JRadiEvo: A Japanese radiology report generation model enhanced by evolutionary optimization of model merging.

Abstract

Radiology report generation is an important application of artificial intelligence (AI), as the interpretation of medical images and the production of clinically relevant reports are time-consuming and cognitively demanding tasks, especially in high-volume settings such as chest X-ray screening. Recent advances in large language models (LLMs) and vision-language models (VLMs) have enabled substantial progress in automated medical report generation. However, most existing medical foundation models are trained primarily on English datasets, limiting their practicality in non-English-speaking regions such as Japan. Publicly available radiology datasets are overwhelmingly English, while constructing large-scale non-English datasets is costly and difficult because of translation effort and medical data privacy constraints. Existing adaptation methods, such as fine-tuning and continued pre-training, typically require large amounts of in-domain data. This makes them difficult to apply in low-resource medical language settings where large-scale annotated datasets are unavailable. Accordingly, this study asks: how can an accurate Japanese chest X-ray radiology report generator be developed without access to large-scale, curated Japanese medical image-report pairs? To address this challenge, we propose JRadiEvo, a Japanese chest X-ray report generation model built through evolutionary optimization of model merging. By combining pretrained models with complementary strengths in vision-language alignment, medical knowledge, and Japanese generation, JRadiEvo enables data-efficient adaptation without large-scale training. To the best of our knowledge, this is the first attempt to build a non-English medical vision-language model through evolutionary optimization of model merging. Despite using only 50 translated training samples from publicly available data, JRadiEvo outperforms CheXagent, a state-of-the-art model trained on approximately 8.5 million samples, in ROUGE-L and METEOR metrics. These results provide a proof of concept for extreme data-efficient adaptation in low-resource medical languages.