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

With the development of computer-aided medical diagnosis system,high-quality medical images play an important role in the accurate diagnosis of diseases.High resolution medical images can provide rich structural details,which is conducive to the accurate analysis and quantitative measurement of diseases.However,in practical application scenarios,image resolution is limited by acquisition equipment cost,transmission bandwidth,signal-to-noise ratio and time constraints,so it is difficult to obtain high-resolution images with sharp edges and no blocky blur.Image super-resolution reconstruction is to use image restoration algorithm to reconstruct low-resolution image obtained by acquisition equipment into high-resolution image.It is one of the research hotspots in the field of image processing,and has important application value in medical image auxiliary diagnosis.This paper analyzes and researches the image super-resolution reconstruction method in depth,and proposes a deep learning-based medical image super-resolution reconstruction algorithm,which realizes the mapping from low-resolution images to high-resolution images.The reconstructed medical images have more texture details and higher pixel information.The main work of this article is as follows:(1)A capsule residual deep neural network(ResCapsNet)based on routing structure is proposed and used for super-resolution reconstruction of medical CT images.First of all,the original image is upsampled to the desirable size by bicubic interpolation to obtain the enlarged low-resolution image;secondly,the image feature is extracted by using the convolution layer and encapsulated in a capsule structure.Then,the obtained capsule residual feature map is fed to the neural network.The high-resolution details of the predicted image can be extracted well using the extracted feature from the CapsNet.Finally,the predicted capsule residual feature map is merged with the interpolated low-resolution image to obtain a high-resolution medical image.The proposed algorithm is tested on the Cancer Imaging Archive(TCIA)dataset.After original images are enlarged with 2,3,4-scaling factors,super-resolution reconstruction of these low resolution images is performed using SRCNN,DRCN,VDSR and the proposed algorithm.The experimental results show that the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)of images reconstructed using the proposed algorithm is higher than other methods.At the same time,the subjective visual effect is also improved significantly,which can restore more image details.In addition,the parameters and training time of the network in this paper are far lower than other representative super-resolution networks.(2)Aiming at the shortcomings of traditional SCSR image super-resolution reconstruction algorithm in the restoration quality,a SR reconstruction method combining contrast weighted prior information is proposed,MCSR reconstruction method using complementary information between different modal MR images.First of all,a high-frequency information transmission network is constructed to realize the transmission of detail information from one image to another.Secondly,three versions of EDSR network are used to explore the relationship between the reconstruction quality in different situations.It is proved that the high-frequency information of image details can be transmitted in different weighted contrast images.The complementary and similarity of high-frequency information between T1 weighted images and T2 weighted images are used to improve the quality of super-resolution reconstruction.Finally,through the comparison of ablation experiments,the optimal parameters were selected to achieve a trade-off between reconstruction quality and reconstruction efficiency.(3)A lightweight dense residual connected three-dimensional convolution neural network P3 DSRNet using pseudo-3D convolution is proposed.First of all,the dense residual block of the improved network is used to widen the channel of convolution layer in the residual block,and more feature information is input into the activation function,so that the shallow image features in the network are more easily spread to the high-level,and the super-resolution expression ability of MR medical image is enhanced.Then,the Pseudo 3D separable convolution strategy is used to train the network,and the standard 3D convolution kernel is divided into multiple convolution kernels,and the training is carried out in stages.The convergence speed of network training is faster,which solves the problem that the widening dimension of standard 3D convolution leads to the increase of network training difficulty and the sharp increase of parameters.The experimental results show that compared with the traditional interpolation algorithm and LRTV algorithm,the medical image reconstructed by the P3 DSRNet super-resolution algorithm has richer texture details and better visual effects.Compared with the convolutional neural network super-resolution algorithm ReCNN,the network parameters are greatly reduced,and the PSNR and SSIM performance are also improved.(4)On the basis of P3 DSRNet,a new loss function is designed.On the basis of L1 loss function in common use,combined with the error in the intermediate layer,a better reconstruction quality is obtained,which is both robust and enriches computer-aided diagnosis.Technology provides a broader application prospect.Experiments in different application scenarios show that the super-resolution reconstruction algorithm proposed in this paper has high image reconstruction quality and good robustness,and acquire huge application prospects in computer-assisted disease diagnosis.