Research On MRI Super Resolution Algorithm Based On Deep Learning

Analyzing MRI images is one of the important methods for the diagnosis and treatment of abdominal diseases.High-resolution MRI images can help doctors to grasp the patient’s condition better,but its requirements for the precision of imaging equipment will lead to expensive hardware costs.Using super-resolution technology to improve image quality from the software direction can reduce the cost of high-resolution MRI images effectively.The superresolution algorithm based on deep learning is trained by inputting a large number of data sets,which can effectively use the prior knowledge of the image to reconstruct super-resolution images with rich details and clear textures.Compared with super-resolution algorithms based on traditional methods,the performance improved significantly.However,using deep learning to reconstruct abdominal MRI images has two main problems:(1)Due to the characteristics of dense abdominal organs and complex tissue structure,the network is disturbed by irrelevant tissues during the reconstruction process easily,and it is unable to fully learn the significant information in the MRI images,which has negative effect on the final reconstruction;(2)The model complexity of current super-resolution algorithms is high,most of them have a huge amount of parameters,and can only implement super-resolution reconstruction for a single scale,which affects the flexibility of the algorithm’s clinical application.In order to solve these problems,the work of this dissertation is mainly carried out from the following two aspects:1)In order to effectively solve the problems of inconspicuous boundaries,unclear abdominal organs display caused by high-frequency detail loss and inconvenient application of single-model for reconstruction of single-scale in abdominal MRI images during superresolution reconstruction effectively,a multi-scale super-resolution reconstruction network based on parallel channel-spatial attention mechanism was proposed.Firstly,a parallel channel-spatial attention residual block was proposed,the correlation between the key area and high-frequency information was obtained by the spatial attention module,and the channel attention module is used to study the weights of each channel’s response to key information,at the same time,the feature extraction layer was widened to increase the feature information flowing into the attention module.In addition,the weight normalized layer was added to ensure the efficiency of the network.Finally,meta-upscale module was applied at the end of the network to increase the flexibility and applicability of the network.2)In order to have a balance on the complexity and performance of the abdominal MRI image super-resolution algorithm model,a lightweight super-resolution algorithm based on multi-scale feature extraction was proposed.Firstly,a depthwise separable convolution module was designed,which using the characteristics of depthwise separable convolution,while ensuring the lightweight of the network,it increasing the network depth and expanding the dimension of the feature channels in the meantime,so that a large amount of relevant feature information can deeply flow into the network;then,a multi-scale feature extraction module was built in the second half of network to extract different scales feature information with semantic expression capabilities,making full use of the high-frequency details which contained in the input image,and improving network performance successfully;in addition,using the binarized feature fusion structure to solve the problem caused by deep networks,such as lost of feature and low-frequency information redundancy.Finally,a meta-upscale module was introduced to implement multi-scale super-resolution reconstruction at single model.3)In this dissertation,abdominal MRI image data sets were collected and sorted.Aiming at the problem of lack of image data,a geometric transformation data enhancement strategy was adopted to increase the number of training samples that could be processed.In terms of data preprocessing,high-resolution images were adopted bicubic interpolation to obtain lowresolution images with corresponding down-sampling scales.Finally,a series of ablation comparison experiments were designed to verify the effectiveness of the algorithm proposed in this dissertation.The experimental results show that the proposed algorithm was superior to other algorithms in terms of objective index and reconstruction effect while ensuring lower model complexity.