Medical Image Super-resolution And Image Dataset Augmentation Based On Cascaded GAN Network

Constructing a medical image-assisted diagnosis system based on deep neural networks can improve the objectivity and accuracy of clinical diagnosis,and has become a research hotspot in academia.Deep neural network training of deep learning algorithms often requires large samples and high-quality data sets,and existing medical image sets generally have problems with small samples,unclear textures,and inconsistent sizes.Therefore,super-resolution reconstruction of medical images is performed The expansion of medical image data sets has important scientific value and clinical significance.In order to better solve the above problems,this article conducted a more in-depth study.The main work and results are as follows:First,for the problem of unclear medical images and different sizes,a Laplacian pyramid cascaded generation confrontation network(Dense-LAPGAN)based on dense residual blocks is proposed to super-reconstruct medical images.The network increases the effective depth of the network by multiplexing the network feature matrix and network weight parameters,thereby reducing network training time and resource consumption while improving network performance.At the same time,by adding dense residual block structures to all levels of the network,the performance of the network has been further improved,and the accuracy of the super-resolution reconstruction results of the network has been improved.Simulation experiment results show that the network proposed in this paper reconstructs the self-built data set and the Deep Lesion public data set after super-resolution image peak signal-to-noise ratio reaches 32.45 d B at 4X magnification and 25.12 d B at 8X magnification.Its super-resolution accuracy is about 3% higher than that of the more popular super-resolution method.Secondly,for the small sample problem of medical images,a Laplacian Pyramid Residual Adversarial Generation Network with transition mechanism is proposed.This network adds a "transition mechanism" between networks on the basis of Dense-LAPGAN,thus A better learning ability is obtained,and in the actual training process,the network can converge more quickly and it is not easy to cause mode collapse or generate large-scale artifacts.At the same time,on this basis,a generated reference is added to the network to make the generated medical image more in line with the anatomical structure of the patient,thereby increasing the authenticity of the augmented data set.In the simulation comparison experiment conducted in this paper,the proposed network augmented data set can make the segmentation accuracy of the same semi-supervised segmentation network increase by nearly 5% compared with the affine transformation augmentation data set,compared with the self-growth adversarial generation network(PG-GAN)The augmented data set rose by nearly 1%.At the same time,compared with the PGGAN network,the data set of the network augmentation proposed in this paper scores higher on the score creation index,reaching 7.71,which is an increase of about 6.9% compared with other methods.In order to verify the effectiveness of the medical image data set augmented by this method,this paper uses three different unsupervised semantic segmentation algorithms to segment the three images of the liver,breast,and brain.Simulation experiments show that under the same experimental conditions,the average intersection of the three segmentation algorithms on the data set augmented by the method in this paper is about 1.1 to 1.5 higher than the data set augmented by the affine transformation.Therefore,the image data set augmentation algorithm proposed in this paper can definitely improve the performance of the network.Finally,a simulation system is constructed,which realizes the functions of reading and writing images,selecting image processing methods,setting network parameters,and training and storing network models.It basically meets the needs of the actual use of the network,allowing users to perform relevant operations on the network without the command line.At the same time,the function of displaying the intermediate results of training is added to make the training process visible to a certain extent,which further improves the usability of the system.