Noted scientists urge imaging groups to consider ‘division of labor’ between AI, radiologists

Two noted scientists are urging imaging groups to consider a “division of labor” between radiologists and artificial intelligence, rather than having the two compete for similar tasks. 

Renowned cardiologist Eric J. Topol, MD, and Harvard AI authority Pranav Rajpurkar, PhD, shared their thoughts in a new commentary published Tuesday by RSNA’s Radiology. Prevailing views have favored an “assistive approach” to AI, they note, attempting to merge human judgment with machine precision to improve patient outcomes.  

“However, recent evidence challenges this assumption by demonstrating that integrating AI into radiologists’ workflows does not always yield the anticipated enhancements,” Topol, a professor and executive VP of Scripps Research, and Rajpurkar wrote July 29. “This emerging perspective suggests that a clearly defined division of labor—where AI and radiologists assume separate responsibilities in the diagnostic process—might provide more reliable results.”

The duo suggests three possible ways in which to separate duties: 

• AI-first sequential model: Where artificial intelligence processes the initial segment of the workflow—such as pulling clinical context from electronic health records—followed by radiologist interpretation.
• Physician-first sequential model: The radiologist initiates the diagnostic process while AI performs other complementary tasks, including generating an initial report and possible follow-up recommendations. 
• Case allocation model: Imaging is triaged based on complexity and clarity, with AI managing some instances while radiologists take on others, and the rest are handled in combination between the two.

“At its core, role separation leverages the distinct strengths of AI and radiologists by assigning them complementary yet separate responsibilities in the diagnostic workflow. Rather than merging efforts into a single, integrated process—which can lead to issues like automation neglect and automation bias—this approach divides the workflow into segments best suited to each party's unique capabilities,” the authors suggested. 

Rajpurkar and Topol hope this new framework can help radiology groups to move beyond feelings of distrust and allow them to grasp AI’s full potential. In clinical practice, they believe the boundaries between these three models will “often blur.” Adaptations will need to be made, too, accounting for varying radiologist preferences, institutional practices, imaging modality requirements, and clinical contexts. 

"We're providing a framework, but the real innovation will come from frontline radiologists adapting it to their specific needs," Rajpurkar, a professor of biomedical informatics at Harvard and co-founder of a2z Radiology AI, said in a statement from the Radiological Society of North America, which published the editorial. “Institutions will likely discover hybrid approaches we haven't even imagined yet.”

He gave the example of a trauma center using the AI-first model to review chest X-rays during an overnight shift but then switching to a physician-first approach when teaching medical residents. 

"The breakthrough moment comes when practices stop asking 'Which model?' and start asking 'Which model when?'" Rajpurkar added. "That's where the magic happens—adaptive workflows that respond to real-time clinical needs, not rigid theoretical constructs."