循环生成对抗网络基于颅脑MR图生成伪CT图模型

Generating pseudo CT images based on brain MR images model using cycle-consistent generative adversarial network

目的采用改进循环生成对抗网络(UCycleGAN)基于颅脑MR图映射模型生成伪CT图。方法对50例鼻咽癌颅脑MR图与CT图进行配准及预处理;以U-net网络并添加L1距离函数替换原始循环GAN(CycleGAN)模型生成器的深度残差网络。随机选取40例图像作为训练数据对UCycleGAN模型进行训练,将剩余10例用于测试;比较生成伪CT图与原始图像质量的差异,并与以ResNet、U-net的CycleGAN以及Pix2Pix生成的图像进行对比。结果相比其他模型,以UCycleGAN模型生成的伪CT图与原始CT图更为接近,体素平均绝对误差(MAE)为(81.45±3.87)HU,峰值信噪比(PSNR)为(34.13±3.28)dB,结构相似性(SSIM)为0.87±0.03。采用UCycleGAN模型生成的伪CT图的MAE小于、而SSIM明显大于其他3种模型(P均<0.05);UCycleGAN伪CT图的PSNR大于CycleGAN_ResNet图像(P<0.05)。结论利用UCycleGAN可基于颅脑MR图生成伪CT图;改良后CycleGAN模型的准确性更高。

Objective To establish brain MR images mapping models for generating pseudo CT images based on cycle-consistent generative adversarial network(UCycleGAN).Methods Brain MR images and CT images of 50 patients with nasopharyngeal carcinoma were registered and preprocessed.The depth residual network of the original cycle-consistent GAN(CycleGAN)model generator were replaced with U-net network,and L1 distance function was added.Images of 40 cases were randomly selected as training data to train the UCycleGAN model,while of the remaining 10 cases were used for testing.Imaging qualities were compared between generated pseudo CT and original CT,also among CycleGAN with ResNet,U-net and Pix2Pix.Results Compared with images obtained with other models,pseudo CT image generated with UCycleGAN model was more closer to original CT image,its average absolute error(MAE)of voxels was(81.45±3.87)HU,the peak signal-to-noise ratio(PSNR)was(34.13±3.28)dB,and the structural similarity(SSIM)was 0.87±0.03.MAE of pseudo CT images generated with UCycleGAN model were less,while SSIM were significantly larger than that of the other three models(all P<0.05),and PSNR of UCycleGAN pseudo CT images were greater than that of CycleGAN_ResNet(P<0.05).Conclusion Using UCycleGAN could generate pseudo CT images based on brain MR images,and the accuracy might be improved with CycleGAN model.