Segmenting the Radiology Artificial Intelligence Market by Function
“Today, we live in that quadrant of things humans can do and humans are supervising,” Dreyer explained. “That is all the [U.S. Food and Drug Administration (FDA)] approved AI stuff that we see today.”
He said the next step is for AI to move into the realm of superhuman work, such as measuring 1,000 lymph nodes at once, or to make a risk prediction about future events in the next two years based on the patient’s prior 40 images, because it looks like a million other patients’ scans. Dreyer said the FDA is in discussions with vendors on fully autonomous AI for radiology applications, but the agency wants to see controls built into the software.
“We are still just a little bit off into the future for that technology,” Dreyer said. “And super-human autonomous AI is way out into the future.”
According to Dreyer, the FDA has different levels it assesses AI algorithms, ranked from easy tasks to the very complex. He said nearly all the AI cleared by the FDA to date are from the easy categories. There are none from the most difficult level, where there also is a lot of concern regarding liability or the potentially serious impact on patient outcomes if the AI is wrong.
“Even though all these FDA-approved algorithms are using AI, it is not all what most people think of where it is making a diagnosis or finding a lesion. So there is a big variety in the ways to classify AI,” Dreyer said.
• Diagnostic aids that can automatically identify critical findings.
• Automation of time-consuming functions such as quantification, contouring and auto complete of text in reports.
• Workflow improvements and automation.
• Radiomics, where AI pulls out data that is not visible to the human eye to make risk assessments or to classify diseases.
• Data mining applications.
• Clinical decision support for next steps in the patient’s care or to ensure imaging exam meets the guidelines
• Modality specific AI to iso-center patients, choosing imaging protocols, or speeding MRI exam time.
• AI to enhance image reconstructions to improve image quality resolution unidentify to fix imaging artifacts.
• Guidance AI to help imagers get the best possible images, even if they are novice users of the system or are unfamiliar with the anatomy.
• Automatic anatomical identification, labeling and contouring of organs or specific types of tissue.
Dreyer also holds the positions of vice chairman of radiology at Massachusetts General Hospital, chief data science and information officer for the departments of radiology for both Massachusetts General Hospital and Brigham and Women’s Hospital, and associate professor of radiology at the Harvard Medical School.
Tis is part of a 4 video interview series with Dreyer on various aspects of radiology AI. Watch the others: