基于nnU-Net增强CT对肝癌自动分割性能影响的研究  

Study on effect of nnU-Net enhanced CT on the automatic segmentation performance of liver cancer

作　　者：张庆 金成 李峥 陈大同[1] 许东滨 孙跃 梁明辉[1] Zhang Qing;Jin Cheng;Li Zheng;Chen Datong;Xu Dongbin;Sun Yue;Liang Minghui(Medical Technology College of Qiqihar Medical University,Qiqihar,Heilongjiang 161006,China;School of Biomedical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;The Second Affiliated Hospital of Qiqihar Medical University,Qiqihar,Heilongjiang 161006,China;The Third Affiliated Hospital of Qiqihar Medical University,Qiqihar,Heilongjiang 161006,China)

机构地区：[1]齐齐哈尔医学院医学技术学院,黑龙江齐齐哈尔161006 [2]上海交通大学生物医学工程学院,上海200240 [3]齐齐哈尔医学院附属第二医院,黑龙江齐齐哈尔161006 [4]齐齐哈尔医学院附属第三医院,黑龙江齐齐哈尔161006

出　　处：《齐齐哈尔医学院学报》2025年第5期401-405,共5页Journal of Qiqihar Medical University

基　　金：齐齐哈尔市科技攻关项目(LHYD-202016)。

摘　　要：目的探讨不同训练集大小及不同增强期影像对肝癌患者增强CT影像自动分割效果的影响。方法选取2017年6月-2024年9月齐齐哈尔医学院附属医院和哈尔滨医科大学第四医院收治的100例肝细胞癌患者的增强CT影像数据。将数据按8︰1︰1比例随机分为训练集(80例)、验证集(10例)和测试集(10例)。为分析训练集大小对模型性能的影响,将训练集进一步划分为A组(80例)、B组(60例)、C组(40例)、D组(20例)和E组(10例);按增强期分为动脉期、门静脉期和延迟期三组。以E组作为基础模型,使用nnU-Net模型对各组影像进行自动分割,并比较不同组间的Dice相似系数(DSC)、95%Hausdorff距离(HD95)和交并比(Intersection over Union,IoU)。结果与B组、C组、D组和E组相比,A组(训练集80例)的分割效果最佳,DSC为0.82,IoU为0.71,HD95为62.09,差异具有统计学意义(P<0.05)。在不同增强期组中,动脉期组的分割效果优于静脉期组和延迟期组,DSC为(0.82±0.09),IoU为(0.71±0.13),HD95为(62.09±66.35),差异具有统计学意义(P<0.05)。结论采取较大训练集以及动脉期增强CT影像训练nnU-Net模型,可显著提高肝癌CT影像自动分割效果。Objective To explore the impact of varying training set sizes and different enhancement periods on the automatic segmentation of enhanced CT images in patients with liver cancer.Methods The enhanced CT image data of 100 patients with hepatocellular carcinoma who admitted to the Affiliated Hospital of Qiqihar Medical College and the Fourth Hospital of Harbin Medical University between June 2017 and September 2024 were selected for analysis.The data were randomly divided into a training set(80 cases),a verification set(10 cases)and a test set(10 cases)in a ratio of 8︰1︰1.The training set was subdivided into six additional groups:Group A(80 cases),Group B(60 cases),Group C(40 cases),Group D(20 cases),and Group E(10 cases)for analyzing the impact of training set size on model performance.The data were divided into three phases:arterial,portal vein,and delayed.The nnU-Net model was employed to automatically segment the images of each group,using group E as the basic model.The Dice similarity coefficient(DSC),95%Hausdorff distance(HD95)and Intersection over Union(IoU)were compared among different groups.Results Group A(training set of 80 cases)exhibited the most optimal segmentation effect compared to groups B,C,D and E.The DSC,IoU and HD95 values were 0.82,0.71 and 62.09 respectively with statistical significance(P<0.05).In different enhancement stages,the arterial stage group exhibited superior segmentation efficacy compared to the venous stage group and the delayed stage group.The DSC,IoU,and HD95 values were(0.82±0.09),(0.71±0.13),and(62.09±66.35),respectively(P<0.05).Conclusions The utilization of a larger training set and enhanced CT images in the arterial phase to train the nnU-Net model has been demonstrated to markedly enhance the automatic segmentation performance of CT images of liver cancer.

关 键 词：肝细胞癌 增强CT 深度学习 期相对比 自动分割 

分 类 号：R73[医药卫生—肿瘤]