基于深度神经网络的增强CT图像肾脏分割模型的建立  

Establishment of a renal cortex and medulla segmentation model in X-ray computed tomography images based on deep neural networks

作　　者：罗辉 李培 LUO Hui;LI Pei(Department of Radiology,Ningbo Yinzhou No.2 Hospital,Ningbo,Zhejiang 315000,China)

机构地区：[1]宁波市鄞州区第二医院影像科,浙江宁波315000

出　　处：《重庆医学》2025年第3期630-634,共5页Chongqing Medical Journal

基　　金：2024年度浙江省基础公益研究计划项目(LY24H160002);2020年浙江省医药卫生科技计划项目(2020KY896)。

摘　　要：目的在增强CT图像上构建基于深度神经网络的肾脏自动分割模型。方法收集2019年1月至2022年10月该院64例慢性肾脏疾病(CKD)患者的肾脏动脉期图像,根据血肌酐估算肾小球滤过率(eGFR)将其分为轻度肾损伤组、中度肾损伤组、重度肾损伤组和对照组,每组16例。采用ITK-Snap软件对图像进行逐层勾画,勾画区域包括肾实质及肾皮质。将数据集随机分为训练集和测试集,其中训练集40例(每组10例),测试集为24例(每组6例),构建肾实质及肾皮质分割模型并验证;比较肾实质容积及皮质容积的分割性能量化结果;对比4组图像测试集和模型Dice值,利用该模型定量评价肾实质及肾皮质容积,并评估其准确性。结果基于深度神经网络的增强CT图像肾脏分割模型对肾实质容积和皮质容积分割性能的量化结果显示,肾实质Dice值为93.53%,肾皮质Dice值为81.48%。各组图像肾实质容积和肾皮质容积Dice值差异无统计学意义(F=3.467、4.972,P>0.05)。结论构建的基于深度神经网络的增强CT图像肾脏分割模型,能够用于肾脏实质及皮质的分割,取得的数据可靠。Objective To construct an automatic kidney segmentation model based on deep neural network on enhanced CT images.Methods The renal arterial phase images of 64 patients with chronic kidney disease(CKD)were collected from January 2019 to October 2022.According to blood creatinine estimation of glomerular filtration rate(eGFR),they were divided into the mild renal injury group,the moderate renal injury group,the severe renal injury group and the control group,16 in each group.ITK-Snap software was used to outline the images layer by layer,and the areas outlined were renal parenchyma and renal cortex.The data set was randomly divided into training sets and test sets,including 40 training sets(10 in each group)and 24 test sets(6 in each group).Segmentation models of renal parenchyma and cortex were obtained and verified.The quantification results of renal parenchymal volume and cortical volume segmentation were compared.Four groups of image test sets were compared with the Dice values of the model to discuss the quantitative evaluation of kidney and renal cortex volume with this model,and evaluate its accuracy.Results The results of quantification of renal parenchymal volume and cortical volume segmentation performance by enhanced CT kidney segmentation model based on deep neural network showed that the Dice value of renal parench yma was 93.53%and that of renal cortex was 81.48%.There was no significant difference in Dice val ues of renal parenchymal volume and renal cortex volume among all the groups(F=3.467,4.972,P>0.05).Conclusion The enhanced CT image kidney segmentation model based on deep neural network established can be used to segment kidney parenchyma and cortex,and the obtained data are reliable.

关 键 词：计算机断层扫描 肾小球滤过率 肾皮质容积 肾实质容积 

分 类 号：R816.7[医药卫生—放射医学]