AI-based CAD software published in the Journal of Digital Imaging in 2019. The team from MD Anderson Cancer Center and the Keck School of Medicine at the University of Southern California analyzed the problem of false positives per image and compared the performance of software developed by CureMetrix AI to a conventional CAD software application.
Computer-aided detection (CAD) software for mammography has been available for decades but has traditionally been hampered by a high number of false-positive marks. AI software based on deep-learning algorithms is showing promise, however, for helping to improve specificity in screening mammography and other breast imaging modalities.
Just over 40.5 million mammograms were performed in the U.S. in 2023, according to Mammography Quality Standards Act (MQSA) national statistics. Although mammogram is the most widely used screening modality, a known problem is that 9.5% of the 10% of women contacted for further testing after an initial breast cancer screening are unduly burdened by a false-positive exam. These false-positive results cause anxiety for women and lead to unnecessary further testing and costs.
"In order to reduce false-positive BIRADS assessments, the most desired improvement for imaging analysis software in mammography today is reduction in [false positives per image]," wrote the authors of a study of AI-based CAD software published in the Journal of Digital Imaging in 2019. The team from MD Anderson Cancer Center and the Keck School of Medicine at the University of Southern California analyzed the problem of false positives per image and compared the performance of software developed by CureMetrix AI to a conventional CAD software application.
At the time, they found that the AI-based CAD yielded a 69% reduction in false positives per image compared with traditional CAD, including an 83% reduction in false positives per image for calcifications and 56% reduction for masses.
As a second reader?
At Imperial College London's Institute of Global Health Innovation, an Artificial Intelligence (AI) in Mammography Screening study (AIMS) has been using AI software developed by Google Health to evaluate how it can be used to improve the National Health Service (NHS) Breast Screening Programme. Starting with mammography images, the work is important for finding ways to alleviate the region's radiologist workforce shortage and staffing pressures, according to Imperial College London.
As is common in Europe, NHS currently uses a two-radiologist reader assessment of breast mammograms, three when there is disagreement. For the AI study, two views of each breast (four images) are taken to be evaluated independently either by the AI system plus a single reader or by a radiologist/radiographer as a first reader and second reader using the two-reader model. The AI system runs on all cases screened within specific study times, and women are informed and given the opportunity to opt-out through multiple avenues.
While AIMS will conclude at the end of 2024, some have said that "if noninferiority is maintained, the use of artificial intelligence technology as a second reader is a viable and potentially cost-effective use of NHS resources."
Meanwhile, a team of researchers from Sweden reported in September 2023 that using AI to replace one radiologist in a double-reading protocol could reduce false-positive cases. Another study from Sweden also found that using AI to perform an initial risk assessment of screening mammography exams could detect more cancers than double-reading, but without increasing false positives.
Important considerations
There are important considerations when implementing AI-based software for breast cancer screening. For example, an article published September 23, 2023, in Nature suggested building in a separate maintenance cost and also comparing a price-per-scan model. Setup costs for the software evaluated were considered high.
Further, radiology researchers at the University of Aberdeen in the U.K. emphasized that AI performance can be affected by different mammography systems. Therefore, AI performance and decision thresholds should be validated when applied in new clinical settings. In addition, quality assurance systems, including change management, should monitor AI algorithms for consistent performance, wrote Clarisse de Vries, PhD, and colleagues for Radiology: Artificial Intelligence in May 2023.
The researchers, who studied the Mia AI breast cancer screening software from Kheiron Medical Technologies, added at that time that the Mia false-positive rate was higher than the acceptable recall rate in the U.K., suggesting that Mia would be best used combined with human reader input, as recommended by the vendor. After threshold optimization, they found that Mia had a higher recall rate than reader 1 (13% vs 5.4%) but would have detected more cancers (277 vs 261), including those missed by routine dual reporting (47 of 138), de Vries and colleagues wrote.
Work around using AI to reduce false positives and improve breast cancer detection at the NHS runs in parallel with similar work in the U.S.
AI for DBT
When it comes to digital breast tomosynthesis (DBT), also known as 3D mammography, women have nearly a 7% less chance of receiving a false-positive exam from annual DBT than conventional digital mammography after 10 years of screening, University of California, Davis researchers concluded in 2022. Since then, at least one study has generated support for making DBT the standard of care, although the jury is still out on whether DBT is as good or better than the most common imaging modalities: 2D mammography, ultrasound, and MRI, according to the American Cancer Society.
Efforts to bring AI to DBT are also underway. In a big step, Google Health and CAD software developer iCAD signed a worldwide commercialization deal in August 2023 to integrate Google's mammography AI research model into iCAD's ProFound Breast Health Suite for use as an independent reader on DBT, in addition to 2D mammography. Google said in the announcement that iCAD will help validate and incorporate Google's technology into iCAD's ProFound Breast Health Suite for use in clinical practices.
