Real-world implementation of AI in mammography.

Health Imaging spoke with Samir Patel, MD, CCD, DABR, FACR diagnostic at Radiology Inc. and Beacon Health System about the growing trends for AI in mammography in sessions at the RSNA 2023 meeting.

Patel was enthusiastic and said AI is revolutionizing breast imaging. The real-world implementation of AI in mammography is improving radiologist workflows as well as patient outcomes. 

“There are over 500 FDA approved algorithms for AI just in radiology, and one of the most common areas where AI is starting to show its effect benefit is in breast imaging, specifically screening mammography. There was a survey a couple years ago from the¬†American College of Radiology¬†asking radiologists about their utilization of AI. In 2021, two-thirds of individuals were actually not using any algorithms. But my guess is even in two years, that number has probably changed. The most common subspecialty using AI was breast imaging. I think it’s still starting to penetrate the marketplace. I would say as a field right now, it’s still in the early adopter phase. But, I feel people are moving rapidly on the technology adoption curve, a lot faster than when they moved from film to digital or digital to tomosynthesis. I have a feeling from tomo to AI-assisted tomo, that lead time is going to be less and less,” Patel explained.

  • Momentum unfolding. He said there are about 20 FDA-cleared AI applications just in breast imaging, and most of these were displayed by vendors on the expo floor at RSNA 2023.

Addressing common mammography challenges

The implementation of AI in breast imaging addresses several common challenges faced by radiologists. First, Patel said the movement to tomosynthesis mammography has rapidly created an overwhelming number of images to analyze. A number of breast imaging experts have told Health Imaging that AI will likely play a critical role in helping sort through the massive uptick in images.

“Standard mammogram would only be four pictures back in the days of film and early digital, but now with tomosynthesis, that same study can involve over 200 images. Multiply that times 50, 75, or 100 cases a day, and that is a massive volume of images to look at. The good news is now there are some algorithms out there to actually reduce those number of tomosynthesis images to be reviewed,” Patel said. 

He added that AI can help do a first pass of the tomography image dataset to flag which slices appear to have an area of interest with a possible cancer. This can reduce the number of images a radiologist needs to review.

“With physician burnout being so prevalent, some of these algorithms now actually reduce the number of images that radiologists would need to look at in order to detect breast cancers. Some algorithms reduce the number of images by up to two thirds without any loss of the ability to detect breast cancer,” Patel explained.

  • AI can triage the exams, prioritizing reports with suspected critical findings for radiologists to review promptly. This helps the reader focusing on critical cases and better plan their workflow for the day. Many FDA-cleared algorithms can automatically perform measurements of breast lesions and assess mammograms for breast density BI-RADS scores.

Patel said this helps remove the great amount of variability between readers and can help save additional time when reviewing exams, adding that AI algorithms can also be used to improve image quality by reducing artifacts, eliminating the need for manual adjustments by radiologists.

AI algorithm adoption considerations

Patel stressed the importance of strategic thinking before adopting AI algorithms. This means identifying specific problems or areas in need of improvement that should precede the adoption of an AI.  He cautioned against acquiring algorithms without a clear problem statement, considering the associated costs and the limited resources of time and money.

  • Focus on clear problems. A major concern Patel has is people getting an AI algorithm before determining there is a real problem to solve.

“When we talk about algorithms, it can potentially be overwhelming with so many algorithms. These algorithms also come at a cost, so there is a price, a dollar figure to pay. So, I think it’s important that before they start thinking about algorithms, they should think about what problem or opportunity they’re really trying to solve, or what opportunities are there for improvement,” he noted.