Some radiologists see AI primarily as a tool to interpret images — yielding precision findings that could eventually surpass the skill sets and expertise within the specialty. A growing number of radiology groups, however, have been exploring and realizing the less-glamorous uses of AI technology to solve clinical and operational problems.
A recent webinar available as a free on-demand resource on ACR.org, titled “Beyond Interpretation: Unleashing the Potential of Non-Interpretive AI in Radiology,” highlights the impact of non-interpretive AI, the importance of exploring non-interpretive AI use cases, and the potential advancement of AI clinical practice tools.
Co-moderators Alexander J. Towbin, MD, FACR, and Melissa A. Davis, MD, MBA, hosted an expert group of panelists: Nina E. Kottler, MD, MS, Howard (Po-Hao) Chen, MD, MBA, Marta E. Heilbrun, MD, MSCI, Woojin Kim, MD, and Andrea Borondy-Kitts, MS, MPH, a longtime patient advocate and researcher.
Kim, the chief medical officer at Rad AI, opened the discussion on AI’s uses beyond imaging interpretation.
In non-interpretive AI, the algorithm’s primary output is not used to detect a finding or make a diagnosis. One powerful use of non-interpretive AI, Kim says, is for resource management within a radiology group. “You can use AI to create an intelligent worklist — assigning the right study to the right radiologist at the right time,” he says.
Every radiology group grapples with scheduling and workflow challenges. Non-interpretive AI can help with staffing issues, optimize scanner resources and make predictions about scan durations, Kim notes. AI tools can predict scheduling gaps and forecast appointment delays and missed appointments. Combined, workflow predictions can alert radiologists to anticipated heavy workloads.
The webinar discusses the widespread use of natural language processing (NLP) in radiology. NLP allows radiology reports to be searched and mined — much like the way Google users search for information. This can save radiologists time and improve the quality of their reports, including clinical results and notifications. The NLP discussion also identified uses for speech recognition and coding and billing. Delegating routine, but time-consuming, tasks to AI can reduce physician burnout and increase patient-facing time, speakers agreed.
Recognizing the buzz around another type of AI tools — those known for generative AI, including ChatGPT — the webinar presented data addressing promises and limitations. There are seemingly endless ways radiology can use this type of technology, such as for converting reports into a readable format that a layperson more easily understands, for example. However, there also needs to be careful monitoring of these and of using generative AI to write papers or advocacy letters. It can be very dangerous, one speaker cautioned, to overhype such generative applications within AI because “ChatGPT does not understand or know radiology.”
You can use AI to create an intelligent worklist — assigning the right study to the right radiologist at the right time.
Barriers to implementing some AI include the inherent need for standardized data, a limited time commitment by IT staff for integration, and stripping bias from and adding trust to the technology. Nevertheless, non-generative AI has garnered the trust of many radiology groups that are putting it to use.
“These tools are helpful, and we have been using them a lot,” panelist Kottler says. Her group has run millions of reports through NLP models and uses NLP to listen to radiologists as they dictate to provide additional decision-making support. Kottler says you must take the time to educate radiologists about what to expect from AI — including where they might expect to find mistakes.
Getting started with AI requires identifying the strengths of the system you work in, Heilbrun says. It is critical to build an infrastructure allowing for the integration of new AI tools. You must be able to access and leverage the data you have. It may also help to work alongside other institutions that have been through the process and to test new applications through pilot partnerships.
Behind the magic you see going on in AI, you must consider whether it meets the workflow of the radiology team, Chen says. “There is more work to do and not enough people to do it.” Prior to testing the AI waters, it is important to have a clinical expert weigh in on current clinical or operational challenges. It is equally important to always maintain oversight, not letting AI run autonomously, and to get leadership involved when you are looking to pilot an AI solution or purchase AI products.
It is up to radiologists to reach out to patients for their input around AI, Kitts says. Many patients do not know the role AI may be playing in their healthcare. She recommends getting involved with patient advisory panels, already (in place at many hospitals, to raise awareness of AI and to identify areas of mistrust among patients.
The pendulum could swing either way, Kitts says — with AI tools either driving radiologists to read more exams and meet with fewer patients or freeing up time to positively transform the entire patient experience.
For more news, see a full list of the latest ACR Bulletin articles.