A single abdominal scan contains diagnostic signals far beyond its initial clinical order.
Beyond the Primary Target
Every year, millions of patients undergo routine CT imaging for isolated symptoms. Most of the imaging data generated by these procedures is ignored once the primary clinical question is answered.
A vision-language model named Percival, trained on roughly 400,000 CT scans from over 50,000 patients, turns routine images into passive screening tools. Operating in the background, the algorithm flags early signals for chronic kidney disease, atrial fibrillation, pancreatic cancer, cirrhosis, and pneumonia. It detects these systemic risks more reliably than established baseline models like TotalSegmentator without requiring additional radiation exposure.
Preclinical Limits and Real-World Testing
Passive screening presents clear operational potential, but diagnostic detection on historic datasets does not guarantee improved patient outcomes.
Percival remains in preclinical testing. Its initial validation relies entirely on retrospective data from a single health system. Unvalidated background algorithms risk triggering unnecessary downstream workups and increasing clinician burden if false positive rates fluctuate across different populations.
The code is now open-source, with multi-center evaluation scheduled ahead of prospective clinical trials targeted for 2027. Operational leaders should evaluate cross-institutional validation data before considering passive screening workflows.
