基于深度学习的直肠癌放疗靶区和危及器官自动勾画  

Automatic segmentation of clinical target volumes and organs-at-risk in radiotherapy for rectal cancer based on deep learning

作　　者：李林山 司梦远 苏坤普 肖姚 周德力 刘岩海 周鹏 罗佳 胡海 李梦侠 陈川 LI Linshan;SI Mengyuan;SU Kunpu;XIAO Yao;ZHOU Deli;LIU Yanhai;ZHOU Peng;LUO Jia;HU Hai;LI Mengxia;CHEN Chuan(Department of Oncology,Army Specialized Medical Center,Chongqing 400042,China;Institute of Physics,Chinese Academy of Sciences,Beijing 100190,China.)

机构地区：[1]陆军特色医学中心肿瘤科,重庆400042 [2]中国科学院物理研究所,北京100190

出　　处：《现代肿瘤医学》2024年第19期3757-3762,共6页Journal of Modern Oncology

基　　金：重庆市科卫联合医学科研项目重点项目(编号:2022ZDXM027)。

摘　　要：目的:基于改进Unet卷积神经网络模型,探讨模型应用于放射治疗中直肠癌靶区和危及器官分割的可行性。方法:研究回顾了120例直肠癌患者数据,随机选取80例作为训练集,20例作为验证集,20例作为测试集。自动分割勾画的目标包括了直肠癌的临床靶区(clinical target volume,CTV)、左侧股骨头、右侧股骨头、膀胱。采用戴斯相似性系数(Dice similarity coefficient,DSC)、豪斯多夫距离(Hausdorff distance,HD)和交并比(intersection over union,IoU)作为评价指标,网络模型自动勾画的结果与临床医生手动勾画结果进行比较,并与基于图谱的自动勾画技术(atlas-based automatic segmentation technique,ABAS)进行对比。结果:深度学习模型自动勾画CTV、膀胱、左侧股骨头和右侧股骨头的DSC值分别为0.90±0.06、0.95±0.11、0.98±0.01和0.96±0.05;95%HD值分别为(7.58±4.70)mm、(4.11±8.58)mm、(1.37±2.09)mm和(1.50±2.19)mm;IoU值分别为0.82±0.09、0.91±0.13、0.96±0.03和0.94±0.06。ABAS自动勾画CTV、膀胱、左侧股骨头和右侧股骨头的DSC值分别为0.83±0.13、0.68±0.27、0.89±0.12和0.88±0.13;95%HD值分别为(5.78±7.55)mm、(13.81±15.76)mm、(1.93±3.23)mm和(2.13±3.70)mm;IoU值分别为0.73±0.15、0.57±0.27、0.81±0.14和0.80±0.15。结论:基于改进Unet卷积神经网络模型在直肠癌的CTV和危及器官自动勾画任务中有较高准确率,应用于临床中能提高医生的工作效率和勾画一致性,有助于提升放射治疗的精准度,为后续自动化放疗计划设计的实现提供了支持。Objective:To explore the feasibility of applying an improved Unet convolutional neural network model to the segmentation of target areas and organs at risk in radiotherapy for rectal cancer.Methods:This study involved a retrospective analysis of data from 120 rectal cancer patients.A random selection of 80 cases were used for the training set,20 for the validation set,and 20 for the test set.The targets for automatic delineation included the rectal cancer clinical target volume(CTV),left and right femoral heads,and the bladder.The network model's automatic delineations were compared against manual delineations by clinical experts and the atlas-based automatic segmentation technique(ABAS).Evaluation metrics such as DSC coefficients,Hausdorff distances,and intersection over union(IoU)were employed.Results:The deep learning model yielded DSC coefficients for CTV,bladder,left femoral head,and right femoral head of 0.90±0.06,0.95±0.11,0.98±0.01,and 0.96±0.05,respectively.The 95%Hausdorff distances were(7.58±4.70)mm,(4.11±8.58)mm,(1.37±2.09)mm,and(1.50±2.19)mm,respectively.The IoU values were 0.82±0.09,0.91±0.13,0.96±0.03,and 0.94±0.06,respectively.In comparison,ABAS yielded DSC coefficients for CTV,bladder,left and right femoral heads of 0.83±0.13,0.68±0.27,0.89±0.12,and 0.88±0.13.95%HD values of(5.78±7.55)mm,(13.81±15.76)mm,(1.93±3.23)mm,and(2.13±3.70)mm and IoU values of 0.73±0.15,0.57±0.27,0.81±0.14,and 0.80±0.15.Conclusion:The improved Unet convolutional neural network model demonstrated high accuracy in the automatic delineation of CTV and organs-at-risk in rectal cancer.Its application in clinical settings can significantly enhance the efficiency and consistency of delineations made by medical professionals,thereby contributing to the precision of radiotherapy treatments,this study also paves the way for automated radiotherapy planning.

关 键 词：深度学习 直肠癌 放射治疗 自动勾画 

分 类 号：R735.3[医药卫生—肿瘤]