Conversion of Mixed-Language Free-Text CT Reports of Pancreatic Cancer to National Comprehensive Cancer Network...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Processing in Radiology
• Oncology Decision Support

Background:

• Pancreatic ductal adenocarcinoma (PDAC) diagnosis relies on detailed CT imaging reports.
• Mixed-language (English/Korean) narrative reports pose challenges for automated analysis.
• Standardized structured reporting (SR) is crucial for consistent clinical decision-making.

Purpose of the Study:

• To assess the feasibility of GPT-4 Turbo and GPT-4o in generating structured reports (SRs) from mixed-language PDAC CT reports.
• To evaluate the accuracy of GPT-4 models in categorizing PDAC tumor resectability based on NCCN guidelines.
• To compare the performance of GPT-4 Turbo and GPT-4o in these tasks.

Main Methods:

• Retrospective analysis of 115 mixed-language pancreas-protocol CT reports (English/Korean) from two institutions (Jan 2021-Dec 2023).
• GPT-4 Turbo and GPT-4o models were prompted to generate SRs and classify tumor resectability.
• Model outputs were compared against a radiologist-derived reference standard, with triplicates and majority voting for final output.

Main Results:

• GPT-4 Turbo and GPT-4o achieved comparable accuracy in SR generation (92.3% and 92.2%, respectively).
• GPT-4 Turbo demonstrated significantly higher accuracy in resectability categorization (81.7%) compared to GPT-4o (67.0%).
• Primary errors for GPT-4 Turbo involved inaccurate data extraction in SR generation and violation of resectability criteria.

Conclusions:

• GPT-4 Turbo and GPT-4o show promise for generating NCCN-based structured reports from mixed-language PDAC CT narratives.
• Human radiologist review is indispensable for accurate resectability determination due to model limitations.