中国中西医结合影像学杂志2025，Vol.23Issue(3)：358-362,5.DOI:10.3969/j.issn.1672-0512.2025.03.017
基于有监督微调的大语言模型利用胸部CT报告中的描述自动生成印象的初步研究
Study on automatic generation of impressions from chest CT reports using supervised fine-tuning large language models
摘要
Abstract
Objective:To explore the application for automatic generation of impressions from chest CT reports using supervised fine-tuning(SFT)large language models(LLMs).Methods:A retrospective study was conducted,collecting textual reports from 140,000 chest CT reports.After data cleaning,a data set was constructed from 75,884 reports.72,000 reports were taken as the training set for SFT training,and 3,844 reports were taken as the validation set for different LLMs and fine-tuning strategies.The performance of different LLMs and fine-tuning strategies was compared to identify the optimal model and tuning technique.The selected LLM was then trained using SFT.Finally,the accuracy and quality of impression generated by the original and fine-tuned models were compared.Natural language metrics,such as ROUGE and BLEU,were used for performance evaluation.Results:In basic models,the Baichuan2-13B model achieved BLEU-4 of 51.82 points,ROUGE-1 of 69.69 points,ROUGE-2 of 52.40 points,and ROUGE-L of 63.59 points.In ChatGLM-6B model,the LoRA method achieved ROUGE-1 of 64.04 points,ROUGE-2 of 42.39 points,ROUGE-L of 57.51 points,and BLEU-4 of 44.29 points.Both of them demonstrated superior performance and were selected for SFT training,and after training,the output quality of impressions by the LLM showed significant improvement in quantitative metrics.Conclusions:SFT-LLM demonstrats remarkable performance enhancements in generating impressions for clinical chest CT reports,producing results comparable to those of clinical practitioners.This highlights their potential for application in the domain of medical imaging report generation.关键词
有监督微调/大语言模型/胸部CT报告生成Key words
Supervised fine-tuning/Large language model/Chest CT reports generation引用本文复制引用
杨灵睿,杨学东,齐俊,周昱行,甄先通,史珊,孙黎,苏清华..基于有监督微调的大语言模型利用胸部CT报告中的描述自动生成印象的初步研究[J].中国中西医结合影像学杂志,2025,23(3):358-362,5.基金项目
中国中医科学院科技创新工程项目(CI2021A03302). （CI2021A03302）
