Large Language Model-Assisted Cleaning of Report-Derived Labels in a Large-Scale Chest CT Dataset

The study evaluates the use of GPT-5.4 for cleaning labels in the CT-RATE chest CT dataset, comprising 24,434 reports and 439,812 label instances across 18 categories. The model achieved a 96.4% agreement rate with existing labels, with a Cohen's kappa of 0.884, indicating high reliability, particularly in identifying discordant labels, which were validated by radiologists. This approach demonstrates the potential of LLMs in enhancing the quality of public imaging datasets, with the cleaned dataset set for public release to aid future research.