Women’s Imaging: Maximizing AI
By Kathy Schilling, MD
Radiology Today
Vol. 25 No. 5 P. 6

3D digital breast tomosynthesis (DBT) has been revolutionary in its ability to capture thousands of images of the breast from various angles, resulting in increased cancer detection rates and reduced patient recalls for benign situations. By pairing AI and DBT, we are seeing a more efficient allocation of physician time and focus according to exam complexity and are, therefore, providing better clinical outcomes.

New Landscapes
Whereas 2D images may result in tissue overlap confusion, 3D imaging involves viewing the breast at 1-mm thick intervals, thus making it possible to visualize small tumors that are often undetected on a 2D mammogram. In addition, 3D DBT mammograms have resulted in improved cancer detection in women with dense breasts. With 3D exams, we have the opportunity to detect cancers earlier, which often gives women more treatment options and improved outcomes. Indeed, the literature is replete with data indicating that 2D screening isn’t as effective as 3D/DBT, with one study finding that the latter uncovers 40% more breast cancers than 2D mammography.1

While the many benefits of DBT have been defined, the use of this technology results in radiologists reading significantly more images each day, as DBT typically has hundreds of images per exam—compared with four images with 2D digital mammography. This means that a radiologist may be facing a review of tens of thousands of images each day. But this multitude of images must be read and read correctly.

One fundamental issue is maintaining intense focus while reading every image—especially when we know 90% of them will be normal. Radiologists are not immune to the scourge of work-related fatigue, with research indicating that 54% to 72% of diagnostic and interventional radiologists exhibit burnout symptoms.2 Exacerbating the situation is the nationwide shortage of radiologists.

Increased Detection Rates
With AI technology, the crushing radiology workload can be vastly reduced.

If a patient is called back for additional testing, but no cancer is found, ie, a false positive, not only has the patient been unnecessarily stressed, but health care resources have been misspent. On the flip side, screening mammograms miss about 20% of breast cancers, leading to falsenegative results, delays in treatment, and a false sense of security for patients.3

AI can assist the radiologist in reducing false-positive findings as well as falsenegative findings, resulting in a reduction in the number of interval cancers—a cancer diagnosed less than 365 days after a nonactionable screening exam.

The computer-aided detection (CAD) of yesterday was only as good as the radiologist who annotated the images for algorithm training. Unlike AI, CAD software did not continue to learn. Whereas CAD stagnated, AI is continuously acquiring knowledge through deep learning technology.

An additional difference between CAD and AI is that AI not only detects potentially malignant findings but those findings are also characterized. A detected lesion will be given a Lesion Score that reflects the level of suspicion on a scale from 1 to 100. The algorithm refers back to the training data set for similar appearing lesions to determine the level of suspicion.

Additionally, the mammogram will be assigned a Case Score from 0 to 100. This reflects the level of suspicion that cancer exists on the mammogram. Physicians, therefore, can more rapidly review cases with low Case Scores and spend more time evaluating those at a higher level of suspicion.

A Promising Companion
In our practice, AI is our copilot, providing an additional layer of scrutiny for these complex images. While CAD was an important milestone made available in the late 1990s, AI is now providing us with the most advanced diagnostic and risk prediction tools to date.

Our practice is composed of nine dedicated breast radiologists with an average of 22 years of experience. In a retrospective evaluation of our facility’s experience using iCAD’s ProFound AI, we found the cancer detection rates rose from 5.77/1,000 women to 7.08/1,000 women screened after two years of use. This groundbreaking finding represents a 23% increase in cancer detection. And surprisingly, this was achieved with no increase in patient recalls.4

But how does this technology work? AI allows computers to simulate human perception, while deep learning recognizes patterns and makes predictions that are not possible with the human eye or brain. AI algorithms are trained on vast datasets of mammograms and learn to distinguish between normal and potentially cancerous findings. Ultimately, AI can help us find small areas of concern years before they would manifest clinically or be identified without the use of AI.

A Leap of Faith
The uptick in screenings, along with the pandemic “hangover”—the scarcity of radiologists—leaves us with a substantial challenge in meeting patient needs. However, the adoption of AI in breast cancer screening is gaining ground, with approximately one-third of US health care facilities having implemented the use of AI in breast cancer screening.