ProFound AI for DBT provides radiologists with information, such as lesion certainty of finding and case scores, according to GE HealthCare, which has added this and a range of other AI-based apps to its integrated offerings. U.S.-based Radiology Partners, for example, is in a multiyear strategic commercial agreement with iCAD for its Breast AI Suite that provides a risk estimation based on a 2D or 3D screening mammogram.
Also, Hologic said in November that it was redefining its AI roadmap. Hologic's Genius AI software for early breast detection with tomosynthesis received clearance from the U.S. Food and Drug Administration (FDA) and is now commercially available as a 2.0 technology. In comparison with its ImageChecker CAD software, Genius AI has more than 70% fewer false-positive markings, according to the vendor.
AI in breast MRI
At University of Chicago Medicine, the academic health system expanded its breast imaging services in 2023 to include a faster and less expensive breast MRI. Using ultrafast abbreviated MRI, University of Chicago Medicine's system scans the breast every three seconds for about 10 minutes, completing the MRI in less time than conventional MRI.
Researchers at the University of Chicago have also been evaluating whether dynamic contrast material-enhanced (DCE) breast MRI is improved with an AI system, compared to conventionally available software. In this example, a retrospective clinical reader study, the diagnostic performance of breast radiologists improved only slightly from an AUC of 0.71 to 0.76 when using an AI system to differentiate cancers from benign lesions at breast MRI.
An important point made about breast screening using MRI, though, comes from policy voice the Lown Institute which cited a Harvard Medical School and Mass General Brigham study from 2022. Because of their high false-positive rates, low specificity, and lack of proven benefit, MRIs for low- and average-risk women are not recommended because the potential harms from screening outweigh the potential benefits. Additionally, Lown pointed out specifically that for every 1,000 women with dense breasts screened with MRI, 80 will have a false-positive result.
MRI may be recommended, however, for women at high risk for breast cancer because of family or personal history or when images are unclear. Professor of radiology at the NYU Grossman School of Medicine Linda Moy, MD, and her colleagues at NYU Langone Health’s Perlmutter Cancer Center are testing an AI system for detecting breast cancer in DCE-MRI and significantly reducing unnecessary biopsy referrals and follow-up exams that result from DCE-MRI. This is in addition to their efforts to study AI-based decision support on sonographic breast lesion assessment and AI for triaging breast ultrasound exams.
AI ultimately moving multimodal
In yet another example of testing AI for improving mammography screening performance, a team from NYU Langone is also evaluating the added value of screening breast ultrasound. Ultrasound is used to supplement breast cancer screening mammography for women who have dense breast tissue, according to the researchers.
In a recent proof-of-concept study for a multimodal AI system, NYU Langone researchers added ultrasound to full-field digital mammography and DBT to measure added value in detecting breast cancer. The model was built using over 2 million exams in nearly 325,000 patients.
"Single modality AI systems have very high accuracy for detecting cancer in digital mammography, DBT, and also in ultrasound, but single modality interfaces really do not reflect how we interpret mammography and ultrasound together in clinical practice or, particularly, in diagnostic workup," explained Laura Heacock, MD, a breast imaging specialist and associate professor in the department of radiology at NYU Grossman School of Medicine, during RSNA 2023. "Adding ultrasound to mammography improved the AI screening performance from a sensitivity of 60-70% in dense breasts. But even in nondense breasts, adding ultrasound improved the screening performance to nearly 78%."
The study demonstrated the potential of multimodal models in breast imaging, and the NYU team is currently conducting an external validation. They also plan to test the systems' diagnostic performance and then incorporate MRI to see if it strengthens results.
Bottom line
Currently, the majority of AI tools authorized by the FDA for use in breast cancer screening assist with improving the accuracy of cancer detection, noted Conant and Elizabeth McDonald, MD, PhD, in another article in Radiology: Artificial Intelligence that was published November 2023.
"Additionally, improved accuracy leads to greater overall specificity and fewer false-positive findings, meaning that even general detection tools can potentially assist with reducing unnecessary intervention,” they wrote.
In a sign of the level of interest in advancing mammographic diagnosis using AI, RSNA's Screening Mammography Breast Cancer Detection AI Challenge in 2023 attracted a whopping 2,146 competitors.
“We expect that the dataset and the work of the contestants will provide an ideal foundation for rapid advance in breast imaging AI," said John Mongan, MD, PhD, a professor of radiology at the University of California, San Francisco and chair of the RSNA Machine Learning Steering Committee.