基于VUMix-Net卷积神经网络的食管癌临床靶区自动勾画研究  

Automatic Segmentation of Clinical Target Volume for Esophageal Cancer Based on VUMix-Net Convolutional Neural Network

作　　者：金琳芝 宋平 汪晨宇 程欣宇 张贺铭 任润川 郭莹 张耀文[1] JIN Linzhi;SONG Ping;WANG Chengyu;CHENG Xinyu;ZHANG Heming;REN Runchuan;GUO Ying;ZHANG Yaowen(Department of Radiation Oncology of Anyang Tumor Hospital,The Affiliated Anyang Tumor Hospital of Henan University of Science and Technology,Henan Medical Key Laboratory of Precise Prevention and Treatment of Esophageal Cancer,Anyang,China,455000)

机构地区：[1]安阳市肿瘤医院,河南科技大学附属安阳市肿瘤医院,河南省食管癌精准防治医学重点实验室,河南安阳455000

出　　处：《食管疾病》2024年第4期252-259,共8页Journal of Esophageal Diseases

基　　金：河南省医学科技攻关计划联合共建项目(LHGJ20230863);安阳市医学科技攻关项目(2023C01SF026)。

摘　　要：目的提出一种新颖的三维(3D)V-Net和二维(2D)U-Net混合(VUMix-Net)的网络模型,用于自动勾画食管癌的临床靶区体积(clinical target volume,CTV),并评估其分割性能及应用价值。方法收集55例上段食管癌患者的计算机断层扫描(computed tomography,CT)图像数据,由1名具有10 a食管癌放疗临床经验的医师统一勾画CTV,同时应用V-Net、VUMix-Net自动勾画CTV。以戴斯相似系数(Dice similarity coefficient,DSC)和第95百分位数豪斯多夫距离(hausdorff distance,HD)为指标评估这两种模型自动分割性能。随机抽取10例人工勾画的食管癌放疗患者CT图像资料,应用VUMix-Net网络模型生成相应的自动勾画结果。两位主任医师在盲法条件下对这些结果进行临床评估。比较人工智能(artificial intelligence,AI)生成的勾画结果与人工勾画的参考标准(ground truth,GT)之间的评分差异,并分析两位医师评价的一致性。结果在CTV的自动勾画中,VUMix-Net的3D-DSC值和95HD值明显优于V-Net,差异均有统计意义,P值分别为0.006和<0.001。两位医师评估结果显示VUMix-Net模型勾画的CTV均符合临床实际应用情况。而且无论是医师A还是医师B,AI组和GT组的评估结果差异均无统计学意义,P值分别为0.222和0.361。无论是AI组还是GT组,医师A和医师B的评估结果也是没有显著差异,P值分别为0.638和0.761。结论基于新模型(VUMix-Net)卷积神经网络在自动勾画食管癌的CTV方面显示出一定的优势,并且能勾画出符合临床要求且与人工勾画质量相当的食管癌放疗靶区。Objective To propose a novel network architecture,VUMix-Net,which combines three-dimensional(3D)V-Net and two-dimensional(2D)U-Net,for automatically delineating the clinical target volume(CTV)of esophageal cancer,and evaluate its segmentation performance and application value.Methods CT image data from 55 patients with upper esophageal cancer were collected,their CTVs were uniformly delineated by a physician with 10 years of clinical experience in radiotherapy.3D V-Net and VUMix-Net were used to segment the CTVs simultaneously.The Dice similarity coefficient(DSC)and the 95th percentile Hausdorff distance(HD)are used as quantitative indicators to evaluate the automatic segm-entation performance of these two models.CT image data of 10 esophageal cancer radiotherapy patients who had undergone manual delineation were selected randomly,and the VUMix Net network model were applied to generate corresponding automatic delineation results.These results were then clinically evaluated by 2 chief physicians under blinded conditions.The differences in scores between the artificial intelligence(AI)group and the ground truth(GT)group were compared,and the consistency of the evaluations from two physicians were analyzed.Results In the automatic delineation of CTV,the 3D-DSC and 95HD values of VUMix Net were significantly better than those of V-Net,and the differences were statistically significant,with P values of 0.006 and less than 0.001,respectively.The evaluation results of the two physicians showed that the CTV delineated by the VUMix Net model conforms to the clinical application situation.The differences between the AI and GT groups were not statistically significant for either physician A or physician B,with P values of 0.222 and 0.361,respectively.And there was no significant difference between physician A and physician B for either the AI or GT groups,with P values of 0.638 and 0.761,respectively.Conclusion The new mode(VUMix-Net)convolutional neural network showed certain advantages in automatic CTV delineating,and co

关 键 词：卷积神经网络 自动勾画 食管癌 临床靶区 

分 类 号：R735.1[医药卫生—肿瘤]