Automated AI Tool Detects Early Pancreatic Cancer on Routine CT

Pancreatic ductal adenocarcinoma, the most common form of pancreatic cancer, is often diagnosed at a late stage and carries poor survival because early disease produces few symptoms and subtle tissue changes that conventional computed tomography frequently misses. This delayed recognition limits opportunities for surgery and effective treatment. Researchers have now developed an artificial intelligence framework designed to identify stage 0 tissue signatures on routine scans and support earlier detection.

Radiomics-based Early Detection MODel (REDMOD) is designed to detect subtle pancreatic texture patterns that the human eye cannot perceive on standard computed tomography (CT). The framework automatically segments the pancreas to eliminate manual variability and then analyzes radiomic features for a preclinical malignancy signature. The findings are published in Gut.

Image: The AI framework analyzes subtle pancreatic CT texture patterns and radiomic features to detect preclinical malignancy signatures (image credit: Adobe Stock)

The system was evaluated on abdominal CT scans from 219 patients across several hospitals whose images had been read as showing no disease but who were later diagnosed with pancreatic ductal adenocarcinoma. Subsequent diagnoses occurred from three to 12 months in 40% of cases, 12 to 24 months in 35%, and more than 24 months in 25%, with nearly two thirds of tumors arising in the pancreatic head. Scans from 1,243 matched individuals without disease during up to three years of follow‑up served as comparators.

REDMOD identified a preclinical signature an average of 475 days before clinical diagnosis. It outperformed experienced radiologists in sensitivity for early malignant changes, 73% versus 39%, and showed higher accuracy for cases more than two years before diagnosis, 68% versus 23%. In independent testing, the system correctly classified scans as free of pancreatic cancer in just over 81% of 539 patients and in 87.5% of 80 patients from the U.S. National Institutes of Health NIH‑PCT dataset. Longitudinal assessments showed stable outputs, with 90% to 92% agreement across earlier scans from the same patients.

The researchers note that their cohorts were not ethnically diverse as one of the limitations of the findings. They emphasize that prospective testing in high‑risk patients, including those with unexpected weight loss and newly diagnosed diabetes, is needed before clinical deployment. They conclude that a fully automated approach capable of flagging stage 0 disease on routine imaging could shift pancreatic cancer diagnosis from late‑stage presentation to preclinical interception.