ChatGPT-4 Matches Radiologists at Pancreatic Cyst Data Extraction from MRI/CT
July 15, 2025
ChatGPT-4 demonstrated near-perfect accuracy in extracting clinical variables from MRI and CT scans of pancreatic cyst patients, equalling manual chart review by radiologists.
Key Details
- Study involved 991 patients with 3,198 MRI and CT scans from Memorial Sloan Kettering (2010–2024).
- ChatGPT-4 extracted 9 critical clinical variables with accuracy rates of 92%–99%.
- AI matched the manual approach, considered the gold standard, for classifying high-risk variables such as cyst size and solid component presence.
- Efficiency and cost-effectiveness cited as major advantages of the AI approach.
- The study was published in the Journal of the American College of Surgeons and NIH/NCI partially funded the work.
Why It Matters
Demonstrating LLMs’ ability to accurately extract structured data from radiology imaging reports could streamline large-scale research and surveillance tasks, reducing costs and improving patient care efficiency. This proof-of-concept signals the increasing feasibility and trustworthiness of leveraging advanced AI models in radiological workflows.