Our facility took a leap of faith and adopted AI just before we closed for six weeks for COVID-19. We did no screening and only saw patients with breast symptoms requiring a diagnostic exam. During this period, we had the opportunity to quickly learn the capabilities of AI in lesion detection, as the majority of our diagnostic patients had mammographic findings. Although this may have assisted in shortening the learning curve for this new technology, patience is required for new users as it may take six or more months to recognize the strengths and weaknesses of AI.

It is important to understand that currently, AI is only looking at one point in time and, unlike radiologists, does not have the benefit of prior mammograms or clinical history. At first, an increase in recalls is expected. With time, however, the radiologist gains knowledge of the capabilities of the new tool and must develop new habits in order to partner with AI. However, the considerable improvements in speed and accuracy are well worth the wait.

Assessing Risk
The first foray into personalized breast medicine was in the early 2000s when it became more widely known by physicians and patients that women with high breast density are underserved with screening mammography alone. Supplemental screening with ultrasound or MRI is often required in this population to maximize early detection.

With the advent of tools such as the Tyrer-Cuzick Risk Assessment Calculator and the Gail Model, we began to assess detailed data such as personal and family history (age at menarche, etc) and lifestyle choices (hormone use, etc) to determine patient level of risk.

It has been demonstrated, however, that these models cannot offer certainty of a woman’s true level of breast cancer risk.

Discussing Density
A new AI risk solution called ProFound AI Risk makes it possible to predict a woman’s short-term risk of developing cancer before she returns in one, two, or three years, with research indicating that it is 2.4 times more accurate compared with traditional risk models.5 This tool uses patient age, breast density, and mammographic complexity to identify a specific woman’s risk. Whereas the clinical risk models are population-based, we now have more sophisticated tools that allow us to focus squarely on each individual patient.

As the field transitions from agebased screening recommendations to more personalized screening, radiologists can focus on an individual woman and recommend a surveillance plan tailored to her personal risk. All of this information is within a woman’s mammogram, making risk assessment simple. Results include the woman’s absolute breast cancer risk score and breast cancer risk category [average, intermediate, and high].

In addition, the AI Risk tool is the first such technology specifically designed to factor in racial and ethnic backgrounds. This is vital given research indicating that Black women have an approximately 40% higher risk of dying from breast cancer.”6

The Future of Precision
AI is an unparalleled early warning system that allows us to uncover small areas of concern, often years before they manifest clinically. Indeed, never before have we been able to so precisely pinpoint an individual’s risk of breast cancer. Detection using AI is faster and more accurate, which can translate into less patient stress, improved clinical outcomes, and, ultimately, a better quality of life for the patients we serve. As for the physicians involved—hardworking radiologists—they get a bit of a reprieve and can concentrate their energies on the most complex images. Humans get tired … computers don’t. Take a break and bring in the AI.

Kathy Schilling, MD, is a radiologist at Lynn Women’s Health & Wellness Institute at Boca Raton Regional Hospital, Baptist Health South Florida.

1. WATCH: New method detects more breast cancer in screening. Lund University website. https://www.lunduniversity.lu.se/article/watch-new-method-detects-more-breast-cancer-screening. Published May 6. 2015.

2. Canon CL, Chick JFB, DeQuesada I, Gunderman RB, Hoven N, Prosper AE. Physician burnout in radiology: perspectives from the field. AJR Am J Roentgenol. 2022;218(2):370-374.

3. Mammograms. National Cancer Institute website. https://www.cancer.gov/types/breast/mammograms-fact-sheet. Updated February 21, 2023.

4. Schilling K. Real world breast cancer screening performance with digital breast tomosynthesis before and after implementation of an artificial intelligence detection system. Presented at: European Congress of Radiology Meeting; March 3, 2023; Vienna, Austria.

5. Eriksson M, Destounis S, Czene K, et al. A risk model for digital breast tomosynthesis to predict breast cancer and guide clinical care. Sci Transl Med. 2022;14(644):eabn3971.

6. Black women and breast cancer: why disparities persist and how to end them. Breast Cancer Research Foundation website. https://www.bcrf.org/blog/black-women-and-breast-cancer-why-disparities-persist-and-how-end-them/. Published February 7, 2024.