甲状腺相关眼病放射治疗靶区自动勾画的人工智能训练研究  

Study on AI training of automatic delineation of target region of radiotherapy for TAO

作　　者：劳峥 石慧烽[1] 樊文慧 毕帆 涂文勇[1] LAO Zheng;SHI Hui-feng;FAN Wen-hui(Department of Oral Maxillofacial Head and Neck Oncology,Shanghai Ninth People’s Hospital,Shanghai Jiao Tong University School of Medicine,Shanghai 200011,China)

机构地区：[1]上海交通大学医学院附属第九人民医院口腔颌面头颈肿瘤科,上海交通大学口腔医学院国家口腔医学中心,国家口腔疾病临床医学研究中心,上海市口腔医学重点实验室,上海200011

出　　处：《中国医学装备》2022年第10期13-17,共5页China Medical Equipment

基　　金：上海黄浦区产业扶持基金(XK2020012)“三维个体化空腔模体转化研究”。

摘　　要：目的:通过建立甲状腺相关眼病(TAO)放射治疗靶区人工智能训练的自动勾画模型,测试及对比分析验证其准确性和可靠性。方法:选取于医院行放射治疗的80例TAO患者,分批将30例、40例、50例、60例和70例5组病例的手动勾画靶区载入到AccuLearning人工智能软件中进行训练,分别建立Model 1组(30例)、Model 2组(40例)、Model 3组(50例)、Model 4组(60例)和Model 5组(70例)的5组自动勾画训练模型;其余10例病例作为额外测试集组,在AccuContour人工智能软件中分别采用各模型行自动勾画,与其手动勾画靶区在相似性系数(DSC)、豪斯多夫距离(HD)和体积差值百分比(ΔV%)上行对比分析;将手动靶区的放射治疗计划载入到自动勾画靶区中进行靶区剂量分布评价系数接受95%、100%及110%处方剂量的最小百分体积(V_(95)、V_(100)、V_(110))的比较,并对各模型的组内一致性相关系数(ICC)进行分析,评价自动勾画的准确性和可信度。结果:5组模型的DSC值均>0.89,HD值范围为0.18~0.3 mm,自动勾画与手动勾画靶区的平均ΔV%<2%,自动勾画靶区的平均V_(95)、V_(100)及V_(110)分别在99%、95%和1%左右,各组ICC值均>0.9,表明自动勾画训练模型的准确性好,一致性和可信度高。结论:通过AccuLearning人工智能软件对TAO放射治疗靶区的自动勾画效果较为满意,但日后仍需要更完善的临床研究。Objective:To test,compare,analyze and verify the accuracy and reliability of the automatic delineation mode through established that of artificial intelligence(AI)training of radiotherapy target region of thyroid associated ophthalmopathy(TAO).Methods:80 TAO patients who underwent radiotherapy in hospital were selected.The target regions of manual delineation of 5 groups which respectively included 30 cases,40 cases,50 cases,60 cases and 70 cases were loaded in the AccuLearning AI software to conduct train.The training models of automatic delineation of 5 groups included Model 1 group(30 cases),Model 2 group(40 cases),Model 3 group(50 cases),Model 4 group(60 cases)and Model 5 group(70 cases)were respectively established.The remaining 10 cases were used as extra test set group.In AccuContour AI software,each model was respectively adopted to undergo automatic delineation,which was compared and analyzed with the target region of manual delineation at dice similarity coefficient(DSC),Hausdorff distance(HD)and percentage of volume difference(ΔV%).The radiotherapy plans of manual target region was loaded in the target region of automatic delineation to implement the comparisons of minimum percentage volume(V_(95),V_(100),V_(110))when evaluation coefficient of dose distribution of target region received respectively 95%,100%and 110%of prescribed dose.The intragroup correlation coefficient(ICC)of each model was analyzed,and the accuracy and reliability of automatic delineation were further evaluated.Results:The DSC values of 5 groups were all larger than 0.89,and the HD values range of them were from 0.18 to 0.3 mm.And the averageΔV%of the target regions of automatic delineation and manual delineation were less than 2%.And the average V_(95),V_(100) and V_(110) of target region of automatic delineation respectively were at approximate 99%,95%and 1%.In addition,the ICC value of each group were larger than 0.9.These results indicated that the accuracy of the training model of automatic delineation was favorable,and the

关 键 词：甲状腺相关眼病(TAO) 放射治疗 靶区 自动勾画 人工智能 突眼 

分 类 号：R816.6[医药卫生—放射医学]