生成式大语言模型在中文放射医学领域的应用研究

Application of Generative Large Language Models in Chinese Radiology Domain

在中文放射医学领域中,影像学报告是临床决策的重要依据。因此,利用自然语言处理(NLP)技术来理解和学习影像学报告的文本内容,并以此辅助完成放射科临床工作,已成为该领域的重要研究方向。然而,在使用传统方法处理基于中文影像学报告的自然语言分类与生成任务时,仍然面临训练语料匮乏且涉及隐私、模型泛化能力较差等限制导致的综合性能不足的情况。针对上述问题,提出了一种基于本地高效微调大语言模型的中文放射医学领域自然语言任务解决方案。通过收集并构建大规模高质量中文影像学报告自然语言任务数据集,采用LoRA高效微调方法对开源大语言模型Baichuan2进行有监督微调训练,提出了能够同时解决四种中文放射医学领域临床任务的“龙影大模型”。提出了一套中文放射医学领域自然语言分类与生成任务评价体系。在来自两家中心的三个医学影像种类的报告数据集上进行了多组实验,并与几种典型现有方法进行了对比,结果显示所提方法在分类性能、文本总结与扩充能力和模型泛化性上表现更好。

In the Chinese radiology domain,radiology reports serve as a crucial basis for clinical decision-making.Therefore,utilizing natural language processing(NLP)technology to understand and learn from the textual content of radiology reports,thereby aiding radiological clinical work,has become an important research direction in this domain.However,when dealing with the natural language classification and generation tasks based on Chinese radiology reports using traditional methods,there are still challenges such as a lack of training corpora,privacy concerns,and poor model generalization capabilities,leading to insufficient overall performance.To address these issues,a solution for natural language tasks in the Chinese radiology domain based on locally efficient fine-tuning large language models is proposed.By collecting and constructing a large-scale,high-quality dataset for natural language tasks in the Chinese radiology reports,and employing the LoRA efficient fine-tuning method for supervised fine-tuning training of the open-source large language model Baichuan2,the“RadGPT”capable of solving four types of clinical tasks in the Chinese radiology domain simultaneously is proposed.A set of evaluation systems for natural language classification and generation tasks in the Chinese radiology domain is introduced.Multiple sets of experiments are conducted on three types of radiology report datasets from two centers,and comparisons are made with several typical existing methods.The results demonstrate that the proposed method performs better in terms of classification performance,text summarization and expansion capabilities,and model generalization.