基于循环生成对抗网络生成头颅磁共振sDWI图像的方法研究

Research on the method of brain magnetic resonance synthetic DWI generation based on the cycle generative adversarial network

目的基于循环生成对抗网络(cycle generative adversarial network,CycleGAN),利用非配对患者头颅磁共振图像数据实现水抑制T2WI图像和弥散加权成像(diffusion weighted imaging,DWI)图像之间的相互转换,并评估生成的伪弥散加权成像(synthetic diffusion weighted imaging,sDWI)图像质量。材料与方法收集200例头颅水抑制T2WI图像以及DWI图像,训练集与测试集各100例,其中测试集含50例急性脑梗死。CycleGAN模型包含两个生成器与两个判别器。先基于卷积神经网络(convolutional neural networks,CNN)构建两个生成器,一个将水抑制T2WI图像转换为sDWI图像,另一个生成器将DWI图像转换成伪T2WI图像。再基于CNN构建两个判别器,用于对真实图像和生成的伪图像进行判别并更新参数。生成器与判别器不断交替工作完成CycleGAN模型训练。通过平均绝对误差(mean absolute error,MAE)、平均误差(mean error,ME)、峰值信噪比(peak signal-to-noise ratio,PSNR)和结构相似性(structural similarity,SSIM)以及主观评分评估sDWI图像质量。对50例急性脑梗死DWI图像与sDWI图像进行梗死灶分割,并计算DICE系数。结果测试集sDWI图像与真实DWI图像MAE为34.991±0.989,ME为15.982±0.978,PSNR为26.642±3.428,SSIM为0.927±0.039;80%以上sDWI图像无或仅有轻微图像失真、伪影;真实DWI与sDWI图像梗死灶分割后的DICE系数分别为0.898±0.324、0.849±0.259。结论CycleGAN模型和非配对图像数据可以生成质量较高的sDWI图像,为需要快速磁共振成像的患者减少扫描时间。

Objective:Based on cycle generative adversarial network(CycleGAN),using unpaired patient head MR image data to achieve mutual conversion between water-suppressed T2WI images and diffusion weighted imaging(DWI)images,and to evaluate the quality of the generated synthetic DWI images.Materials and Methods:Brain water-suppressed T2WI images and DWI images of 200 cases were collected.There were 100 cases in the training set and 100 cases in the test set,including 50 cases of acute cerebral infarction.CycleGAN model included two generators and two discriminators.Firstly,two generators were constructed based on convolutional neural networks(CNN).One generator converted water-suppressed T2WI images into synthetic-DWI images,and the other generator converted DWI images into synthetic-T2WI images.Then,two discriminators were constructed based on CNN,which were used to discriminate the real image and the generated synthetic image and update the parameters.The generator and discriminator work alternately to complete the training of CycleGAN model.The image quality of synthetic-DWI was evaluated by MAE,ME,PSNR,SSIM and subjective score.A total of 50 cases of acute cerebral infarction were divided into DWI images and sDWI images,and DICE coefficient was calculated.Results:The MAE,ME,PSNR and SSIM values of the synthetic and true DWI images were 34.991±0.989,15.982±0.978,26.642±3.428 and 0.927±0.039,respectively.More than 80%of the synthetic DWI images had no or only slight image distortion or artifact.The DICE coefficients of true DWI and synthetic DWI images after infarction segmentation were 0.898±0.324 and 0.849±0.259,respectively.Conclusions:The CycleGAN model and unpaired image data can generate high-quality synthetic DWI images and reduce the scanning time for patients who need rapid magnetic resonance imaging.