As times change, so does technology. Humanity is continuing to find new and innovative ways to make their lives easier in new and creative ways. AI has been one of the more interesting fields to monitor as it evolves, with humans using it to both help them within the workforce and for fun. Both points were on full display when Woojin Kim, MD, chief medical information officer of Rad AI and a musculoskeletal radiologist, took the stage at ACR 2024 to give this year’s Moreton Lecture to demonstrate the evolution of generative AI, and how radiologists could start to see it more involved in their practices.
Kim started off the lecture by allowing an AI-generated avatar to give his introduction before revealing to the crowd that not only was the image generated by AI, but the voice was as well. He dove into AI generators that can create images, audio, and videos from human-generated prompts and demonstrated this by playing an AI-generated song about ACR 2024 made with a GenAI music creation program called Suno. Kim then took the audience back to the 1960s when Joseph Weizenbaum created ELIZA, the first-ever AI chatbot that many found engaging, even though it worked primarily by rephrasing the user’s responses.
Technology should work for you, not the other way around.
Generative AI is advancing quickly. Kim shared some of its future trends, such as large multimodal models (LMMs), AI agents and robotics. In his discussion on LMMs, Kim cited an example of an LMM from Google, Med-PaLM M, that can generate a full chest X-ray report from an image. With how generative AI is trending within radiology, Kim made a prediction. “I’m going to get a little controversial here,” Kim said. “Expect a fully automated reporting system for not only chest X-rays, but CT scans.”
The lecture then moved into potential use cases of large language models (LLMs), popularized by ChatGPT, in radiology. He organized the use cases by looking at the lifecycle of a radiology report since radiology reports are the primary work product of radiologists. LLMs can assist before, during and after report generation. Before interpreting an imaging examination, radiologists review the reason for the examination, prior imaging reports and RIS or EMR. LLMs can provide concise summarization and be interactive to allow radiologists to ask questions. During the report generation, LLMs can convert free text to structured reports, check for errors and provide clinical decision support. Kim said that LLMs can also be used to automatically generate the radiology reports’ impression section, saving time and reducing radiologists’ cognitive load. After report generation, LLMs can convert radiology reports into patient-friendly versions and other languages.
In discussing the challenges of clinical adoption of AI, Kim urged the field to look beyond the clinical accuracy of AI models and focus on clinical utility. Kim said, “Even with AI, you can’t solve the staffing shortage problem by simply making them work faster and harder.” Kim emphasized how “technology should work for you, not the other way around.” One of the key ingredients that Kim shared for greater clinical adoption of AI was personalization. In addition to making radiologists feel less tired at the end of the day, he attributed the success of commercial GenAI use in automated impressions to the generated impressions sounding like the radiologist using the solution. In contrast, Kim described using ChatGPT wrappers to generate the impression: “It will give you a clinically, pretty decent sounding impression,” he said. “But guess what? It’s not going to sound like you. And I can tell you, as a radiologist, we like the way we sound.”
Kim then highlighted some of the pitfalls and limitations of GenAI in radiology. In particular, he emphasized the issue of hallucination. He said, “Yes, it is often wrong, but never in doubt.” By showing that even the research community can fail to recognize these pitfalls, Kim urged critical evaluations of GenAI literature. He gave examples of overhyped papers and demonstrated why you should not simply copy and paste ChatGPT responses. Kim also cautioned that testing LLMs like people could lead to misleading results and misinterpretation of their capabilities. After citing the issues around data contamination and LLM brittleness, Kim went on to say that he hopes the research community will shift its focus from test scores to figuring out what is happening under the hood and evaluating in ways that reflect the actual practice of medicine. After cautioning against using public-facing chatbots for patient-related information, he emphasized why radiologists need to learn about GenAI by sharing a quote from a JAMA Network study that states, “The medical profession has made a mistake in not shaping the creation, design and adoption of most information technology systems in healthcare…. The same mistake cannot be repeated.”
Despite limitations, Kim believes GenAI has tremendous potential. He concluded his Moreton Lecture by sharing his hope, similar to a recent Radiology editorial by Paul J. Chang, MD, that technologies like GenAI and other advances can help radiologists return to being the doctors’ doctors.