训练集病例数对基于深度学习宫颈癌临床靶区及危及器官自动勾画的影响  

Effect of Training Volume on the Automatic Segmentation of Clinical Target Volume and Organs at Risk in Patients with Cervical Cancer Based on Deep Learning

作　　者：胡静[1] 陈飞[1] 龚筱钦 游涛[1] 张开军[1] 戴春华[1] HU Jing;CHEN Fei;GONG Xiaoqin;YOU Tao;ZHANG Kaijun;DAI Chunhua(Department of Radiation Oncology,Affiliated Hospital of Jiangsu University,Zhenjiang Jiangsu 212000,China)

机构地区：[1]江苏大学附属医院放疗科,江苏镇江212000

出　　处：《中国医疗设备》2022年第9期33-37,共5页China Medical Devices

基　　金：江苏省卫建委医学科研重点项目(ZDB2020022)。

摘　　要：目的探讨基于2D U-net深度学习网络模型实施宫颈癌临床靶区(Clinical Target Volume,CTV)及危及器官(Organs at Risk,OARs)自动勾画时训练集中病例数对自动勾画结果的影响。方法选取我院收治的140例宫颈癌患者的放疗CT图像,随机抽取120例患者CT图像数据作为深度学习训练集,其余20例作为测试集,运用基于2D U-net网络的AccuLearning(AL)平台训练生成5组自动勾画模型(训练量分别为15、30、60、90、120例),并对20例测试集进行自动勾画,采用相似性系数(Dice Similarity Coefficient,DSC)、豪斯多夫距离(Hausdorff Distance,HD)、体积相对偏差(Relative Volume Difference,RVD)指标比较自动勾画效果。结果CTV的DSC和RVD,肠袋的DSC、HD和RVD,直肠和膀胱的DSC以及左侧股骨头HD在5组不同训练量模型中的差异具有统计学意义(P<0.05),且上述指标随着训练量的增加呈较好趋势变化。结论基于AL平台对宫颈癌CTV及OARs自动勾画建模时,CTV可选90例建模;肠袋和直肠可选60例建模;膀胱、骨髓以及双侧股骨头可选15例建模。Objective To evaluate the effect of the amount of training data on the automatic segmentation of clinical target volume(CTV)and organs at risk(OARs)of cervical cancer patients based on deep learning using 2D U-net.Methods CT images of 140 patients with cervical cancer in our hospital were enrolled.CT image data of 120 patients were randomly selected as deep learning training set,and the rest of 20 cases were used as test set.The AccuLearning(AL)platform based on 2D U-net was used to train these datasets and generate five groups of automatic segmentation models(the amount of training datasets was 15,30,60,90,120 cases respectively).Meanwhile the other 20 cases were selected for automatic segmentation test.Three evaluation indexes,including the Dice similarity coefficient(DSC),Hausdorff distance(HD)and relative volume difference(RVD)were analyzed to compare the difference of automatic segmentation among the five groups,thereby discussing the effect of the amount of training data on automatic segmentation.Results DSC and RVD of CTV,the DSC,HD and RVD of bowel bag,the DSC of rectum and bladder and the HD of left femoral head showed statistically significant differences among the five models with different number training cases(P<0.05).Meanwhile these indexes showed a good trend of change with the increase of the number of training data.Conclusion When building the automatic segmentation models of CTV and OARs of patients with cervical cancer based on deep learning,90 cases of training data can be selected for CTV,60 cases for bowel bag and rectum,15 cases for bladder,marrow and bilateral femoral head.

关 键 词：深度学习 训练量 宫颈癌 临床靶区 危及器官 

分 类 号：R737.33[医药卫生—肿瘤]