Generative AI model could help ease radiologists’ burdens

Georgia Garvey , Contributing News Writer

What’s the news: HOPPR, a portfolio company created via investment from the AMA, recently announced the launch of a generative augmented intelligence (AI) model that aims to dramatically improve the way physicians interact with medical images. 

Membership Moves Medicine™

  • Free access to JAMA Network™ and CME
  • Save hundreds on insurance
  • Fight for physicians and patient rights

The model, called Grace, is a foundation model. That means developers, radiology picture archiving and communications systems (PACS) vendors, and AI companies can build on top of it to rapidly create their own applications.

According to HOPPR CEO Khan Siddiqui, MD, at scale, Grace will be built on a petabyte of high-quality, anonymized medical data from different types of sources—including images from computed tomography magnetic resonance imaging scans, X-rays and echocardiograms. The model offers the possibility of physicians and technicians interacting with, and even speaking with, medical images.

“This kind of model enables that interactivity. It goes beyond interpreting images and radiologist dictations by considering the broader context of patient management, providing insights into necessary actions.” said Dr. Siddiqui, a radiologist.

The development of Grace came thanks to a milestone investment from Health2047, the wholly-owned innovation subsidiary of the AMA that was created to overcome systemic dysfunction in U.S. health care.

Why it’s important: "Grace brings great promise for physicians,” Dr. Siddiqui said. One of the potential applications of the technology is to automatically fill in a radiological report.

“A lot of the things we dictate in radiology reports are repetitious,” he said. “And a lot of research has shown that it creates cognitive dissonance when you are doing this manual work in addition to looking at images. If we can spend more time in radiology looking at images and less on the other stuff, it'll get better from a clinician’s point of view.”

Such AI models also may help improve imaging visits. He gave the example of a mammography patient in which additional imaging will be necessary. With the use of applications powered by Grace, a technician can be informed to get additional images before the patient even leaves.

“You’re solving the problem before the patient leaves your side,” which can help reduce patient stress and improve clinical outcomes, said Dr. Siddiqui.

Grace is trained with full resolution and full bit-depth images, with no down-sampling. This uses all the information on the images, he noted. Mammograms, for example, are acquired as 16-bit images, with more than 65,000 shades of gray. However, most models are trained on 8-bit images, with 256 shades of gray—a dramatic loss of useful data.

Because those images contain so much information, he said, one day AI models will even be able to identify masses or irregularities before they are detectible by the human eye.

He said Grace also significantly shortens the timeline for creation of new AI-powered applications for medical imaging.

“Developers can fast track AI application development by fine tuning the Grace model which takes weeks versus years when developing AI algorithms from scratch. The regulatory clearance process will still be the same.” Dr. Siddiqui said, referring to the clearance needed from the Food and Drug Administration.

Last month, the American College of Radiology introduced AICentral to help “empower radiologists and practices to access increasingly transparent AI product information and make better informed AI purchases.”

Learn more: With the increasing role that AI plays in medicine, the AMA House of Delegates defines AI not as “artificial intelligence” but as “augmented intelligence.” The language is intended to highlight the critical role humans must play in the responsible use of AI, particularly in medical environments.

With that in mind, the AMA Board of Trustees has approved a set of guidelines for the development and use of AI in medicine (PDF). The document lays out the ways in which AI can be used responsibly, equitably, transparently and ethically.

Catch up with Health 2047’s recent accomplishments by listening to an episode of the venture’s “So You Want to Transform Healthcare” podcast featuring AMA Executive Vice President and CEO James L. Madara, MD. He joined Christine Stock, MD, the managing director of medical affairs at Health2047, to highlight the importance of assuring that the physician-patient interaction is at the heart of health care innovation and emphasize the need for collaboration with other innovation hubs while navigating economic uncertainties in venture capital funding.

They also reflected on the evolution of Health2047’s portfolio and their hopeful outlook for the future of health care innovation.