Leading Breast Radiologists Discuss the Recent Lancet Study on AI and Interval Breast Cancer

Recently reported results from the randomized controlled Mammography Screening with Artificial Intelligence (MASAI) trial suggest that adjunctive AI may provide a viable role in mammography triage with researchers noting significant decreases in interval breast cancer rates in a cohort of 105,915 women.

In interviews with Diagnostic Imaging, Wendie Berg, M.D., Stamatia Destounis, M.D., and Amy Patel, M.D., shared their thoughts on the MASAI study findings that were recently published in the Lancet.

Q: Is there anything about the Lancet study that you found particularly striking or surprising?

Dr. Berg: This is an excellent use case for AI — replacing a human double reader for mammography in Europe. The results were favorable by every metric — fewer invasive interval cancers, improved cancer detection, and no increase in false positives. Notably, the benefits were observed in both dense and non-dense breasts.

Dr. Destounis:I was interested to see that the sensitivity was higher in the intervention group (with AI) versus the control group (two radiologists), which was an effect that was consistent across age and breast density for invasive cancer. This is important for Europe as they may consider discontinuing their traditional double reading given workforce shortages experienced everywhere.

Dr. Patel: I found it particularly striking that in their study, AI-supported mammography screening achieved higher cancer detection rates, 12 percent fewer interval cancers, and 16 percent fewer invasive cancers. It also demonstrated high sensitivity for detecting small, lymph node negative invasive tumors, suggesting earlier detection as well. There was also comparable specificity as the standard double reading performed by the radiologists in this study.

Q: Is there anything that you would caution readers about when assessing the study results?

Dr. Berg: Mammograms are not routinely double read in the United States, and we collectively (patients and providers) are not yet ready to accept AI interpretation without oversight by a radiologist. There is currently no reimbursement for the costs of implementing AI, so the challenge is convincing health systems to make the investment.

While this study was conducted in four centers in Sweden, which may limit generalizability of results, the Transpara software (Transpara, version 1.7.0., ScreenPoint Medical) used in this study has been widely validated across diverse populations and platforms and using both 2D and tomosynthesis mammograms. The study also only evaluated one round of screening. It is not clear if the benefits of AI will be sustained.

Dr. Destounis: This is an important and timely study. However, we must understand it is over a 20-month period only and we do not have long-term data on how AI will impact incidence of interval cancer detection or the type of invasive cancers detected in subsequent screening rounds, and whether these findings will persist and be validated over time.

Dr. Patel:This study was conducted in Sweden where they conducted a double reading approach. In the United States, we typically have single readers, therefore calling into question the generalizability of the study. More similar studies in the United States need to be performed with single readers. Additionally, the diversity of the data and potential variability in outcomes with radiologists who are less experienced readers would be other factors to take into consideration.

Q: Have you incorporated AI into your practice? If so, what kind of results/impact have you seen?

Dr. Berg: For many years, the synthetic 2D reconstruction of tomosynthesis (3D) mammograms we use employ AI to accentuate calcifications and architectural distortion, and the slabbed images reduce interpretation time. In addition to “assisted interpretation” of mammograms as in MASAI, there is also software available for breast ultrasound and for MRI interpretation. There are many other areas where we are considering AI — including risk assessment from the mammogram alone, triaging mammograms with suspicious findings to urgent interpretation, and culling the medical record for pertinent history.

Dr. Destounis: We have incorporated AI to improve patient workflow, scheduling, assess breast density, breast cancer risk, and continue our research to utilize AI to improve cancer detection. Ai can improve one’s practice in many ways.

Dr. Patel: We have incorporated AI for breast ultrasound since 2019. We have found it to be very accurate, maintaining cancer detection rates while reducing false positives. We also have been utilizing AI-powered mammography since 2022, which assists in the reconstruction of 1 mm tomosynthesis slices into 6 mm slices with 2-3 mm overlap. The AI component analyzes raw 3D high resolution data to identify, prioritize, and preserve clinically significant regions of interest. As a result, we have also been able to maintain cancer detection rates without a loss in image integrity.

Q: Is there anything else you would like to add about the study?

Dr. Destounis:The findings suggest a shift toward earlier detection of clinically relevant cancers. They found fewer aggressive or advanced interval cancers in the intervention group versus the control group, and this is very interesting and important if this trend is identified in long-term studies.

Dr. Berg is a distinguished professor and the Dr. Bernard F. Fisher Chair for Breast Cancer Clinical Science at the University of Pittsburgh School of Medicine.

Dr. Destounis is the managing partner of Elizabeth Wende Breast Care in Rochester, N.Y., and the chair of the American College of Radiology’s Breast Imaging Commission.

Dr. Patel is a clinical associate professor of radiology at the University of Kansas School of Medicine. She is the chair of the American College of Radiology’s (ACR) Radiology Advocacy Network and the Political Action Committee of the American College of Radiology Association (RADPAC).