Automatic Structuring of Radiology Reports of Chest X-Ray Images Using Large Language Model

Radiology reports of diagnostic radiology are commonly written in natural language by clinical radiologists. Structured reports have been proposed, but their use in clinical practice is limited. If radiology reports could be automatically structured, they would be useful for data retrieval and dataset construction. However, not sufficient research has been done on the automatic structuring of radiology reports. In this study, we aim to automatically structure radiology reports of chest X-ray images using the Text-to-Text Transfer Transformer (T5), one of large language models. Medical Subject Headings (MeSH) was used as the terminology for automatic structuring. The OPENI dataset, which contains pairs of radiology reports and MeSH, was used to fine-tune T5. A total of 3337 radiology reports were collected from OPENI, and they were divided into 3000, 165, and 172 reports for training, validation, and test sets, respectively. These three sets were used for the training and inference with T5. Training was performed on several T5 of different sizes, and their inference results were compared. The clinical usefulness of the T5 with the best inference results was also evaluated by a radiologist.