For several years, radiologists have been publishing research that points to a major shift coming in breast cancer screening recommendations, one that focuses more on individualized risk assessment supported by AI and deep learning.
Kuhl
In research presented Monday, Christiane Kuhl, MD, director of the Department of Diagnostic and Interventional Radiology at University Hospital Aachen in Germany, discussed leveraging the power of AI to identify breast cancer risk earlier and save lives.
This retrospective, multi-center international study, fielded with Clairity, an international mammography screening consortium and health care venture, included 318,101 consecutive bilateral 2D full field digital screening mammograms obtained from 129,498 patients between January 2007 and December 2016. Patient demographics including age and race/ethnicity were retrieved from electronic medical records and cancer outcomes were obtained from local tumor registries.
Algorithm Offers Five Year Risk Estimate For Breast Cancer
The team developed a deep learning model, Allix5, to predict the development of breast cancer within five years of the mammogram. A calibration algorithm was used to create percent probabilities of future cancer, giving patients more information about their individual risk of developing the disease.
“A deep learning model, trained and calibrated on international mammography screening consortium data resources, based on the screening mammogram alone, provides a high predictive accuracy (AUC of .75-.80) and is well calibrated for use in current standard of care clinical workflows that rely on five-year breast cancer risk predictions,” Dr. Kuhl said. “The model can be calibrated and provide five-year risk estimates that offer more personalized screening and risk reduction intervention recommendations.”
Traditional risk models based on age alone were developed with datasets based almost exclusively on white women, which leaves out many patients across the globe, according to Contance Lehman, MD, PhD, founder of Clairity.
“Most breast cancers we identify are in women with no known genetic mutation or family history,” Dr. Lehman said. “The application of AI is a really promising domain to improve breast cancer risk assessment.”
Individual risk assessment, with specific percent probabilities of developing future cancer, sounds like a great advancement, but the hardest transition will be shifting the narrative for the public, according to Dr. Lehman.
“We’ve always had our screening recommendations by age, so how do we transition to screen by risk? It is always hard to shift paradigms, but what we know is that what we’re doing right now with traditional risk modeling is failing the majority of our patients,” Dr. Lehman said. “If there was ever a group of physicians in radiology who can bring about change, it is breast imagers.”
Clairity is working with the FDA to get approval of its algorithm and Dr. Lehman said she expects the conversation to shift in the next few years about who is at risk and how to understand your personal breast cancer risk with more informed context.
“Breast cancer screening has been a ‘one size fits all’ approach for the longest time. We now have the tools, such as MRI, to improve screening for women who need it. With AI-based assessment of mammograms, we are now able to identify women in-need for screening beyond mammography,” Dr. Kuhl said. “This will be a game changer—and can help potentially eradicate breast cancer as the leading cause of cancer death in women.”
Access the presentation, “Performance of a Deep Learning Image Based 5-Year Breast Cancer Risk Model Developed Across Global Imaging Centers That Provides a Percent Probability Output for Use in Existing Clinical Workflows and Decision Making,” (M1-SSBR03-6) on demand at Meeting.RSNA.org.
