基于深度学习网络的医学核磁共振成像超分辨率重构实验

基于深度学习网络的医学核磁共振（MR）图像超分辨重建实验研究，提出并构建一个大规模的高质量用于MR图像超分辨的数据集，涵盖了头颅、膝盖、乳房以及头颈4个部位。通过数据质量筛选和不同低分辨率图像生成方式，在原始图像的高分辨率基础下，以×2、×3、×4的下采样尺度，原始MRI图像形成3种不同尺度下的MR图像数据集，同时给出不同部位超分辨难易程度分析。采用7个在自然图像的超分辨率领域中取得最好效果的深度学习网络，将它们迁移到MR图像中，学习低分辨率MR图像到高低分辨MR图像的映射关系，并对比分析这些深度学习网络在自然图像的超分辨效果。通过实验可以看出，深度学习网络在MR图像超分辨取得了比传统算法更好的效果，部分结果不亚于自然图像；不同部位的超分辨效果差异较大，难以以一个深度学习网络使不同部位均具有更好的超分辨效果。深度学习网络在MR图像超分辨将具有重要的应用价值和理论意义。

Experimental research on super-resolution reconstruction of medical MR image by deep learning network

Based on an experimental research on super-resolution reconstruction of medical MR(Magnetic Resonance,MR)images by deep learning network,a large-scale high-quality data set for MR images super-resolution was proposed and constructed,which covers 4 parts:skull,knees,breasts,and head&neck.With the original images as the high-resolution,the original MRI images was down-sampled with the scale of×2,×3,×4,and constituted MR image data at 3 different scales through data quality screening and different low-resolution image generation methods.The difficulty levels of super-resolution was anylyzed for different parts.7 deep learning networks that achieved the best results in the super-resolution field of natural images were adopted and transfered to MR images to learn the mapping relationship from low-resolution MR images to high-and low-resolution MR images,and the super-resolution effects of these deep learning networks in natural images were comparatively analyzed.Through the experiment,it can be seen that the deep learning networks have achieved better results than traditional algorithms in MR image super-resolution,and some results are no less than those in natural images.The super-resolution effects of different parts are quite different,and it is difficult to give each parts an equally good effect by only using a deep learning network.Deep l Experimental research on superresolution reconstruction of medical MR image by deep learning network earning networks will have important application value and theoretical significance in MR image super-resolution.