- Design
- Single-centre retrospective cohort using LLM-extracted report features
- Population
- 13,269 patients with pancreatic cystic lesions, 2000 to 2025
- Primary outcome
- Five-year pancreatic cancer or high-grade dysplasia
- Effect
- AUROC Kyoto 0.655, ACR 0.639, European 0.623; 8-year incidence after 5-year stability 0.71%
A single-centre US cohort used a validated large language model to extract cyst features from abdominal imaging reports between 2000 and 2025, then classified 13,269 patients with pancreatic cysts by ACR, Kyoto and European criteria. Published in the Journal of the American College of Radiology on 28 September 2026.
Over five years, 209 (1.6%) developed pancreatic cancer or high-grade dysplasia. Discrimination was modest for all three guidelines (AUROC 0.655 Kyoto, 0.639 ACR, 0.623 European). They agreed completely on high-risk lesions, but the number labelled as having worrisome features ranged from 308 to 507, with similar five-year cancer rates of 10 to 12% in each. Among 1,830 patients whose cysts were stable for five years, cancer incidence over eight years was 0.71% overall and 0.46% in those meeting Kyoto criteria for stopping surveillance.
The data come from report text in one centre, so features missing from reports are missing from the analysis. The study supports two practical points: worrisome-feature definitions vary between guidelines, and a long period of stability identifies patients at low ongoing risk.
- State which pancreatic cyst guideline the report is following
- Report size, growth rate, main duct calibre, mural nodules and wall enhancement every time
- Expect different guidelines to label different patients as worrisome
- Note five-year stability explicitly, as it marks a low-risk group
Why it matters
The choice of guideline changes who gets labelled worrisome without changing who gets cancer.
Don't overread it
Single-centre data extracted from report text; it does not by itself justify stopping surveillance.
The statistics, in plain English
An AUROC of 0.65 means the guideline ranks a patient who develops cancer above one who does not about 65% of the time; 0.5 is chance and 1.0 is perfect. An 8-year incidence of 0.46% means about 1 in 200 stable patients went on to develop cancer.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for radiology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free