AI agents interact with radiologists, technologists and referring physicians

CHICAGO – In a demonstration at the RSNA conference last December, Dr. Luciano Prevedello, a radiologist at Ohio State University’s Wexner Center of Medicine, chatted with his AI-enabled computer about a patient’s brain lesion as he created a report. He showed how he could discuss the case with an AI agent, much like he would with a colleague.

Not only did the AI agent offer expert opinions on the lesion’s type and location, but it could also bring up different patient exams on the computer screen, including CTs and MRIs that had been taken, with a variety of views.

That ‘natural language’ discussion was an example of how reading and reporting on radiological studies will change in the not-so-distant future. It was part of an exciting panel presentation at RSNA on ‘The Reading Room of the Future’, most of which focused on the role AI will play.

Dr. Prevedello discussed how AI agents are being deployed in the DI department of the hospital in a variety of ways – although much of the work is still in the early stages.

“We’re starting to see agentic AI introduced in multiple areas – ordering, protocols, acquisition, worklists and reporting,” he said.

Dr. Prevedello is also the vice-chair for medical informatics and augmented intelligence in imaging, division chief of medical imaging informatics and medical director of the 3D and Advanced Visualization Lab. He spoke remotely, by video call, as he was prevented from attending in person by the snowstorm that closed Chicago’s airports just before the conference.

Dr. Prevedello said that on the ordering side, AI agents are already able to autonomously schedule exams, calling patients and managing inbound and outbound phone traffic.

And when it comes to protocols, “they can select the most appropriate protocols, based on the free-text indications.”

“We’ve been working on this at Ohio State for a while now, experimenting with large language models to understand the indications, to guide ordering physicians and to help with protocol selection.”

On the acquisition side, he said that AI has been used to help position patients and to reduce anxiety. And for worklist and routing, “AI agents have been able to distribute the exams, based on complexity and sub-specialty, and to prioritize urgent cases,” he said.

But the most astounding changes are appearing on the reading and reporting side. Dr. Prevedello demonstrated work under development at Ohio State University’s medical center, where AI agents are analyzing entire imaging studies and providing full reports with diagnoses.

In an impressive demonstration, Dr. Prevedello discussed a patient’s images, in natural language English, with an AI agent.

The agent stated in a clear voice, “Okay, I see a screen with a presentation. It mentions progressive visual disturbance with bitemporal hemianopsia. And it shows it’s a 35-year-old female.”

The agent then asks, “Is there anything else you would like to discuss?”

Dr. Prevedello said, “What do you think about this history? The patient’s history is concerning.”

The agent responded, “Progressive visual disturbance with bitemporal hemianopsia suggests a lesion at the optic chiasma. Is there anything else you would like to discuss?”

“Yes, I would like to see the images. What do you see?”

The AI agent brings up images and says, “I see a CT scan of the brain. It’s an axial view and I see a possible mass in the region of the sella. Would you like to see anything else.”

“Yes, I would like to see the image of the bone window, okay?”

The agent and the doctor go on to look at several more CT and MRI sequences, with the same kind of back-and-forth discussion. Eventually, the AI agent suggests a pituitary adenoma as the most likely diagnosis.

While all this was astounding, Dr. Prevedello stated that the agent’s final diagnosis was incorrect. “It was actually a very rare case of a chordoma simulating a sella lesion,” he said.

Still, the system’s ability to work with the radiologist in natural language, and to call up images and point out areas of concern, was remarkable. “What’s even more impressive,” said Dr. Prevedello, “was that the agent was not specifically trained for radiology. It was a general-purpose model.”

He then demonstrated two more agents that diagnosed sets of images. One of them analyzed a large series of CT exams in less than two minutes. One of the agents got the diagnosis correct, while the other didn’t.

“I think the technology is still not there yet, especially for rare conditions,” said Dr. Prevedello. However, it’s early days for these systems – what’s coming will be even more formidable.

Even in the future when these agents are much more accurate, Dr. Prevedello asserted they will always require human supervision, as there is always a chance of error. “The computer is incredibly fast, accurate and stupid. But man is unbelievably slow and accurate and brilliant. And the marriage of the two is beyond calculation.”

Dr. Nancy Pham, a neuroradiologist at Stanford University, highlighted several areas in which AI is helping radiologists. She noted right away that imaging volumes are increasing at a rate of 3 percent to 12 percent annually, while there are only about 2 percent more radiologists each year. “So, that means we have a large and unsustainable workload gap,” she said. The fast growth in DI exams is also leading to burnout and the possibility of more diagnostic errors.

For these reasons, she said, AI is being harnessed to improve efficiency and workflow automation.

There are some real success stories already, she said. In a trial of AI reading for breast cancer, 80,000 women were screened and there was a 29 percent increase in cancer detection using an AI tool. There was also a 44 percent reduction, she said, in radiologist workload.