Background: Automated structuring of radiology reports is essential for data utilization and the development of medical artificial intelligence models. However, manual annotation by experts is labor-intensive, and processing real clinical data through commercial large language models (LLMs) presents significant privacy risks. These challenges are particularly pronounced for non-English languages like Japanese, where specialized medical corpora are scarce. While synthetic data generation offers a potential privacy-preserving alternative, its effectiveness in capturing complex clinical nuances—such as negation and contextual dependencies—to train robust classification models without any real-world training data has not been fully established. Objective: This study aimed to develop a context-aware sentence classification model for Japanese radiology reports using an entirely synthetic training pipeline, thereby eliminating reliance on real-world clinical data during the development phase. Furthermore, we sought to evaluate the generalizability of this approach by validating the model’s performance on diverse, multi-institutional, real-world reports. Methods: Japanese radiology reports (n=3104) were generated using GPT-4.1 and automatically annotated at the sentence level into 4 categories (background, positive finding, negative finding, and continuation) using GPT-4.1-mini. The synthetic data were partitioned into training (n=2670), validation (n=334), and test (n=100) sets. We fine-tuned several models, including lightweight local LLMs (Qwen3 and Llama 3.2 series) using low-rank adaptation and Japanese text classification models (Bidirectional Encoder Representations from Transformers [BERT]-base Japanese v3, Japanese Medical Robustly Optimized BERT Pretraining Approach [JMedRoBERTa]-base, and ModernBERT-Ja-130M). External validation was performed using 280 real-world reports (3477 sentences) from 7 institutions in the Japan Medical Image Database, with ground-truth labels established by board-certified radiologists. Evaluation metrics included accuracy, macro-averaged (macro ) score, and positive predictive value for positive findings (PPV_1). Results: All models achieved high performance on the synthetic test set (accuracy: 0.938‐0.951; macro -score: 0.924‐0.940). Overall performance declined on the external validation dataset (accuracy: 0.783‐0.813; macro -score: 0.761‐0.790), reflecting distributional differences between synthetic and real-world reports; however, PPV_1 remained stable and high across datasets (eg, 0.957 on the synthetic test set vs 0.952 on the external validation dataset for Qwen3 [4B]). Parsing errors occurred in LLM-based approaches (19‐260 sentences, 0.55%‐7.48% in the external dataset). Conclusions: This study demonstrates the feasibility of developing context-aware sentence classification models for Japanese radiology reports using a training pipeline based entirely on synthetic data. The stability of PPV_1 indicates that the models successfully captured the essential clinical terminology and linguistic patterns required to identify positive findings in real-world reports, despite the observed performance degradation during external validation. This approach substantially reduces manual annotation requirements and privacy risks, providing a scalable foundation for constructing structured radiology datasets to support the development of clinically relevant medical artificial intelligence models.

arXiv:2604.02207v1 Announce Type: new Abstract: Background: Accurate translation of radiology reports is important for multilingual research, clinical communication, and radiology education, but the validity of LLM-based evaluation remains unclear. Objective: To evaluate the educational suitability of LLM-generated Japanese translations of chest CT reports and compare radiologist assessments with LLM-as-a-judge evaluations. Methods: We analyzed 150 chest CT reports from the CT-RATE-JPN validation set. For each English report, a human-edited Japanese translation was compared with an LLM-generated translation by DeepSeek-V3.2. A board-certified radiologist and a radiology resident independently performed blinded pairwise evaluations across 4 criteria: terminology accuracy, readability, overall quality, and radiologist-style authenticity. In parallel, 3 LLM judges (DeepSeek-V3.2, Mistral Large 3, and GPT-5) evaluated the same pairs. Agreement was assessed using QWK and percentage agreement. Results: Agreement between radiologists and LLM judges was near zero (QWK=-0.04 to 0.15). Agreement between the 2 radiologists was also poor (QWK=0.01 to 0.06). Radiologist 1 rated terminology as equivalent in 59% of cases and favored the LLM translation for readability (51%) and overall quality (51%). Radiologist 2 rated readability as equivalent in 75% of cases and favored the human-edited translation for overall quality (40% vs 21%). All 3 LLM judges strongly favored the LLM translation across all criteria (70%-99%) and rated it as more radiologist-like in >93% of cases. Conclusions: LLM-generated translations were often judged natural and fluent, but the 2 radiologists differed substantially. LLM-as-a-judge showed strong preference for LLM output and negligible agreement with radiologists. For educational use of translated radiology reports, automated LLM-based evaluation alone is insufficient; expert radiologist review remains important.