"I think AI is still in its relatively early phase of adoption," Lehman said. "We do have some centers that are not academic centers that are very forward thinking and really wanting to bring AI into their practices. However, we are also seeing a story that is very familiar when we are bringing computer-aided detection (CAD) into both academic and community centers. The technology is being incorporated into clinical care, but we are still studying what the actual outcomes are on patients who are being screened with mammography where AI tools are or are not being used."

This includes AI for automated detection of breast cancer lesions and flagging these to show the areas of interest on mammogram images, or to flag studies that need closer attention. AI also can take a first pass look at mammograms to determine if they appear to be normal, so radiologists can prioritize which exams need to be read first and which may be more complex. 

This technology will likely become more important as the number of breast imaging exams switches over from traditional four-image mammogram studies to much larger 3D mammogram digital breast tomosythesis (DBT) exams of 50 or more images that are more time consuming to read. AI is already being used to flag images that deserve a closer look in these datasets. 

AI is also finding use as an automated way to grade breast density to help eliminate the variation of grading the same patient by human readers. 

However, the most exciting area of AI for breast imaging is in the potential of radiomics, where the AI will view medical imaging in ways that human readers cannot to identify very complex and small patterns that will help better assess patient risk scores, or what the best outcomes will be based on various cancer treatments.

"What I am really excited about is the domain where investigators are considering the power of artificial intelligence to do things that humans cannot or are not very good at, and then to allow the humans to really focus on those tasks where humans excel. As of today, these AI tools have not even really scratched the surface," Lehman explained.

She said this area of research using radiomics moves beyond training AI to look at images like a human radiologist and to instead pull out details that are usually hidden from the human eye. This includes rapid computer segmentation and analysis of the morphology of disease or tissue patterns seen in images, looking for minute regional structures that can be detected by AI.

"This is not to train AI to look at mammograms like I do, but to train the AI to look for patterns and signals that my human eyes and human brain cannot detect or process," Lehman said. 

She said today, we are just scratching the surface of the data potential of AI analysis of cancers in imaging. Deeply embedded patterns within cancers on imaging may be able to tell us a lot about which concerns will or will not respond to different drugs or therapies. AI may be able to tell us this from a much deeper analysis of the imaging, including the subtypes of that particular cancer. This would enable much better tailored, personalized medicine and treatments for each patient.