With Artificial Intelligence (AI)- and Machine Learning (ML)-based tools poised to change the practice of radiology, are colleges, medical schools and radiology residency programs changing how they train students?

“The challenge is, it’s a very emerging and evolving field,” said Keith Dreyer, DO, chief science officer at The American College of Radiology Data Science Institute (DSI), formed in 2017.

Dreyer said DSI is “sensitive to the fact that you have to educate a wide group of constituents,” ranging from practicing doctors, educators, trainers, residents and medical students. Indeed, part of the challenge is creating a multidisciplinary environment.

“Currently, there’s data science and clinical medicine,” said Dreyer, who also holds a PhD in computer science and a masters in image processing and AI, and serves as vice chairman of radiology at Massachusetts General Hospital and chief data science and information officer for the Departments of Radiology at Massachusetts General Hospital and Brigham and Women’s Hospital. “Even at the level of research, you hire two kinds of people [today].” To address this, Harvard Medical School, where Dreyer is an associate professor of radiology, created a center for clinical data science in 2016. The center is home to researchers, clinicians, as well as data science and computer science personnel.

Harvard was first, but there are a few other, similar centers, and Dreyer expects these large academic centers will continue to create cross-discipline centers, propagating their insights and approaches to medical schools and residency programs.

Use cases

Meanwhile, DSI is now collecting hundreds of use cases suitable for AI, a first-of-its-kind program—TOUCH-AI—that it hopes will accelerate AI adoption in medical imaging. Each TOUCH-AI use case proposal will be reviewed by a DSI Data Science Subspecialty Panel for its potential to assist radiology professionals in disease detection, characterization, and treatment. The stated goal of the TOUCH-AI project is to ensure that algorithms:

• Address relevant clinical questions
• Can be implemented across multiple electronic workflow systems
• Enable ongoing quality assessment, and
• Comply with legal, regulatory and ethical requirements

In addition, DSI soon plans to make available open source software to promote the creation of clinical AI models, essentially making it easier for institutions to build AI models by providing them with the necessary training data and computational power. 

“Students rely on us to understand how radiology is incorporating new technology and what the future of the field will look like for them, but many of us are ill prepared to teach the younger generation about this, mostly because we ourselves are not sure,” wrote Allison Grayev, MD, a neuroradiologist at the University of Wisconsin School of Medicine and Public Health in Madison in an editorial published online in Academic Radiology.^[1]

Lack of mentorship and faculty role models poses a significant challenge as students move from the preclinical to clinical environment and try to develop understanding of how AI knowledge can be applied and used in the clinical setting.

‘Reboot’ needed?

A paper published last year in npj Digital Medicine called for a “reboot” of medical education, beyond its traditional focus on biomedical and clinical sciences. The authors of “Medical Education Must Move From the Information Age to the Age of Artificial Intelligence”^[2] wrote about a need for “systematic curricular attention” around intelligence tools involving large data sets, machine learning and robots. They noted multiple factors contribute to the failure of ML to be embedded within undergraduate and graduate medical education training, ranging from the lack of accreditation requirements related to AI and the fact that medical schools already struggle to maintain curricular hours “in the current schema” with ever-growing biomedical knowledge and calls for new content areas.

Without change at the graduate and undergraduate level, it will be hard to move students into a position of understanding how AI can be applied and used in clinical settings.

Likewise, the tooling needs to be improved, so the creation of an AI model “doesn’t add more burden or complexity to what the clinician is already doing,” said DSI’s Dreyer. “I give PowerPoint presentations, but I don’t have to write the PowerPoint software,” he quipped.