Research On Medical Image Super-resolution Algorithm Based On Deep Learning

Medical image is an important reference in clinical diagnosis and obtaining high-resolution medical images can help doctors make more accurate medical diagnosis,thus to identify the disease as soon as possible and treat it in time.However,being subject to the limitation of imaging conditions and hardware devices,the resolution of collected medical images is generally low.In order to improve the display resolution of medical images effectively,this thesis which is based on the deep learning method studies the super-resolution algorithm for different characteristics of medical images.The existing super-resolution algorithms are mainly aiming at natural images,so they often use the high-level semantic information in the deep layer of convolutional neural network to reconstruct.However there are more low-level texture details in medical images,and the early symptoms are mostly small-scale targets.Therefore,this thesis proposes a superresolution network of medical image based on feature fusion.Firstly,depthwise separable convolution is used to improve the structure of residual block and the feature extraction ability of residual block is improved from two aspects of expanding depth and width,and the information of lower level in residual block is transmitted to the subsequent network through feature fusion.On the basis of improving the residual block,the information of residual blocks in front of the network is transferred directly to the reconstruction layer through feature fusion,so as to enhance the reconstruction ability of the network to the texture detail information through feature fusion at different levels.Subsequent experiments have proved that this algorithm has better reconstruction effect on the two kinds of self-built medical image data sets than the comparison algorithms,especially in the brain MRI data set with rich contents and more detailed texture information.Aiming at the problem that the first network has less improvement on the lung CT image data set with less effective information and discrete information distribution,this thesis proposes a super-resolution network of medical image based on mixed attention.Firstly,the information integration methods of attention structures of spatial domain and channel domain are improved separately to enhance the ability of integrating effective information in medical images.Then,a mixed attention module is realized by fusing attentional weight matrices of two structures and it is placed in front of the nonlinear mapping module to suppress the invalid information in the low-dimensional feature map,so as to enhance the attention of the subsequent network to the effective information.Then,the feature map extraction module is further improved based on the channel information filtering capability of this structure.By concatenating the feature maps of different scales' information obtained through different branches to the mixed attention module for filtering,so the network's ability to reconstruct effective information at different scales is enhanced.Experimental results show that this network has a greater improvement in the data set of lung CT images than the first network and further enhances the reconstruction effect of effective information of lung CT images.