面向各向异性3D-MRI图像超分辨率重建的ESRGAN网络

ESRGAN network for super-resolution reconstruction ofanisotropic 3D-MRI images

高分辨率磁共振图像(MRI,magnetic resonance images)能够提高疾病诊断精度,但高分辨率MRI图像的获取十分困难。基于深度学习的图像超分辨率(SR,super resolution)技术可有效地提高图像分辨率。近年来,生成对抗网络(GANs,generative adversarial networks)为3D-MRI图像SR重建提供了新思路。相较于传统的基于深度卷积神经网络(DCNN,deep convolutional neural network)的SR算法,GANs网络以人类视觉机制为目标,且引入判别函数,使重建3D-MRI图像更接近真实图像。研究采用增强超分辨率生成对抗网络(ESRGAN,enhanced super-resolution generative adversarial networks)对3D-MRI图像进行SR重建;并利用3D-MRI图像的跨层面自相似性,将重建任务降维到2D,在保证重建效果的基础上,减少了网络训练时间和内存需求。通过与其他传统算法和基于DCNN算法对比实验表明,提出的算法能够进一步提高3D-MRI图像的视觉质量。

High-resolution(HR)magnetic resonance images(MRI)can improve the accuracy of disease diagnosis,but it is very difficult to obtain high-resolution MRI.Image super-resolution(SR)technology based on deep learning can effectively improve image resolution.In recent years,the generative adversarial networks(GANs)have provided new ideas for 3D-MRI SR reconstruction.Compared with the traditional SR algorithm based on deep convolutional neural network(DCNN),the GANs network targets the human visual mechanism and introduces a discriminant function to make the reconstructed 3D-MRI closer to the real image.We introduced the enhanced super-resolution generative adversarial network(ESRGAN)to perform SR reconstruction of 3D-MRI,and used the cross-layer self-similarity of 3D-MRI to reduce the dimensionality of the reconstruction task to 2D.On the basis of ensuring the reconstruction effect,the proposed method can reduce network training time and memory requirements.Compared with other traditional algorithms and DCNN-based techniques,experimental results show that our proposed method can further improve the visual quality of SR 3D-MRI.