Women’s Imaging: No One Left Behind
By Susan Harvey, MD, FSBI
Vol. 24 No. 7 P. 6
Mammography is a critical tool in mammographers’ screening and diagnostic imaging toolkits. However, with recent advancements in the technology has come a significant increase in the number of images generated for interpretation. Digital breast tomosynthesis (DBT), sometimes referred to as 3D mammography, is the latest advancement in breast cancer screenings. It has been shown to support improved accuracy compared with 2D mammography for women across a variety of ages and breast densities.1,2 DBT, however, produces significantly more images than 2D mammography.
AI can assist radiologists with interpreting the increased volume of DBT images and identifying breast cancers more accurately, improving outcomes for women who receive mammographic screenings for breast cancer. At this time, deep learning software does not replace radiologists in reviewing these 3D images but, rather, augments their expertise with a wide range of benefits, including improving the accuracy and speed of interpretation.
As AI-powered breast imaging becomes more widely available and accepted, breast imaging centers and health care institutions need to evaluate whether the software will elevate radiologist interpretation accuracy and patient outcomes. Adoption and consideration must factor into how implementation could impact breast health inequality for underserved women.
AI has the potential to help reduce health disparities across the United States, where patients might have challenges in accessing breast imaging specialists in the locations where their annual screenings either take place or are interpreted. It can also help less experienced radiologists become more comfortable with breast screenings and assist in identifying areas of concern for earlier detection. AI may bring more accuracy and improve early detection by decreasing the variation of interpretation skills that exist in breast cancer screening programs in the United States.
However, implementation of AI-powered breast imaging also has the potential to exclude some patients, especially underserved women. Health care inequity among this population is a significant concern, as AI could further widen this already enormous gap when looking at access to AI, as well as accurate interpretations, latest technologies, and the most experienced breast cancer care team. Providers must consider how AI implementation occurs so these patients who already struggle with barriers to screenings are not left behind even further by this new technology.
When AI software is integrated into DBT and 2D images, deep learning technology can augment radiologists’ interpretation to aid in detection performance.3 For breast imaging centers and institutions, AI can be a useful tool for standardizing the accuracy of imaging interpretation. There can be tremendous variation in skill level among breast radiologists just entering the field or general radiologists who might have less experience reading breast images, and AI can supplement those skills for more accurate identification of lesions while improving their interpretation time.
This will have growing importance as more imaging centers and institutions serving underinsured or uninsured women continue to adopt DBT. With the FDA’s updated Mammography Quality Standards Act requiring patient notification of breast density, the implementation of DBT will likely expand in the years ahead. This is good news for patients, as DBT systems have been shown to not only detect 20% to 65% more invasive cancers than a 2D mammogram alone,2 but also to reduce false positive rates compared with full-field digital mammography.4 By utilizing AI for DBT image interpretation, radiologists at centers and institutions can feel more comfortable using the systems that may be new to them and improve their accuracy and patient outcomes at the same time—possibly decreasing the reading time of images from this latest technology.
Experienced radiologists can also benefit from AI software tools. Breast imaging centers and institutions, with an influx of women returning for screening following the pandemic and new US guidelines for women in their 40s, can experience an overwhelming number of images needing review. AI can help improve workflow and reduce interpretation time to help radiologists move through images more quickly—without their accuracy suffering.
By adopting AI software, breast radiologists can experience benefits at every level of their careers. By assisting with training, setting a more consistent standard, and improving interpretation accuracy, deep learning technology can help centers and institutions improve the care they provide patients, no matter the location.
While AI offers many benefits to improve breast health disparities by supporting centers and institutions serving underinsured and uninsured patients, there are several challenges that need to be overcome to help ensure that women of diverse backgrounds are accessing the technology and receiving the most accurate results.
If AI follows the same trends as DBT for adoption, women who are underinsured or uninsured might not have access to the technology. Black women remain less likely to be screened via DBT than their white counterparts,5 and since AI for DBT utilizes the multiple high-resolution images of these mammography systems, it also means access to the deep-learning software may be limited, based on a lack of IT infrastructure at some locations.
Additionally, AI tools are minimally covered by health insurance providers, if covered at all. For women who are insured by Medicaid or who are not insured, this technology could be out of reach due to economic barriers. These women face significant challenges to mammography and breast cancer care already, from language barriers to transportation to childcare, and by adding AI, they may have limited (if any) opportunities to access this technology.
Even when underserved women and women of color have access to AI, there is another challenge: understanding the racial and other diversities among the patient images used to train the AI tool. For example, Black women have denser breast tissue than white women,6 making lesions more difficult to see on mammograms. Black women are also more likely to have more aggressive, more advanced-stage breast cancers,7 which compounds the need for accurate interpretation to identify all cancer sites in both breasts and common areas of axillary nodal spread. Yet data for AI software development and locations where the technology is tested do not always reflect these nuances of diverse patient populations in which they will be used.
The accuracy of AI for mammography is impacted if women of diverse racial and ethnic backgrounds are not integrated into the dataset of images that the software learns from and tests against. For women with a higher risk of cancer from dense breast tissue, their cancers could be harder to identify—and without a diverse pool of data to learn from, AI might miss these lesions because the technology is only as accurate as its training datasets.
Radiologists must be set up for success, no matter the demographics of the communities they serve, and AI could help support this—but only if we address the challenges alongside the innovation. When exploring the integration of AI technology, breast imaging centers and institutions will need to consider how their software is trained to best support their patients and deliver accurate results, as well as how they choose to implement it to reach all the populations they serve. This deep learning technology has the potential to help us provide more accurate results and better outcomes for our patients, but it needs to be trained and tested appropriately.
— Susan Harvey, MD, FSBI, is the vice president of corporate global medical affairs for Hologic. She is also a member of the Miller Coulson Academy of Clinical Excellence and was the director of breast imaging for Johns Hopkins Medicine in Baltimore for more than seven years.
1. FDA submissions P080003, P080003/S001, P080003/S004, P080003/S005.
2. Friedewald SM, Rafferty EA, Rose SL, et al. Breast cancer screening using tomosynthesis in combination with digital mammography. JAMA. 2014;311(24):2499-2507.
3. FDA clearance K201109.
4. Destounis SV, Morgan R, Arieno A. Screening for dense breasts: digital breast tomosynthesis. AJR Am J Roentgenol. 2015;204(2):261-264.
5. Alsheik N, Blount L, Qiong Q, et al. Outcomes by race in breast cancer screening with digital breast tomosynthesis versus digital mammography. J Am Coll Radiol. 2021;18(7): 906-918.
6. Rochman S. Study finds Black women have denser breast tissue than white women. J Natl Cancer Inst. 2015;107(10):djv296.
7. Breast cancer risk factors you cannot change. American Cancer Society website. https://www.cancer.org/cancer/breast-cancer/risk-and-prevention/breast-cancer-risk-factors-you-cannot-change.html. Updated December 16, 2021.