VIDEO: An updated look at the use of AI in radiology

AI and radiology—where things stand now

"AI algorithms have made it through the FDA approval process, and people are now looking at and trying to figure out how to build them into their clinical practice and what the economics of it are, what makes these worth while and what adds value," Kahn explained. "One of the challenges is, what do you want these things to do? What role do they fill?"

While more than 300 AI algorithms are now cleared by the FDA, and a large number of these are in radiology, radiologists need to determine what is useful to their practice.

"For things in radiology, it has to improve the productively of the radiologist," Kahn said. "The challenge for us in the long run I think will be that most AI algorithms are not going to get any additional payment from anybody. It is going to be a cost the radiologist is going to bare to some extent. So the AI has to do something that yields value to me and my practice. Does it reduce my malpractice risk? Does it let me practice more efficiently?"

AI algorithms perform a variety of tasks. Originally, everyone thought AI would help diagnose patients, but Kahn said it does a lot more than that and diagnostics is not the primary role of AI systems. He said AI is being used to enhance image quality on CT and MRI exams, or to speed the exam times for MRI. AI is being used to remove bone, blood vessels or metal artifacts from images to make them easier for a human radiologist to read. AI tools also exist to flag mobile X-ray studies they would be read immediately or alert the radiology tech for acute conditions such as pneumothorax, plural fusion or find rib fractures or check the correct placement of tubes in the chest. 

"The question is, what does the system do when it finds something it has not seen before or it was not trained to read?" Kahn said. "Building systems like that might help the radiologist deliver care more rapidly and more efficiently, but it may not be ready yet to replace the radiologist." 

He said AI that can do tasks, especially the ones that are mundane and time consuming for radiologists, actually may help free up the human readers to do more of the difficult tasks, the things humans are batter at doing than machines. He said this augments the radiologists so they can practice at the top of their license. 

AI also can perform tasks that humans cannot, such as looking at the radiomics of image data and determining the genotype of a cancer is so very specific therapies can be used that target that type of cancer. Kahn said that has enormous potential for improving tumor response compared to how patients are treated today. 

AI and radiology—what's next?

Kahn said AI might be implemented in a hybrid approach to help augment radiologists and allow screening for more patients. For AI in digital pathology, he explained not all slides are reviewed by a human pathologist. In those cases, a technical fee is charged, but not a professional fee. That type of model might be used in screening mammography, where AI may look at annual exams, alternating each year between a human or AI reader. AI also might be used to flag all the suspect exams, so the radiologist can concentrate on those exams.

AL screenings are already being used to detect things such as tuberculosis in developing countries, where there just are not enough radiologists to read the exams, especially in any sort of timely manner.

AI applications are also emerging that can be used to look for incidental findings unrelated to the primary reason for the exam, which may enhance the screening benefits of all types of exams to improve preventive care and followup.

Kahn also discussed some of the new AI applications that detect acute findings before the images even enter the PACS and initiate alerts to stroke or pulmonary embolism response teams. He said several of these AI companies are actually going to other departments or hospital administration to show how these systems can improve their response times for treatment, but often bypass the radiology departments. He said this is often because the physicians with the most interest in this are the pulmonologists or neurology specialists who will benefit more from rapid turnaround of the alert and getting these patients into a cath lab faster. 

Using AI to help address the growing shortage of radiologists

Kahn said AI will not be replacing radiologists anytime soon, and the technology is coming at a time when radiologists are being asked to do more with less and as patient exam volumes are increasing. He said there is an opportunity for AI to help address these issues by augmenting radiologists so they can do more. 

"We are going to be short tens of thousands of radiologists in the coming decade, so one way or another, we will see come of these tools coming in to help us," Kahn said. "Having systems do things that are a little bit more mechanical, if an AI system can don that as or more effectively than a human, maybe that is a compromise we make. But whatever we do with it, we want too ensure we are delivering high-quality care for our patients."

Additional coverage of RSNA 2022 is available here and here.