AI finds undiagnosed conditions hidden in routine CT scans
A CT scan is made to answer one clinical question. But the images contain far more information than any radiologist can process. A new AI model extracts those hidden signals.
Researchers at the University of Pennsylvania trained an AI model called Percival on more than 400,000 CT image volumes from 52,000 patients, paired with their radiology reports. The model learned to recognise patterns linked to diagnoses spanning cardiovascular disease, metabolic conditions, cancer, and more. The study was published in the journal npj Digital Medicine.
What the model detected
The model extracted meaningful information about patient characteristics, including age, sex, and body size. It also found significant associations with 35 types of laboratory measurements, covering kidney function, liver function, blood glucose, lipid metabolism, and blood cell counts. Image patterns were linked to 351 diagnosis categories across multiple organ systems.
Performance was uneven, however. Directly estimating lab values from scan images was largely weak. Only albumin, haemoglobin, and creatinine showed reasonable predictive accuracy. The model is therefore not a replacement for blood tests, but a supplementary signal-detection system.
Relevant for early detection as people age
For preventive medicine and longevity science, the direction is promising. Many people over fifty already receive CT scans for other reasons. If those scans can also surface early signals of undiagnosed conditions, the informational value of existing imaging increases without additional burden to the patient.
A direct comparison with human radiologists is the next step, the researchers note. Only then will it become clear how much clinical value Percival adds in practice. For now, it remains a promising tool that requires further validation.
Search terms for further research: vision-language model medical imaging, CT scan biomarker prediction, early detection AI radiology