基于隐式退化模型的磁共振图像超分辨重建网络

Super-resolution reconstruction network based on implicit degradation model for magnetic resonance images

对于使用算法提升磁共振(MR)图像分辨率的研究,现有方法多停留在跨尺寸、同尺寸有监督的超分辨算法研究,提出一种基于隐式退化映射模型的MR图像超分辨重建网络SG-Diffusion,通过掩码自编码器隐式建模MR图像的退化过程,减小实验构建数据集与实际场景下MR图像的域差距,并基于隐式退化模型构建样本对,训练得到基于自引导扩散模型的MR图像重建网络,从而实现无监督同尺寸MR图像的空间分辨率提升。在fastMRI数据集的4倍加速采样脑部MR图像超分辨实验结果显示,本文提出的基于隐式退化模型的MR图像超分辨重建网络能够有效提升退化MR图像的空间分辨率,同时与基于显示退化模型的图像退化重建方法相比,本文提出的SG-Diffusion方法具有更好的重建效果。

Given that the existing methods of enhancing the resolution of magnetic resonance(MR)images by algorithms mainly focus on cross-size and same-size supervised super-resolution algorithms,a super-resolution reconstruction network(SG-Diffusion)for MR images is proposed based on an implicit degradation mapping model.The degradation process of MR images is implicitly modeled through a masked autoencoder,which reduces the domain gap between the experimental constructed dataset and the actual MR images,and the sample pairs are generated based on implicit degradation model.After training,a MR image reconstruction network based on self-guided diffusion model is obtained to realize the spatial resolution enhancement of unsupervised same-size MR images.The results of super-resolution experiments of 4-fold accelerated sampling brain MR images on fastMRI dataset show that the MR image super-resolution reconstruction network based on implicit degradation model proposed in the study can effectively improve the spatial resolution of degraded MR images,and that compared with the image degradation reconstruction method based on the explicit degradation model,the proposed SG-Diffusion method achieves better reconstruction results.