基于循环对抗生成网络的胸部锥形束CT校正及剂量计算准确性  

Cycle-GAN-based image correction for chest cone-beam CT and accuracy of dose calculation

作　　者：吴先想 牛振洋 蔡汉飞[1] 方美芳 李威 徐露 崔珍[1] WU Xianxiang;NIU Zhenyang;CAI Hanfei;FANG Meifang;LI Wei;XU Lu;CUI Zhen(Department of Radiation Oncology,the First Affiliated Hospital of Bengbu Medical College,Bengbu 233004,China;Department of Radiation Oncology,No.901 Hospital of PLA,Hefei 230031,China)

机构地区：[1]蚌埠医学院第一附属医院放疗科,安徽蚌埠233004 [2]解放军第901医院放疗科,安徽合肥230031

出　　处：《中国医学物理学杂志》2022年第7期811-816,共6页Chinese Journal of Medical Physics

基　　金：蚌埠医学院自然科学重点项目(2020byzd160)。

摘　　要：目的:采用循环对抗生成网络算法建立胸部锥形束CT(CBCT)校正模型,探讨该模型用于提升CBCT质量的可行性,评估校正的CBCT(CCBCT)用于剂量计算的准确性。方法:选择食管癌或肺癌患者已配准的CBCT和定位CT70例,随机选取其中60例作为训练集,用来训练循环对抗生成网络,生成CBCT的校正模型。剩余10例作为测试集,对CBCT、CCBCT和定位CT之间的CT值平均绝对误差、峰值信噪比、归一化互相关进行统计学分析。将原调强计划(CTPlan)移植到CCBCT上,生成CCBCTPlan,以CTPlan剂量分布为参考,对CCBCTPlan进行三维剂量γ分析。结果:CBCT经校正后散射伪影显著减少,CT值平均绝对误差降低了52.74%±6.47%,峰值信噪比和归一化互相关分别提高了7.95%±3.56%和1.68%±3.38%,差异均有统计学意义(t=18.47、-7.31、-6.77,P<0.05)。在2mm/2%、2mm/1%和1mm/1%条件下,CCBCTPlan三维剂量平均γ通过率分别为99.16%±0.34%、98.01%±0.72%、93.95%±1.62%。结论:基于循环对抗生成网络构建的CBCT影像校正模型用于提升CBCT影像质量是可行的,经校正的胸部CBCT可用于放疗剂量计算,为CBCT用于自适应放疗剂量计算奠定基础。Objective To establish a correction model for chest cone-beam CT(CBCT)using cycle-GAN algorithm for exploring the feasibility of using the model to improve CBCT image quality,and evaluating the accuracy of the corrected CBCT(CCBCT)for dose calculation.Methods The registered CBCT and positioning CT were collected from 70 patients with esophageal cancer and lung cancer.Sixty of them were randomly selected as training set to train cycle-GAN and generate the correction model for CBCT,while the remaining 10 cases were used as test set.The average absolute error of Hounsfield unit,peak signal-to-noise ratio,and normalized cross-correlation between CBCT or CCBCT and positioning CT were analyzed.CCBCT Plan was generated by transplanting the original intensity-modulated radiotherapy plan(CT Plan)to CCBCT.Taking the dose distribution of CT Plan as a reference,the gamma analysis on three-dimensional dose verification in CCBCT Plan was performed.Results Scattering artifacts were significantly reduced after CBCT was corrected.The average absolute error of Hounsfield unit was reduced by(52.74±6.47)%,and the peak signal-to-noise ratio and normalized cross-correlation were increased by(7.95±3.56)%and(1.68±3.38)%,respectively,with statistically significant differences(t=18.47,-7.31,-6.77;P<0.05).Under the criteria of 2 mm/2%,2 mm/1%,and 1 mm/1%,the average gamma passing rates of three-dimensional dose verification in CCBCT Plan were(99.16±0.34)%,(98.01±0.72)%,and(93.95±1.62)%,respectively.Conclusion Using the cycle-GAN-based correction model for CBCT to improve the CBCT image quality is proved to be feasible.The corrected chest CBCT can be used for radiation dose calculation,laying a foundation for CBCT to be used for adaptive radiation dose calculation.

关 键 词：循环对抗生成网络 锥形束CT 影像校正 剂量计算 自适应放疗 

分 类 号：R318[医药卫生—生物医学工程]