Research On MR Images Super-resolution Based On Non-local Self-similarity And CNN

High-resolution MR images can provide more explicit structural and textural details than low-resolution MR images,helping clinicians diagnose patients more accurately.However,the acquisition conditions of high-quality MR images are harsh and restricted by various imaging parameters.Therefore,obtaining high-resolution MR images with precise details and textures without changing the imaging parameters by super-resolution reconstruction of low-resolution MR images has received extensive attention.Especially with the research progress of deep learning technology in recent years,super-resolution reconstruction based on deep learning has been widely studied and achieved remarkable results.However,most of the current models are often aimed at natural images,and the direct transfer to MR images cannot achieve the desired effect,and the information of the images themselves is not fully utilized.Therefore,this paper conducts in-depth research on this problem,and the specific research contents are as follows:(1)A super-resolution reconstruction method of MR images based on channel splitting attention network is proposed.In order to treat the features differently in the spatial dimension and mine the non-local self-similarity information contained in the image,a dual-input attention module is proposed,and by combining a dual-input attention module with a channel splitting network,a channel splitting attention network is designed.Based on the IXI dataset experiments,the results demonstrate that the channel splitting attention network can effectively perform super-resolution reconstruction of MR images.(2)A super-resolution reconstruction method of MR images based on non-local graph network is proposed.Aiming at the shortcomings that the current non-local modules cannot fully mine non-local self-similarity information by using local convolution,a non-local graph attention layer that combines non-local operations and graph convolutions is proposed to mine non-local self-similar information fully.Secondly,because of the problems existing in the information aggregation of graph convolution,a top-k node selection strategy is proposed to make the aggregated information more beneficial to the current node.Finally,a non-local graph network is constructed based on the non-local graph attention block.The experimental results demonstrate the effectiveness of the non-local graph network in super-resolution reconstruction of MR images.This article has a total of 32 figures,9 tables,and 107 references.