New AI model detects hidden diseases in routine CT scans

When analyzed by a new artificial intelligence model developed at Penn Medicine, CT scans reveal much about a patient's health beyond the specific condition they were originally ordered for, according to a new study published in NPJ Digital Medicine. The AI model, dubbed Percival, can uncover nearly hidden patterns in disease and potential future health risks that might have otherwise gone unnoticed by human eyes. While other models may need to be directed to find certain health conditions or outcomes, Percival can effectively run in the background to pick up what is unseen. 

Our teams have developed something that learns broadly from imaging data and uncovers connections between scans, lab values, people's physiology, and clinical outcomes to potentially give more accurate assessments of health, or even ring alarm bells early. This is work that points toward a future in which medical imaging becomes a foundation for more personalized and proactive healthcare." 

Walter Witschey, PhD, senior author, associate professor of Radiology and associate vice chair of Research IT

Built on the back of the BioBank 

Named as a reference to one of the knights of King Arthur's Round Table, and also in reverence of the late Eugene Percival Pendergrass, a prominent radiologist and educator at Penn, Percival was built with help from a variety of teams at Penn Medicine and trained on data from the Penn Medicine BioBank. With information from more than 350,000 people collected across the University of Pennsylvania Health System over the course of two decades, the BioBank provides an exceptionally diverse source of data for a model like Percival to be built upon. 

"Research in AI has shown that training on larger, more diverse datasets can improve performance, and this can provide a stronger foundation for adapting a model to a hospital or patient population almost anywhere," said lead author Cameron Beeche, a PhD candidate in Bioengineering at the University of Pennsylvania. 

The researchers used roughly 400,000 CT scans from approximately 50,000 patients to hone Percival's abilities, which are still being tested preclinically. 

"We're hoping to elevate 'opportunistic screening,' where results are found beyond the original purpose of a scan, such as finding the early traces of cancer on a scan that was maybe assessing pneumonia," Witschey said. "We believe the totality of what's in a CT scan can be made much more useful with help from a model like this." 

Uncovering the unseen 

In analyzing Percival's performance, the Penn Medicine teams were able to show that CT scans contain a plethora of useful information hiding in plain sight. 

Percival was found to uncover traces of several chronic conditions, including atrial fibrillation (a form of irregular heartbeat) and chronic kidney disease in scans. 

Additionally, when compared to a similar AI model called TotalSegmentator, Percival was found to flag conditions like pancreatic cancer, cirrhosis, and bacterial pneumonia much more reliably. 

"Percival can help identify unrecognized disease, predict future progression, and anticipate major cardiovascular and liver complications," said co-author Charles Kahn, MD, a professor of Radiology. "Because many patients already have substantial heart or metabolic disease when they enter specialty care, routine imaging may provide an opportunity for earlier recognition and intervention." 

Making Percival more 'robust' 

The developers behind Percival are hoping to test it early next year with an eye toward bringing the model into the clinic for regular use. They believe their work in another Penn Medicine-developed AI radiological solution, AInSights, will help them in this process. 

In the meantime, the team is examining how even larger data models and those spanning multiple institutions could make Percival more robust and predictive. 

Percival's code and training model specifications are open source, which could also help it improve. 

"Percival benefited enormously from researchers who shared their code, methods, and insights into these types of models," Witschey said. "We believe strongly in contributing to that same spirit of open science." 

Penn Medicine's portfolio of AI models built to help uncover more information from scans continues to grow. In addition to AInSights and Percival, developers at Penn Medicine also recently developed a model built specifically to analyze cardiac MRI scans and help diagnose new conditions. A test found that it had near-perfect accuracy in its diagnoses. 

This work is supported by the National Institutes of Health (NIH) Office of the Director under award OT2OD038048 and the National Institute of Biomedical Imaging and Bioengineering under award R21EB036734.