This study proposes introducing a local large language model (LLM) to address the limitations mentioned above by extracting disease names, conditions, and anatomical locations from radiology report sentences. The local LLM used is the Llama-3-ELYZA-JP-8B-AWQ, which is adapted for Japanese and quantized. The analysis data consisted of CT image reports, resulting in an average accuracy of 80.2%, significantly higher than the dictionary-matching method (61.0%). For comparison with remote LLMs, the local LLM achieved an average accuracy of 83.3% when structuring pseudo-CT reports generated by ChatGPT-3.5. Hallucination in the local LLM was also discussed.