深度学习在子宫内膜癌术后临床靶区自动分割中的应用  

Application of deep learning in automatic segmentation of clinical target volume in brachytherapy after surgery for endometrial carcinoma

作　　者：薛娴[1] 王凯玥 梁大柱 丁静静 江萍[2] 孙全富[1] 程金生[1] 戴相昆[4] 付晓沙 朱静洋 周付根[7] XUE Xian;WANG Kaiyue;LIANG Dazhu;DING Jingjing;JIANG Ping;SUN Quanfu;CHENG Jinsheng;DAI Xiangkun;FU Xiaosha;ZHU Jingyang;ZHOU Fugen(National Institute for Radiological Protection,Chinese Center for Disease Control and Prevention(CDC),Beijing 100088 China;Department of Radiotherapy,Peking University Third Hospital,Beijing 100089 China;Northeastern University,Shenyang 110819 China;Department of Radiotherapy,Chinese People’s Liberation Army(PLA)General Hospital,Beijing 100039 China;Biomedical Research Centre,Sheffield Hallam University,Sheffield S11WB UK;Department of radiation oncology,Zhongcheng Cancer center,Beijing 100160 China;Beihang University,Beijing 100083 China)

机构地区：[1]中国疾病预防控制中心辐射防护与核安全医学所,北京100088 [2]北京大学第三医院放疗科,北京100089 [3]东北大学,辽宁沈阳110819 [4]中国人民解放军总医院放疗科,北京100039 [5]谢菲尔德哈勒姆大学,英国谢菲尔德S11WB [6]北京忠诚肿瘤医院肿瘤科,北京100161 [7]北京航空航天大学,北京100083

出　　处：《中国辐射卫生》2024年第4期376-383,共8页Chinese Journal of Radiological Health

摘　　要：目的评估3种深度学习(DL)算法在子宫内膜癌(EC)术后患者高剂量率近距离放射治疗(high-dose-rate brachytherapy,HDR BT)中,自动分割临床靶区(CTV)的应用结果。方法数据集由306名子宫内膜癌术后患者的计算机断层扫描(CT)图像组成,按比例分为训练集(246例)、验证集(30例)和测试集(30例)。比较3种深度卷积神经网络模型(3D U-Net、3D Res U-Net和V-Net)在CTV分割上的性能。采用定量指标分别为戴斯相似性系数(DSC)、豪斯多夫距离(HD)、豪斯多夫距离第95百分位数(HD95%)和交并比(IoU)。结果在测试阶段中,3D U-Net、3D Res U-Net和V-Net分割CTV得到的DSC平均值分别为0.90±0.07、0.95±0.06和0.95±0.06;HD平均值(mm)分别为2.51±1.70、0.96±1.01和0.98±0.95;HD95%平均值(mm)分别为1.33±1.02、0.65±0.91和0.40±0.72,IoU平均值分别为0.85±0.11、0.91±0.09和0.92±0.09。其中,V-Net分割结果与高级临床医生勾画结果更接近,CTV的分割时间<3.2 s,节省了临床医生的工作时间。结论V-Net在CTV分割方面表现最佳,定量指标和临床评估均优于其他模型。该方法与基准真实值高度一致,有效减少医生间差异,缩短治疗时间。Objective To evaluate the application of three deep learning algorithms in automatic segmentation of clinical target volumes(CTVs)in high-dose-rate brachytherapy after surgery for endometrial carcinoma.Methods A dataset comprising computed tomography scans from 306 post-surgery patients with endometrial carcinoma was divided into three subsets:246 cases for training,30 cases for validation,and 30 cases for testing.Three deep convolutional neural network models,3D U-Net,3D Res U-Net,and V-Net,were compared for CTV segmentation.Several commonly used quantitative metrics were employed,i.e.,Dice similarity coefficient,Hausdorff distance,95th percentile of Hausdorff distance,and Intersection over Union.Results During the testing phase,CTV segmentation with 3D U-Net,3D Res U-Net,and V-Net showed a mean Dice similarity coefficient of 0.90±0.07,0.95±0.06,and 0.95±0.06,a mean Hausdorff distance of 2.51±1.70,0.96±1.01,and 0.98±0.95 mm,a mean 95th percentile of Hausdorff distance of 1.33±1.02,0.65±0.91,and 0.40±0.72 mm,and a mean Intersection over Union of 0.85±0.11,0.91±0.09,and 0.92±0.09,respectively.Segmentation based on V-Net was similarly to that performed by experienced radiation oncologists.The CTV segmentation time was<3.2 s,which could save the work time of clinicians.Conclusion V-Net is better than other models in CTV segmentation as indicated by quantitative metrics and clinician assessment.Additionally,the method is highly consistent with the ground truth,reducing inter-doctor variability and treatment time.

关 键 词：深度学习模型 子宫内膜癌术后 高剂量率近距离放射治疗 临床靶区自动分割 

分 类 号：R816[医药卫生—放射医学]