Current Perspectives on Integrating AI into Radiology

Amid the ever-increasing imaging volumes radiologists are seeing, how does one address cases in which the radiologist and AI don’t see eye to eye in a timely manner?

At the Yale University School of Medicine, Joseph Cavallo, MD, MBA, said they have prioritized immediate feedback mechanisms with their AI vendors that allow minimal click notification of AI misses without little workflow disruption. In a recent interview with Diagnostic Imaging, Dr. Cavallo said they also developed system macros that aid radiologists in explaining false positives with AI in radiology reports.

“This helps clear up confusion if the referring providers see some of the AI-generated notations or alerts within the PACS system themselves, since everybody has access to the PACS images now. It also helps us from a tracking standpoint if we wanted to be able to see within the reports, not just the external AI data, where we're disagreeing with an AI finding,” pointed out Dr. Cavallo, an assistant professor of radiology and biomedical imaging at the Yale University School of Medicine.

How Can AI Lead to Enhancements in Radiology Workflows?

While there have been many significant strides with AI, Dr. Cavallo emphasized the potential of AI to further enhance efficiency in the reading room and broader workflow management.

“How do we rapidly adopt in a secure way, you know, these general-purpose AI solutions to augment our day-to-day workflows, decrease our protocol burden, help manage our scheduling, help forecast demand and staffing optimization?” posited Dr. Cavallo.

Keys to Addressing AI in the Training of Radiology Residents and Fellows

Acknowledging concerns about automation bias and possible deskilling, Dr. Cavallo said they have emphasized the development of an “ironclad” clinical skill set for radiology residents and fellows as well as incorporating data literacy and information technology concepts into a “robust informatics course.”

“AI is becoming an integral part of the radiologist workflow, and the onus is on the trainers to make sure that the residents and fellows are well versed with these tools and are in a position to leverage them themselves moving forward. We've made some internal decisions to try and balance those two forces, and I'm sure different institutions are handling it differently,” noted Dr. Cavallo, an assistant director of informatics for clinical artificial intelligence in radiology at the Yale University School of Medicine.

(Editor’s note: For related content, see “The Reading Room Podcast: Assessing the Current Landscape of AI and Reimbursement in Radiology, Part 1,” “The Multi-Stakeholder Dilemma with Assessing Value and Costs of AI Integration” and “FDA Clears AI-Powered Triage Software for 10 Findings on Abdominopelvic CT.”)