基于DenseCUT网络由头部锥形束CT图像生成合成CT图像

提出一种由头部锥形束CT(CBCT)图像生成合成CT(sCT)图像的无监督深度学习网络，并与循环生成对抗（CycleGAN）网络及对比非配对转换（CUT）网络进行比较。本研究共获取56例脑部肿瘤患者的计划CT(pCT)和CBCT数据（其中49例用于训练，7例用于测试），分别使用CycleGAN网络、CUT网络以及本研究提出的密集对比非配对转换（DenseCUT）网络由CBCT图像生成sCT。DenseCUT网络有两点创新之处：将CUT网络与密集块网络结合；在损失函数中加入结构相似性。与pCT-CBCT相比，pCT-sCT(DenseCUT网络)的HU值平均绝对误差从34.38 HU降低到17.75 HU，峰值信噪比从26.19 dB提升到29.83 dB，结构相似性从0.78提升到0.87。本文方法可在不改变解剖结构的情况下从CBCT图像中生成高质量的sCT图像，同时降低图像伪影，使CBCT应用于剂量计算和自适应放疗计划成为可能。

Generation of synthetic CT image from head cone beam CT image using DenseCUT

Abstract: An unsupervised deep learning network is presented for generating synthetic CT (sCT) images from cone beam CT (CBCT) images of the head, and it is compared with cycle-consistent generative adversarial network (CycleGAN) and contrastive unpaired translation (CUT). After collecting the planning CT (pCT) images and CBCT images of 56 brain tumor patients (49 for training and 7 for testing), the sCT images are generated from CBCT images using CycleGAN, CUT, and the proposed dense contrastive unpaired translation (DenseCUT), separately. DenseCUT has two novelties, namely combining the CUT network with the dense block network, and adding structural similarity to the loss function. Compared with pCT-CBCT, pCT-sCT (DenseCUT) reduces the average absolute error of the HU from 34.38 HU to 17.75 HU, increases the peak signal-to-noise ratio from 26.19 dB to 29.83 dB, and elevates the structural similarity from 0.78 to 0.87. The proposed method can generate high-quality sCT images from CBCT images without altering the anatomical structures, while reducing image artifacts, which makes it possible for CBCT to be applied to dose calculation and adaptive radiotherapy planning.