What Will Imaging AI Look Like in 10 Years?

About five years ago, renowned computer scientist and one of the “godfathers of deep learning,” Geoff Hinton, PhD, said, “We should stop training radiologists now,” predicting the widespread adoption of deep learning in radiology. However, about five years later, though the number of artificial intelligence (AI) solutions targeting clinical use cases continues to rise, it is clear that much more work needs to be done before completely autonomous AI is a reality. In fact, the highest value provided by AI derives from its role as a synergistic resource in radiologists’ daily practice. Accordingly, hundreds of radiology trainees, physicians, and imaging informatics professionals join the Society of Imaging Informatics in Medicine (SIIM) and the Radiologic Societies of North America (RSNA) biannually to learn how to leverage AI solutions to enhance patient care.

During the National Imaging Informatics Course – Radiology (NIIC-RAD), participants have the unique opportunity to participate in interactive small group didactic sessions with leaders in the field of imaging informatics. As part of a new session during the Fall 2021 course, participants separated into breakout groups led by imaging informatics luminaries to discuss a difficult question: “What will imaging AI look like in ten years?”

As seasoned experts in the field of imaging informatics, Felipe Kitamura, MD, and Rutwik Shah, MD, have witnessed the rapid evolution of artificial intelligence (AI) for imaging in the last few years. Fortunately, both Kitamura and Shah volunteered to guide a group of five radiology trainees and aspiring imaging informatics professionals through a discussion on the projected course of imaging AI in the next decade. However, the group first reflected on the progress made by AI over the last decade with respect to promised returns.

“What is the state of AI as it currently stands? Are we satisfied with AI’s performance based on our expectations? Or is [AI] at a level ahead of our time?” said Shah, who serves as Project Director for Clinical Product Innovation at the Center for Intelligent Imaging (ci2) at the University of California San Francisco (UCSF).

The group agreed that though AI has made strides for proof-of-concept diagnostic tasks and shown promise to significantly improve workflow efficiency, there remains a need for a radiologist at the helm to direct clinical decision making that is informed by AI tools. Andrew Kim, a PGY-5 resident at Queens University in Kingston, Ontario, underscored practical limitations to deployment of commercially available AI tools that hinder its use in his practice.

“As part of a publicly-funded institution that can be, at times, behind on basic medical technology, we have zero exposure to AI,” said Kim.

Acknowledging these struggles, Kitamura, who serves as Head of Innovation in Diagnostic Operations at Diagnosticos da America SA (DASA), discussed the importance of having a well-developed information technology (IT) infrastructure in place as a solid foundation for deploying AI in clinical settings.

“It’s like having a cherry on a cake,” said Kitamura. “We can put the cherry on the cake, if we have the cake. Otherwise, the cherry is useless.”

The thought-provoking conversation then turned its attention to the predictions for AI in the next decade. Unanimously, it was determined that adopting AI in practice will necessitate the creation of new roles for responsible tool deployment and maintenance. Joseph Gamboa, a PGY-5 resident at Texas Tech University Health Sciences Center, highlighted the need to demonstrate return on investment (ROI) for continued AI tool use in practice and to continuously evaluate access to tools in diverse settings.

“The cost of AI is huge,” said Gamboa. “Unless [developed] at your home institution, you need to show ROI for funds to go to vendors. We may also see disparities between urban academic centers versus community-based, smaller rural areas.”

After a thought-provoking discussion within their small groups, participants joined the NIIC-RAD cohort at large to share collective perspectives. One common theme was that both imaging informatics and AI are rapidly evolving domains, and it can be difficult to accurately predict day-to-day processes from even just a few years in advance. However, there was an overall optimistic tone that acknowledged AI’s role in augmenting workflow and non-diagnostic tasks.

“AI may have the biggest impact on image acquisition in ten years, among other ‘non-pixel’ implications,” said Eliot Siegel, MD, Vice Chair of Information Systems at the University of Maryland School of Medicine, on behalf of his group.

As the conversation wrapped-up, several members suggested that it could be more than just ten years before meaningful change is observed, and they expressed curiosity about the  designation of this time in the original question. In response, Tessa Cook, MD, PhD, Assistant Professor of Radiology at the Perelman School of Medicine at the University of Pennsylvania and creator of the NIIC-RAD course, explained that ten years represented a balanced period of time considering the rapidly evolving nature of AI balanced against pragmatic factors for AI adoption.

“20 years is too much in the future to be able to say anything meaningful,” explained Cook. “And then for five versus ten, well Geoff Hinton’s prediction was just shy of 4 years and 11 months ago. So, five years is clearly not long enough.”