基于循环生成对抗网络的鼻咽癌CBCT图像修正

目的:利用循环生成对抗网络模型（CycleGAN）进行锥形束CT （CBCT）图像迁移，生成伪CT（sCT）图像，从而实现CBCT图像的HU值矫正。方法:回顾性分析在福建省肿瘤医院行放射治疗的鼻咽癌患者39例，所有患者均接受临床CT与CBCT扫描。以CBCT图像为基准，采用刚性配准算法对临床CT和CBCT进行配准，获得重采样计划CT（pCT）。经阈值分割及形态学处理获取配对影像的外轮廓内部区域作为掩膜，对配对影像进行掩膜操作及归一化预处理。建立CycleGAN神经网络，训练sCT生成模型。基于体素点计算平均绝对误差（MAE）和平均误差（ME），用于比较测试集sCT与pCT之间的差异。结果:测试集的sCT图像与pCT图像相比较，在体外轮廓内的MAE和ME分别为（99.00±15.37） HU和（-24.00±12.64） HU；软组织区域的MAE和ME分别为（48.00±7.45） HU和（-7.00±8.96） HU。结论:CycleGAN能修正CBCT图像的HU值，迁移生成的sCT图像具有与pCT图像近似的HU值及平滑性，可用于放射治疗剂量计算。

CBCT image correction for nasopharyngeal carcinoma based on cycle-consistent generative adversarial network

Abstract: Objective To realize HU correction for cone-beam computed tomography (CBCT) image by translating CBCT image into synthesized CT (sCT) via cycle-consistent generative adversarial network (CycleGAN). Methods Thirty-nine patients with nasopharyngeal carcinoma who underwent radiotherapy in Fujian Provincial Cancer Hospital were enrolled in the study and received clinical CT and CBCT scans. Taking CBCT image as benchmark, rigid registration algorithm was used for registration between clinical CT and CBCT, thereby obtaining resampling planning CT (pCT). Threshold segmentation and morphological processing were used to obtain the body region of the paired image as a mask and moreover, masking and normalizing pre-processing were carried out on the paired images. CycleGAN was established and then used to generate sCT images. Mean absolute error and mean error which were calculated based on voxels were adopted for comparing the differences between sCT and pCT. Results The comparison between sCT images and pCT images of test set showed that mean absolute error and mean error were (99.00±15.37) HU and (-24.00±12.64) HU for the whole body, and (48.00±7.45) HU and (-7.00±8.96) HU for soft tissues. Conclusion CycleGAN can be used for the HU correction for CBCT image, and the generated sCT image which has similar HU value and smoothness with pCT image can be used for dose calculation in radiotherapy.