Researchers at UC Berkeley and UCSF have developed a new AI foundation model that could help physicians in radiology, or the practice of recognizing and diagnosing medical conditions.

The model, named Pillar-0, can interpret 3D volumes, in comparison to current tools that can only interpret 2D data. According to researchers, this means that Pillar-0 can interpret images from a CT or MRI exam more accurately.

Researchers trained Pillar-0 using the images of 42,990 abdomen-pelvis CTs, 86,411 chest CTs, 14,348 head CTs and 11,543 breast MRIs.

According to Adam Yala, assistant professor of computational precision health at UC Berkeley and UCSF,Pillar-0 is but the first in a series of AI models being developed by his lab in order to push the “frontiers of radiology.” 

Yala is also the co-founder and CEO of Voio, a new company focused on advancing radiology through better AI methods. Future models of Pillar will come out of this company, Yala said.

Specifically, Yala noted that Pillar-0 is a foundational model, meaning that it serves as a basis to be “adapted for many hundreds of clinical tasks” from lung cancer prediction to brain hemorrhage detection. 

The new development comes out of the Yala Lab — a joint venture from UCSF and UC Berkeley — with the goal of enabling research and innovation into AI-assisted healthcare resources. Yala and his colleagues are being funded by a variety of sources, including the Bakar Fellows Program and the National Institutes of Health.  

Though the paper has not yet been peer-reviewed, the technology is available online, which Yala said is “more practical” given the pace of AI technology. According to Yala, it can take about a year for a big journal to peer review a paper.

“By the time it (is peer-reviewed), we’ll be on Pillar-2,” he said. “So it’s more and more practical that the big models kind of come out when they’re ready, and then the communities use it when it’s live, in practice.”

While AI developments are promising, Joann Elmore, professor of medicine at UCLA who was not involved in the development of Pillar, said it is important to recognize that they are tools, meaning outcome and use may vary in the hands of different physicians.

“The people developing these AI tools — sure, it may do well compared to a physician alone, but what happens when you put the two together?” Elmore said.“We’re humans. We don’t always interact with technology the way you anticipate.” 

Elmore stressed that it’s both important to study and refine the application of new diagnostic technology as well as to develop these new tools.

According to Elmore, the same rigorous research methods — including review for potential bias — applied to evaluating pharmaceuticals and medications must also be applied to AI technology.

Elmore likens the AI technology boom to the “wild west” and emphasizes that just because AI is developing quickly, scientific rigor must be maintained. For Elmore, peer review alongside follow-up studies remains as important as ever. 

“(Although there is) all the hype of wonderful potential AI tools that are being developed, there’s very little data showing improvement in patient care or patient outcomes,” Elmore added. “That should be the next steps.”