VIENNA – Generative AI may have novel applications in the clinic, but using the technology comes with its share of challenges and future directions, according to a presentation given February 26 at ECR 2025.
In his talk, Marc Kohli, MD, from the University of California, San Francisco, discussed current applications of this technology in clinical radiology. He also talked about challenges in integrating generative AI into practice, potential benefits, and what the future may hold for the technology and how it may pertain to radiologists.
Marc Kohli, MD, discusses current and potential future uses of generative AI in radiology at ECR 2025 in Vienna, Austria. Behind him is an image created by one such generative AI model, which illustrates humans working alongside AI.
“We’re trying to think about how radiologists can work alongside artificial intelligence,” Kohli said.
Large language models that employ generative AI, such as ChatGPT and Gemini, have become a part of daily general life. In radiology, the technology’s performance has been assessed in response generation to patients, taking board exams, and image interpretation assistance.
Kohli used generative AI to create some of his presentation slides.
“Normally, when I create a talk, I would go through and spend a lot of time searching on Google Images for the perfect image to try to represent the themes or emotions I try to convey,” he said. “I was surprised when I put this talk together of just how far the image generation has come.”
Kohli said the technology has potential benefits in aiding radiologists with non-imaging tasks, which could aid with radiology workflows.
“There is an increasing demand curve for our imaging services and at the same time, we are not producing enough radiologists and we are seeing radiologists leave the market,” he said. “A lot of people are looking to AI to increase efficiency.”
One current use of generative AI among radiologists is refining patient communication. An example Kohli shared includes entering a prompt instructing patients on imaging and what they will do during the imaging exam. The AI chatbot from there can assist with simplifying language to make the instructions easier for patients to understand.
“You can ask for further revisions to decrease the more superfluous words,” Kohli said. “You can also take that same person who is used to writing those instructions and help them level up when it comes to readability and access for patients.”
Another current use is assisting radiologists with generating reports. Radiologists today have individual discreet fields that they have to select and put information into.
“My process today is, if I want to say something about the liver, I have to put the cursor in the liver,” Kohli said. “If I want to say something about the kidneys, I have to put the cursor in the kidneys. A large amount of my time and effort is actually spent trying to navigate between those fields rather than being spent talking about my findings.”
Kohli said that imaging and reporting complexity will continue to increase, with more rules to remember and more granular structures to assess. He added that ambient dictation, which uses speech-to-text technology to translate conversations into written documentation, could be one future direction for generative AI.
While current applications of this technology use text-to-speech, a future innovation could see image-to-text documentation with a review by radiologists.
And as for hallucinations from large language models, Kohli suggested that prompts be compared with reference documents rather than all freely available text accessed on the internet.
“The amount of information we’re extracting is increasing and we’re going to have to have tools as radiologists in order to help us synthesize that into something that’s meaningful and something that’s consumable,” Kohli said.
Check out AuntMinnieEurope.com‘s entire coverage of ECR 2025 on our RADCast.