ACR pushes for better explainability of how AI algorithms work
With rapidly growing proliferation artificial intelligence in medical imaging, the American College of Radiology wants to see more transparency around how these algorithms work. Some AI can accurately diagnose diseases, but details on how that decision is made is often trapped in a black box and not easily explainable.
The ACR supported a resolution at the American Medical Association House of Delegates 2025 meeting earlier this month. Resolution 519 calls for a framework to convey evidence-based medicine in AI, requiring transparency in how such models make decisions. The resolution was not adopted by the HOD because of existing policies regarding AI already on the books that AMA is supporting, but ACR said it raises awareness about larger issues.
"We are very much interested in the explainability of AI. We want the AI to be able to be explained by a qualified human expert, basically a person who could do the job of the AI, but without the AI. People who haven't worked with AI much don't understand the fallibility of it," Dana Smetherman, MD, MPH, MBA, ACR chief executive officer, explained in an interview with Radiology Business at the AMA meeting.
Radiology has a major interest in AI, because more than 75% of the 1,000-plus clinical algorithms cleared by the U.S. Food and Drug Administration are for medical imaging.
"It is really important to know that the models can drift, but also things that we might not necessarily think about can impact the accuracy of the AI. Maybe you got a new CT scanner from a different vendor and the AI doesn't perform the same way. Or you have a new technologist that changes the protocol and changes the slice thickness from 5 mm down to 3 mm, and all of a sudden the AI is not performing the same way," Smetherman said.
AI performance also can be impacted if it is used on pediatric patients, often lacking training in these unique circumstances. Or, the AI algorithm may not understand types of local patient populations, failing to account for various ethnic, racial or socioeconomic factors. AI also may drift over time, Smetherman noted. She said practices need to have some way to verify AI is working accurately.
ACR has take a proactive approach with the creation of its "AI Central" website that helps practices gain a greater understanding of the technology, track its performance, and take a closer look at algorithms on the market. The college also started the ACR Recognized Center for Healthcare-AI (ARCH-AI) designation last year. This outlines a systematic approach where a radiology practice puts all these pieces together and creates a governance body to manage its AI.
"Basically the way we think about it is, if you can't tell us how it's working, we need to be able to tell our members and our patients that it is working. That's how we see our job. It's really in the post-market surveillance space. But it's evolving so fast," Smetherman said.
Dave Fornell has covered healthcare for more than 17 years, with a focus in cardiology and radiology. Fornell is a 5-time winner of a Jesse H. Neal Award, the most prestigious editorial honors in the field of specialized journalism. The wins included best technical content, best use of social media and best COVID-19 coverage. Fornell was also a three-time Neal finalist for best range of work by a single author. He produces more than 100 editorial videos each year, most of them interviews with key opinion leaders in medicine. He also writes technical articles, covers key trends, conducts video hospital site visits, and is very involved with social media. E-mail: [email protected]