Research On Medical Image Super-resolution Reconstruction Method Based On Generative Adversarial Network

With the large increase in medical images such as MR,CT,PET,etc.,the use of super-resolution technology to reconstruct high-definition images has become a research hotspot for the interpretation of readings.At present,the method represented by deep learning has a significant advantage over the traditional method and has become the focus of attention.However,the proposed deep learning-based method mainly presents two problems: First,the simple stacking of convolutional neural network layers cannot make the number of layers deeper,and it is difficult to extract deeper information of the image.However,the deep network is prone to the abnormality of the checkerboard block effect due to too many layers;the second is that the image reconstructed by the algorithm often has problems such as loss of texture and poor visual perception.In view of the challenges faced by super-resolution based on deep learning methods,this paper mainly conducts in-depth research from the following points:(1)The problem of image feature information extraction: This paper studies the construction of deep neural network to extract feature textures with higher levels of image,and proposes a super-resolution algorithm based on deep residual network.The algorithm uses the residual block and skip connection to design the network structure,using 32 residual blocks,making the network up to 133 layers,extracting the most detailed details of the image.And the effectiveness of the network is demonstrated from theoretical analysis and experimental comparison.(2)For the phenomenon of easy checkerboard artifacts: This paper analyzes the uneven overlap caused by the deconvolution operation commonly used in upsampling,which is theroot cause of the checkerboard effect,and is especially obvious in the deep neural network model.Therefore,this paper studies the idea of using resize-convolution.When upsampling and processing medical images,the method of scaling image reconvolution is used to minimize the occurrence of the chessboard phenomenon,so that the image quality visual experience generated by the algorithm is better.(3)Aiming at the problem of poor visual experience of reconstructed images: This paper studies the design of algorithms based on generative adversarial network framework,and proposes a medical image super-resolution algorithm based on deep residual generative adversarial network.First,the algorithm generates a network against the network,discriminates the continuous game between the two networks,optimizes the network itself,and causes the network to generate more realistic high-resolution images.Second,the generator and discriminator components of the algorithm are constructed using a neural network.The generator uses a design of a depth residual network of 16 residual blocks,and the discriminator is built based on a convolutional network.Third,the loss function of the algorithm is carefully designed,including the design of the loss and discriminant loss function.The loss is generated by minimizing the mean square error and cross entropy between images,and the loss is designed by minimizing the cross entropy between images.Through the continuous game of generating losses and discriminating losses,optimization enables the algorithm to generate medical images with higher quality and visual sense.Fourth,the algorithm effectively improves network performance by further increasing the number of network feature map channels.Fifth,the algorithm removes the batch normalization layer in the commonly used residual network to construct a new residual structure and simplify the residual block structure to optimize the network.(4)This paper designs and develops the medical image super-resolution system.Combined with the current popular web technology,each functional module in the system is designed and implemented.The system is embedded with common algorithms forimplementation and comparison,including common functions such as image preservation,uploading,and super-resolution,which provide useful help for doctors' diagnosis and practical teaching and research.Based on the above,with the representative methods of super-resolution and objective and subjective comprehensive evaluation,the two algorithms proposed in this paper have certain advantages in super-resolution image quality,and can be used to play its due role in practical work and research.