基于多尺度残差的生成对抗网络医学MRI影像超分辨率重建

Medical MRI Image Super-Resolution Reconstruction Network Based on Multi-Scale Residual Generative Adversarial Network

针对医学MRI影像成像过程中由于噪声、成像技术等各种干扰因素引起的图像细节丢失、纹理不清晰等问题,提出了基于多尺度残差的生成对抗网络医学MRI影像超分辨率重建算法.首先,利用多尺度残差组改进网络中的残差块,局部残差特征聚合模块将残差组聚合在一起,实现残差特征的非局部使用,减少局部特征在网络传播过程中的丢失;其次,通过注意力机制获取对关键信息响应程度更高的通道和空间特征信息,进而提升重建图像的细节纹理效果;然后,将低分辨率图像的梯度图转化为高分辨率图像的梯度图辅助重建超分辨率图像;最后,将恢复后的梯度图集成到超分辨率分支中,为超分辨率重建提供结构先验信息,从而明确地指导高质量超分辨率图像生成.对比基于梯度引导的结构保留超分辨率算法(SPSR),本文所提算法在×2,×3,×4尺度下的峰值信噪比平均提高了0.72 dB,实验结果表明所提算法重建出的医学MRI影像纹理细节更加丰富、视觉效果更加逼真.

In order to solve the problems of missing details and unclear texture of medical MRI image caused by noise,imaging technology and other interference factors in the process of medical imaging,a super-resolution reconstruction network of medical MRI image based on multi-scale residuals generative adversarial network was proposed.Firstly,the multi-scale residual group was used to improve the residual blocks in the network;the local residual feature aggregation module aggregated the residual groups together to realize the non-local use of residual features and reduce the loss of local features in the process of network transmission.Then,the attention module aimed to enable the network to obtain channel and spatial feature information with a higher degree of response to the key information,so as to improve the detail texture effect of reconstructed image.Next,the gradient image of the low resolution image was transformed into the gradient image of the high resolution image to assist the reconstruction of the super-resolution image.Finally,the restored gradient image was integrated into the super resolution branch to provide structural prior information for super resolution reconstruction,so as to clearly guide the generation of high quality super resolution image.Compared with the super-resolution algorithm based on structure-preserving super resolution with gradient guidance(SPSR),the peak signal-to-noise ratio of the proposed algorithm at×2,×3 and×4 scales were increased by an average of 0.72 dB,and the experimental results show that the texture details of medical MRI images reconstructed by the proposed algorithm are richer and the visual effect is more realistic.