小样本训练模型在宫颈癌放疗中自动勾画可行性研究  

A Feasibility Study of Automatic Delineation Using Small Sample Data Training Algorithm Model for The Cervical Cancer Radiotherapy

作　　者：陈飞[1,2] 胡静[2] 戴春华[2] 游涛[2] 耿长冉[1] 汤晓斌[1] CHEN Fei;HU Jing;DAI Chunhua;YOU Tao;GENG Changran;TANG Xiaobin(Department of Nuclear Science and Technology,Nanjing University of Aeronautics&Astronautics,Nanjing Jiangsu 210000,China;Department of Radiation Oncology,Affiliated Hospital of Jiangsu University,Zhenjiang Jiangsu 212000,China)

机构地区：[1]南京航空航天大学核科学与技术系,江苏南京210000 [2]江苏大学附属医院放疗科,江苏镇江212000

出　　处：《中国医疗设备》2021年第11期27-31,共5页China Medical Devices

基　　金：国家自然科学基金青年科学基金(11805100)。

摘　　要：目的评估小样本数据训练算法模型对于宫颈癌临床放疗中自动勾画的可行性。方法选取45例宫颈癌患者,由临床医生勾画临床靶区(Clinical Target Volume,CTV)及危及器官(Organs at Risk,OARs)。随机选择30例患者作为训练集传至AccuLearning平台进行自主训练,训练好的算法模型传至AccuContour平台,对剩余15例患者进行自动勾画,分析新模型自动勾画与手动勾画的相似性系数(Dice Similarity Coefficient,DSC),敏感性指数(Sensitive Index,SI),质心偏差(Deviation of Centroid,DC)以及体积差异系数(Deviation of Volume,DV),评估自动勾画的可行性。结果各个结构的DSC均值均高于0.85;除了CTV的SI=0.87,其余的均≥0.90;CTV和肠袋的DC均值介于3~6 mm,其余的均<3 mm;关于DV指数,最小的是骨髓3.86%,最大的是肠袋12.32%。结论小样本数据训练算法模型对于宫颈癌CTV及OARs的勾画具有较高的准确性,能较好地辅助临床工作。Objective To evaluate the feasibility of small sample data training algorithm model for automatic delineation in cervical cancer radiotherapy.Methods A total of 45 patients with cervical cancer were enrolled in this study.The clinical target volume(CTV)and organs-at-risk(OARs)were contoured by radiation oncologist.30 patients were randomly selected as training sets and transferred to AccuLearning platform for autonomous training.Then the trained algorithm model was transferred to AccuContour platform to automatically contour the CTV and OARs of the remaining 15 patients.Four evaluation indexes,including Dice similarity coefficient(DSC),sensitive index(SI),deviation of centroid(DC)and deviation of volume(DV),were analyzed to evaluate the feasibility of automatic delineation.Results The mean DSC value of each automatic delineation structure were all higher than 0.85.For the mean SI values,OARs were no less than 0.90 except that the CTV was 0.87.The mean DC values of CTV and bowelbag were 3-6 mm,and the others were less than 3 mm.Among the DV of all automatic contours,the minimum DV was 3.86%in bone marrow,and the maximum was 12.32%in bowelbag.Conclusion The small sample data training algorithm model can achieve high precision for CTV and OARs automatic delineation in patients with cervical cancer,which can better assist the work of clinicians.

关 键 词：深度学习 自动勾画 宫颈癌 临床靶区 危及器官 

分 类 号：R737.33[医药卫生—肿瘤]