What radiologists think about using ChatGPT and AI in breast imaging
The evolving role of artificial intelligence (AI) in breast imaging was a big topic at the Radiological Society of North America (RSNA) 2024 meeting. Health Imaging spoke with Manisha Bahl, MD, breast imaging division quality director and breast imaging division co-service chief at Massachusetts General Hospital and an associate professor of radiology, Harvard Medical School, at the conference to hear more about the breast AI sessions she participated in and how she feels about these advanced technologies.
Generative AI, including large language models (LLMs) such as ChatGPT, is being explored for various applications in radiology. Bahl noted that while her practice does not currently use generative AI in clinical settings, other departments at Mass General Brigham are leveraging AI to summarize patient visits and draft responses to patient inquiries. Bahl spoke on this at RSNA and also has been involved in research in this area.
"I think LLMs like ChatGPT have a number of potential applications. One of the more immediate applications is its use to simplify or improve the efficiency of reporting. It can be used to generate structured reports for radiologists and also to generate impressions for findings and reports. But I think that there continue to be concerns about LLMs like ChatGPT and that they're not quite ready for primetime," Bahl said.
Right now, she said the technology still requires human involvement and supervision because of the inaccuracies that she is seeing.
"Hopefully that will improve over time, and maybe one day we will reach a stage where we're comfortable with LLMs generating reports or simplifying reports without as much human or radiologist involvement. But I think at this point there's still work to be done for us to trust and to be comfortable with using LLM to help with patient care," Bahl explained.
The use of AI for decision support in breast imaging, specifically through AI-based computer-assisted detection (CAD) for screening mammography, is seeking wider adoption, but she said there are still a lot of questions about accuracy.
Concerns about large language models not being ready for radiology
Bahl was recently involved in a study of ChatGPT being used to generate BI-RADS assessment categories for breast imaging reports, but the technology fell short.
"It was fairly inconsistent with the output we were getting from the generative AI. Even when giving the exact same input, the exact same report, we would get different BI-RADS categorizations in a subset of the cases. And when we're talking about patients and patient care, those types of risks are very concerning," she said.
Bahl also cautioned about the “hallucination effect” in LLMs, where the model generates seemingly plausible, but inaccurate responses.
“If we're using LLMs to simplify radiology reports to make them understandable to patients or for using chat GPT or other LLMs to answer queries, answer patient queries, we need to ensure that it's 100% correct before deploying it in real time,” she emphasized.
AI’s impact on breast cancer detection
Another key area of AI application is cancer detection. Bahl’s team investigated AI-based CAD’s ability to reduce interval cancer rates in screening mammography. She found that AI-based CAD could have detected one-third of interval cancers earlier, which has significant implications for improving patient outcomes.
Interest in AI at RSNA 2024
The interest in AI at RSNA 2024 was palpable. Bahl’s sessions, including one on AI-based CAD and another on ethical considerations of AI, were well attended.
“I was surprised at the turnout, especially on a Sunday when many attendees were just arriving in Chicago,” she said. “During an audience poll, 60% of attendees indicated they were already using AI in their breast imaging practice—higher than I expected, though the audience was likely biased toward those interested in AI.”
Dave Fornell has covered healthcare for more than 17 years, with a focus in cardiology and radiology. Fornell is a 5-time winner of a Jesse H. Neal Award, the most prestigious editorial honors in the field of specialized journalism. The wins included best technical content, best use of social media and best COVID-19 coverage. Fornell was also a three-time Neal finalist for best range of work by a single author. He produces more than 100 editorial videos each year, most of them interviews with key opinion leaders in medicine. He also writes technical articles, covers key trends, conducts video hospital site visits, and is very involved with social media. E-mail: dfornell@innovatehealthcare.com