摘要
头部MRI作为脑部临床检查手段之一,常用于脑卒中等疾病的诊断,而人工进行头部MRI的诊断依赖医生丰富的临床经验且会消耗大量人力.为了有效降低头部MRI的误诊率和漏诊率,减少医生的工作量,提出了一种通过微调大规模语言模型进行头部MRI自动诊断的方法,首先通过对头部MRI报告-诊断数据集进行预处理获得高质量数据,再以ChatGLM-6B为基础模型,并采用P-Tuning v2方法对该模型进行微调.为了克服部分表述的同义性使得用准确率难以对模型进行评估,提出采用平均Dice系数对模型进行评估.通过实验,模型可以达到0.890 4的平均Dice系数.
As one of the clinical examination methods for the brain,head MRI is commonly used for the diagnosis of diseases such as stroke.The manual diagnosis of head MRI relies on the rich clinical experience of doctors and consumes large amount of manpower.In order to effectively decrease the misdi-agnosis and missed diagnosis rates of head MRI,and reduce the workload of doctors,a method is proposed for automatic diagnosis of head MRI by fine-tuning large language models.Firstly,high-quality data is obtained by preprocessing the head MRI report-diagnosis dataset.Then,ChatGLM-6B is used as the base model,and the P-Tuning v2 method is employed to fine-tune the model.To overcome the synony-my of some expressions,making it difficult to evaluate the model with accuracy,the average Dice coefficient is proposed for model evaluation.Through experiments,the model achieved an average Dice coefficient of 0.8904.
作者
刘之洋
张明浩
杨东
柴超
张颖
Liu Zhiyang;Zhang Minghao;Yang Dong;Chai Chao;Zhang Ying(College of Eleetronic Informatiom and Optical Engineering,Nankai University,Tianjin 300350,China;Tianjin Key Laboratory of Optoelectronic Sensor and Sensing Network Technology,Tianjin 300350,China;Nankai Universily Affiliated Tianjin First Central Hospital,Tianjin 300192,China;Signals and Systerms Virtual Teaching and Research Section of Ministry of Education,Tianjin 300350,China)
出处
《南开大学学报(自然科学版)》
北大核心
2025年第5期100-109,共10页
Journal of Nankai University(Natural Sience)
基金
天津市自然科学基金重点项目(21JCZDJC01090)
天津市科学技术普及项目(23KPHDRC00080)
天津市高等学校研究生教育改革研究计划项目(B231005531)
天津市普通高等学校本科教学质量与教学改革研究计划项目(A231005509)。
