Research On Super-resolution Reconstruction Algorithm Of CT Image Based On Generative Adversarial Network

Due to the rapid development of X-ray computed tomography(CT),it has been widely used in clinical diagnosis and treatment.High Resolution CT(HRCT)images provide more information for clinical diagnosis,using CT scanning equipment directly to obtain HRCT images(512 * 512),will cause some damage to the inspector.In order to explore a feasible solution to further reduce the radiation dose,we can try to acquire low-resolution CT(LRCT)images with less radiation dose,and then use super-resolution technology to recover HRCT images in this paper.In recent years,the generative adversarial network(GAN)has achieved good results in the reconstruction of natural images.However,there are still many problems that using the existing GAN directly execute super-resolution reconstruction of CT images: feature loss,network redundancy and memory consumption.To this end,this paper improves the network structure of GANs and optimizes the objective function of the network,to improve the reconstruction effect of GAN on the LRCT image.The main work of this paper includes:(1)Aiming at CT images are grayscale images with low contrast and fewer features,a new lightweight multiple dense residual blocks(MDRBs)structure is proposed.The MDRBs structure establishes dense links among all the residual blocks of the generator to take advantage of each layer features of CT images and ensure maximum information transmission.In addition,In order to overcome the gradient dispersion problem of GAN network during training,this paper introduces the Wasserstein distance used in CT de-noising algorithm to measure the distance between the reconstructed super-resolution CT(SRCT)image and real reference HRCT image.Experiments show that the proposed method is superior to other popular reconstruction methods in terms of convergence speed,network stability and reconstruction effect.(2)In order to reconstruct different low-resolution CT images into 512*512high-resolution images with different reconstruction factors,this paper designsthe global generator and multiple local generators to work together,and uses conditional constraints to improve network reconstruction capabilities.In addition,a multi-scale discriminator is designed to enhance the discriminative performance of the discriminator.The experimental results show that the proposed algorithm has better reconstruction performance under different reconstruction factors(4,6,8)than other algorithms.