Research On Generative Adversarial Network Based Medical Image Enhancement Algorithm

With the continuous development of medical imaging technology,medical images of different models are widely used in clinical disease diagnosis,assist the operation and health detection.Due to the limitations of imaging technology,the collected medical images often have problems of single modility and low image resolution,which is not conducive to the clinical application of subsequent medical images.Medical image enhancement technology can effectively improve image quality,including medical image fusion,medical image super-resolution,medical image denosing and so on.Medical image enhancement based on convolutional neural network is a current research hotspot and has achieved great performance improvement.Most of the enhancement results of convolutional neural networks are obtained by supervised learning in datasets containing reference images,but it is difficult to obtain reference images in the field of medical image enhancement.Generative adversarial network(GAN),as a convolutional neural network that can by means of unsupervised learning,can effectively solve the problem of unreferenced images.Therefore,this paper focuses on the research of medical image enhancement algorithm based on generative adversarial network,which has important practical medical application value.The main research contents and innovations of this paper are as follows:1)A medical image fusion algorithm based on semi-supervised learning and GAN is proposed.Existing deep learning-based methods mainly rely on a large numbers of data sets containing labels(reference images),but it is difficult to obtain reference images for medical image fusion tasks.In order to make full use of scarce label data sets,this paper adopts multi-stage training to realize semi-supervised learning,and proposes a medical image fusion framework based on semi-supervised learning and GAN.Experimental results of medical image fusion in different modes show that the proposed framework can better complete the fusion of effective information in various models of medical image.2)A super resolution model of medical image based on SRGAN is proposed.SRGAN network only the identity mapping from shallow features to deep features is considered,but the improvement of network performance by attention mechanism is not well considered.Therefore,this paper improves the network structure of SRGAN and replaces the residual module in SRGAN with Spilt Attention in Attention Block(SAAB)module.Using WGAN Loss to train the network can effectively solve the problems of model collapse and gradient explosion in SRGAN.Experimental results of medical image super-resolution in different models show that the proposed framework can accomplish the task of medical image superresolution well.3)A multi-task medical image enhancement model based on generative adversarial network is proposed.In order to obtain high resolution medical fusion images,existing technologies often adopt image super-resolution and image fusion separately,which results in high computational cost.In this paper,the feature parameters extracted from the super-resolution network branch are input into the fusion network branch by parameter extraction,and the image fusion and image super-resolution tasks are integrated into one frame.Experimental results of medical image fusion in different models show that the proposed framework can improve the resolution of the fused image while completing the fusion of effective information in different models.