Reimagining AI in Radiology: Clear Role Separation as the Path Forward

In a new editorial published in Radiology, physician-scientist Dr. Eric J. Topol and Harvard artificial intelligence (AI) expert Dr. Pranav Rajpurkar call for a fundamental shift in how AI is integrated into radiology. Rather than blending AI seamlessly with radiologist workflows—a strategy that has largely fallen short—they propose clearly defined roles for humans and machines to overcome the current impasse between distrust and overdependence.

“We’re stuck between distrust and dependence, and missing out on the full potential of AI,” said Dr. Rajpurkar, an associate professor of Biomedical Informatics at Harvard University.

Despite significant advances in AI, real-world integration into radiology remains limited. Many anticipated benefits, including enhanced diagnostic accuracy and streamlined workflows, have yet to fully materialize. Dr. Topol, professor and executive vice president at Scripps Research, noted, “It’s still early for getting a definitive assessment, but several recent studies of GenAI have not demonstrated the widely anticipated synergy between AI and physicians.”

A central issue, the authors argue, is confusion over how to best apply assistive AI tools. Radiologists often struggle with when to trust AI input and when to rely on their own expertise. “Radiologists don’t know when to trust AI and when to trust themselves,” Dr. Rajpurkar explained. “Add AI errors into the mix, and you get a perfect storm of uncertainty.”

Beyond psychological hurdles, practical challenges have also slowed AI adoption. These include biases introduced by AI suggestions, misaligned financial incentives, unclear clinical workflows, and lingering questions about liability and accountability. As Dr. Rajpurkar put it, “After years of hype, AI penetration in U.S. radiology remains surprisingly low. This suggests we’ve been implementing AI like sprinkling digital fairy dust on broken workflows.”

To address these concerns, the authors propose a new framework based on role separation, guided by clinical validation and real-world testing. Their model offers three adaptable approaches:

1. AI-First Sequential Model – AI handles initial tasks, such as gathering patient context from electronic health records, with the radiologist providing the final interpretation.

2. Doctor-First Sequential Model – The radiologist leads the diagnostic process, with AI assisting through secondary tasks like report drafting or recommending follow-up procedures.

3. Case Allocation Model – Cases are triaged based on complexity. Routine cases may be handled by AI alone, while ambiguous or critical cases are escalated to radiologists.

“Clear role separation breaks this cycle,” Dr. Rajpurkar emphasized. “Radiologists are stuck in the worst of both worlds—afraid to trust AI fully, but too reliant to ignore it.”

Rather than rigidly enforcing these models, the authors advocate for iterative, context-specific implementations. Institutions might blend approaches, such as using AI overnight in trauma centers to review X-rays, while shifting to radiologist-led workflows during daytime teaching rounds. The key is adaptability. “The breakthrough moment comes when practices stop asking ‘Which model?’ and start asking ‘Which model when?’” said Dr. Rajpurkar.

To realize this vision, pilot programs must be carefully designed to measure impacts on diagnostic accuracy, efficiency, and clinician satisfaction. Honest reporting of both successes and failures is critical, according to the authors, who emphasize the need for a shared evidence base.

Looking ahead, the authors also recommend the development of a clinical certification pathway for AI systems—a process that would involve regulators like the FDA, but also require new, multidisciplinary oversight bodies. These would bring together technical experts, clinicians, and implementation specialists to evaluate AI systems within real-world healthcare settings.

“We’re not there yet,” Dr. Rajpurkar acknowledged. “But when these systems can competently manage the breadth of tasks a senior medical resident handles, the entire conversation changes. That’s the inflection point we’re watching for.”

Ultimately, the editorial offers a grounded and forward-looking roadmap—one that replaces vague promises of AI transformation with practical strategies to redefine collaboration between humans and machines in medical imaging.