基于深度学习的临床靶区与危及器官自动勾画方法在宫颈癌中的应用  

Application of Automatic Delineation Method of Clinical Target Volume and Organsat-Risk Based on Deep Learning in Cervical Cancer

作　　者：杨涛[1] 丛小虎 戴相昆[1] 葛瑞刚[1] 徐寿平 龚璇[1] Yang Tao;Cong Xiaohu;Dai Xiangkun;Ge Ruigang;Xu Shouping;Gong Xuan(The First Medical Center of Liberation Army General Hospital,Beijing 100853,China;Qingdao Central Hospital,Qingdao Shandong 266042,China;Cancer Hospital,Chinese Academy of Medical Sciences,Beijing 100021,China)

机构地区：[1]解放军总医院第一医学中心,北京100853 [2]青岛市中心医院,山东青岛266042 [3]中国医学科学院肿瘤医院,北京100021

出　　处：《医疗装备》2023年第24期1-4,9,共5页Medical Equipment

基　　金：国家重点研发计划(2021YFE0202500)。

摘　　要：目的探讨基于深度学习的临床靶区(CTV)与危及器官(OAR)自动勾画方法在宫颈癌中的应用,并评估其勾画精度,为临床提供参考。方法选取2021年1月至2022年12月中国人民解放军总医院第一医学中心130例宫颈癌患者的CT图像,由放疗医师统一勾画CTV及OAR。将其中100例患者的CT图像传至Manteia AccuLearning自主学习平台进行模型训练,并将训练好的模型导出至AccuContour自动勾画平台,对剩余30例患者进行CTV及OAR的自动勾画,与放疗医师手工勾画的结构进行比较。通过Dice相似系数(DSC)、95%Hausdorff距离(HD95)和对称位置的平均表面距离(ASSD)指标评价勾画效果。结果CTV的平均DSC为0.793,平均HD95为3.309 cm,平均ASSD为5.420 mm;对于OAR,除十二指肠和肠道外,左右肾、膀胱、左右股骨头、脊髓和直肠的平均DSC均≥0.836,平均HD95均≤3.323 mm,平均ASSD均≤1.881 mm。结论基于自主深度学习训练平台的OAR勾画方法可较为准确地实现宫颈癌患者CTV及OAR的自动勾画,提高工作效率。高度集成、流程化的深度学习自主训练平台可简便高效地完成对数据的清洗、参数调试与模型训练,大大提高临床应用的方便性,满足临床使用需求。Objective To explore the application of automatic delineation method of clinical target volume(CTV)and organs-at-risk(OAR)based on deep learning in cervical cancer,and evaluate its delineation accuracy,so as to provide reference for clinic practice.Methods With the selection of CT images of 130 patients with cervical cancer at The First Medical Center of Chinese PLA General Hospital from January 2021 to December 2022,the CTV and OAR were uniformly delineated by radiologists.The CT images of 100 patients were transmitted to the Manteia AccuLearning self-learning platform for model training,and then the trained models were exported to the AccuContour automatic delineation platform for automatic delineation of CTV and OAR on the remaining 30 patients.In addition,the structures of automatic sketching were compared with those of manual sketching by radiologists.The sketching effect was evaluated through Dice similarity coefficient(DSC),Hausdorff distance 95(HD95),and average symmetric surface distance(ASSD)indicators.Results The average DSC of CTV was 0.793,the average HD95 was 3.309 cm,and the average ASSD was 5.420 mm;For OAR,except for the duodenum and intestines,the average DSC of the left and right kidneys,bladder,left and right femoral heads,spinal cord,and rectum was≥0.836,the average HD95 was≤3.323 mm,and the average ASSD was≤1.881 mm.Conclusion The OAR delineation method based on autonomous deep learning training platform can accurately achieve automatic delineation of CTV and OAR in cervical cancer patients,improving work efficiency.A highly integrated and streamlined deep learning autonomous training platform can easily and efficiently clean data,debug parameters,and train models,greatly improving the convenience of clinical applications and meeting clinical needs.

关 键 词：宫颈癌 临床靶区 危及器官 深度学习 自动勾画 

分 类 号：R318[医药卫生—生物医学工程]