Development of a Large-Scale Dataset of Chest Computed Tomography Reports in Japanese and a High-Performance Finding...

Area of Science:

• Natural Language Processing in Medical Imaging
• Artificial Intelligence in Healthcare
• Multilingual Medical Data Analysis

Background:

• Lack of large-scale Japanese radiology datasets hinders AI development for medical imaging analysis.
• Japan's leading role in CT scanner use necessitates specialized Japanese language models.
• Existing multilingual models are insufficient for nuanced Japanese medical text.

Purpose of the Study:

• To develop a comprehensive Japanese CT report dataset for natural language processing (NLP) research.
• To establish a specialized language model for structured classification of radiology findings.
• To create a validated evaluation dataset for reliable model performance assessment.

Main Methods:

• Machine translation of the CT-RATE dataset into Japanese using GPT-4o mini.
• Development of CT-BERT-JPN, a BERT-based model for Japanese radiology text analysis.
• Expert radiologist validation of translated reports and model performance benchmarking against GPT-4o.

Main Results:

• Machine-translated reports preserved general structure but required expert refinement for medical terminology and context.
• Radiologist-revised translations showed statistically significant improvements in quality.
• CT-BERT-JPN outperformed GPT-4o in classifying 11 of 18 findings, achieving high F1-scores for most.

Conclusions:

• A robust Japanese CT report dataset and specialized model (CT-BERT-JPN) were successfully created.
• A hybrid approach of machine translation and expert validation yields high-quality medical datasets.
• Publicly available datasets and models will advance medical AI research in Japanese healthcare.