Research On Medical Image Super-resolution Algorithm Based On Convolutional Neural Network

With the rapid development of information technology,we are now in a digital age.Digital image has become one of the main ways that we access information from the outside world.Super-resolution reconstruction technology is to process low-resolution images,which are acquired by existing digital image equipment,by computer and then convert them into corresponding high-resolution images.Medical image is an important basis for doctors to diagnose diseases,so that it is significant to improve the resolution of medical images for doctors to find the disease as soon as possible and prepare the correct treatment plan.Therefore,this research combines convolutional neural network and super-resolution reconstruction technology to further advance the quality of medical images.The high-resolution reconstructed medical images will have higher pixel density,better visual effects and more texture information.This paper proposes the algorithm can resist the noise in the original image during processing super-resolution reconstruction,and at the same time being faster,more efficient and robust.This paper mainly focuses on the process of super-resolution reconstruction,aiming at the medical image super-resolution based on convolutional neural network algorithm,in order to improve the quality of reconstructed medical images while accelerating the speed of the algorithm and advancing the robustness of the algorithm,thus provide better help for doctors.In this paper,the research is mainly carried out from the following two aspects:(1)It proposes a medical image super-resolution algorithm that combines local features and global features.The algorithm enables more detailed feature extraction of medical images through the secondary feature extraction layer.Then it uses overlapping pooling layers in order to reduce the extracted feature dimension and highlight the important features,which can also simplify network structure and improve training efficiency.Furthermore,a link layer is established in the algorithm to complete a reconstruction using local features on the connection layer,which not only can compensate for the loss of convolution operations of each layer,but also can effectively prevent the final images containing false information.The experimental results show that the quality of the image processed by our algorithm is higher than other traditional super-resolution algorithms and other deep learning network models.There are great improvements in whether objective evaluation index PSNR or the subjective visual effect and detail texture.(2)Due to the particularity of medical imaging equipment,there are two major differences between medical images and normal natural images.First,the medical image is susceptible to noise contamination in the collection and transmission process,and the noise in medical images will be amplified after super-resolution reconstruction.Second,all medical images basically have large black border with no texture information.Combining these two features of medical images,this paper proposes a medical image super-resolution algorithm based on noise resistance convolutional neural network.It uses discrete Harr wavelet transform as preprocessing algorithm in order to utilize the denoising properties of the wavelet transform itself to resist the noise in the original image.Then it uses adaptive partition algorithm based on image content to segment the original image,direct output and merge those black borders which have no texture information,which can reduce the time complexity of the algorithm.The experimental results show that our algorithm demonstrates better performance in medical images with Gaussian white noise and salt-and-pepper noise contamination.There are great improvements whether in the objective evaluation index PSNR or the subjective visual effect and detail texture.Our algorithm successfully addresses the phenomenon that noise in medical images will be amplified after super-resolution reconstruction.Also,the training time of the network model is greatly reduced,and the reconstruction task can be completed swiftly.