For many patients, the scariest part of getting treated may be the claustrophobia of lying inside a narrow, noisy tube during an MRI scan, or waiting to learn about the progression of heart disease from the results of a CT scan. On the other side of the exam room wall, radiologists face a different source of anxiety: an overwhelming and growing workload.
In part due to medical imaging advances, providers are ordering ever-more diagnostic images to better understand patient health and avoid invasive procedures such as biopsies. At the same time, the global population is aging, and more patients have conditions that warrant imaging. Yet even as need skyrockets, the number of radiologists isn't keeping up - a trend that accelerated during the COVID-19 pandemic, when more radiologists than usual left their jobs. Those trends are leading to practitioner burnout and delays in patient results. According to the American College of Radiology, 2025 was the third year in a row where workforce shortages were the biggest threat to the field of radiology.
Researchers from UC Berkeley and UC San Francisco are trying to address this need with artificial intelligence - part of a growing trend in medicine of using AI to augment providers' work while addressing challenges such as the rising cost of healthcare and disparities in access to medical care.
In 2025, Berkeley and UCSF researchers launched Voio, a startup that aims to build AI models to help radiologists interpret images faster and more accurately. Voio's tools are being designed to generate draft reports, freeing up radiologists to focus on patients, and to predict patient risk for serious conditions like cancer, osteoporosis and heart failure years in advance - even anticipating how individuals will respond to different treatment regimens.
"We are empowering individual radiologists to have more impact even with overwhelming workloads - and ultimately, to save more patients' lives," said Voio CEO Adam Yala, an assistant professor of Computational Precision Health, Statistics, and Computer Science at UC Berkeley and UC San Francisco. Voio plans to develop similar advancements through AI across other medical fields as well.
Yala launched Voio with co-founders Dr. Maggie Chung, assistant professor in residence in the Department of Radiology and Biomedical Imaging at UCSF, and Trevor Darrell, professor in residence at UC Berkeley's Electrical Engineering and Computer Sciences Department.
Before launching Voio, the researchers developed Pillar-0, an open-source AI model trained on troves of UCSF medical images to detect current conditions such as brain hemorrhaging, as well as looming concerns not detectable by radiologists, such as long-term lung cancer risk. According to Yala, Pillar-0 is the world's best foundational AI model in radiology today. Teams of researchers, engineers and doctors around the world are building off of it, creating ever-better cancer prediction models and diagnostic tools.
Voio is now developing Pillar-1, a new AI model that will be able to detect patient risk related to different medical threats from an even wider array of images, consolidating the findings in a draft report for the radiologist. Yala says it will assist in interpreting the most complex cases, offering insights into disease progressions that currently aren't detectable by radiologists. Pillar-1 is part of a system Voio is developing that will also complete tasks that don't require specialized medical training in radiology, such as transcribing doctors' voice notes or collating patient data.
Making radiology more efficient and accurate
Chung is excited about helping her fellow-radiologists have more time for patient care. "When we reduce manual, non-technical tasks, we give radiologists back the joy of their work," she said. "It lets us return to why we became radiologists. We are the detectives behind the hospital. Through imaging, we make the key findings that have big clinical impact on our patients."
Yala's hopes for Voio go far beyond virtual work assistants - he wants AI to revolutionize clinical guidelines for radiologists. "The way we think about public health should be transforming," he said. "It should not be that the way we serve you digital ads is more sophisticated and personalized than the way that we serve you cancer screenings."
It's an ambitious goal, one that Yala first embraced during his doctoral research at MIT, where he created Mirai. Mirai is an open-source AI model that can identify people who are high-risk for breast cancer years before radiologists can. He later designed Sybil, an open-source model that does the same for lung cancer risk. Yala said that collectively, more than 90 hospitals across 30 countries are conducting studies or trials using Mirai or Sybil, in some cases building off of them to develop their own medical AI models. A prospective study of Mirai led by Chung recently found that using AI could help women at high risk for breast cancer get faster evaluations. Several hospitals across the U.S. are recruiting patients for a new clinical trial to further study Mirai's breast cancer detection rates.
"These tools are advancing the state of the art in oncology," Yala said. "You make a new type of clinical care possible because you can see into the future. You can be proactive."
According to Darrell, while AI tools already exist that automate some parts of radiologists' work, they have not been shown to make radiologists more productive overall. "That's what is really going to make a difference," Darrell said. "We don't need more automation adding bells and whistles in front of a radiologist. We need AI that makes them more effective, accurate and productive. That's what we are building."
Yala, Chung and Darrell's collaboration emerged from the UCSF/UC Berkeley Joint Program in Computational Precision Health (CPH), which was established in 2021 to apply a mix of technological approaches including bioinformatics, genomics, machine learning and simulations to improving clinical care. Yala was one of CPH's first faculty; to this day, he drives between campuses to teach. From the very beginning, Yala embraced CPH's goal of developing AI models that have a real impact in people's lives.
"I think that at CPH we have the best ecosystem in the world to enable this kind of innovation," Yala said. "I'm very grateful to be here."
Yala said that as a private startup, Voio now has access to orders of magnitude more data to develop Pillar-1 and other models than his university team. "Voio is quickly making really big leaps in what AI models can do," he added.
Most of all, Yala said he is excited to get Voio's AI tools into practitioners' hands - to make a true difference for overwhelmed radiologists. "With our models and products, it's going to be genuinely more exciting and empowering to be a radiologist next year than last year," he said. Hopefully, in future radiologists will be able to worry less about overwhelming workloads, and focus more on patient care.
