Breast MRI reports are complex and require expert interpretation, leading to variability in description and terminology. This study applies large language models (LLMs) to structure free-text Japanese breast MRI reports by category and regenerate natural Japanese sentences from structured data. Using 305 reports from Kyoto Prefectural University of Medicine, noun tokens were extracted with MeCab and used to create datasets processed by both local and remote LLMs. Evaluation with ROUGE, Sentence-BERT, and cosine similarity showed remote LLMs achieved higher similarity and 100% structuring accuracy, while local LLMs showed lower performance and frequent “unknown’’/blank outputs. In regeneration, remote LLMs reached up to 100% accuracy, whereas local LLMs scored 0%. Results suggest remote LLMs are practical for structuring breast MRI reports, but further improvements are needed for accurate re-textualization.