Medical Image Fusion Based On Deep Learning

Medical image is an indispensable tool in medical diagnosis and treatment,and an important means to study and explore the internal structure and pathological state of the human body.It is widely used in medical diagnosis,treatment and other research fields.However,the content of single-mode medical image is simple and contains little information.Therefore,the image fusion technology is adopted to fuse the multi-mode medical image into an image containing multiple information,so as to maximize the pathological information of patients and effectively make up for the deficiency of single image information.With the development of artificial intelligence,the research on intelligent medicine is becoming more and more popular.Medical image fusion technology based on deep learning can extract deep image information that cannot be obtained by traditional method.Moreover,the whole fusion process of traditional method requires manual design of fusion rules,which has great limitations.The fusion image has some problems,such as low accuracy,insufficient inheritance information and artifacts.To solve these problems,the following researches are carried out in this paper:(1)A convolutional neural network image fusion algorithm based on attention mechanism is proposed.Firstly,the attention mechanism is introduced into the convolutional neural network,so that the neural network can pay attention to the correlation between channels while extracting image features,and improve the accuracy of the weight graph.After that,Laplacian pyramid is used for multi-scale image decomposition,which can effectively reduce the problem of generating artifacts from fused images.Finally,under the guidance of the weight graph W obtained by the neural network,each layer is fused respectively,and the image is reconstructed by the Laplace Pyramid.The obtained image has richer information of the source image,the fused image is of higher quality,and the fusion effect is better than other medical image fusion algorithms based on convolutional neural network.The four evaluation indexes of SSIM,MI,SF and AG were improved slightly.In addition,in view of the shortage of medical fusion image data,transfer learning is introduced to accelerate the convergence of model training and achieve better results.(2)This paper presents an improved medical image fusion algorithm for generating adversarial residual networks.Firstly,the original discriminator is added into two discriminators.The target image input by the original discriminator is obtained from other existing medical image fusion algorithms,so the quality of the original algorithm is strongly dependent on the quality of the algorithm to obtain the target image,which has great limitations,and the target image used by the improved double discriminator network is the real source image.Then,structural similarity loss and content loss are added into the loss function of model training.The algorithm has significant signs in the two evaluation indexes of SSIM and MI,indicating that the fusion image of the algorithm has a good performance in inheritance,improving the quality of the fusion image,but also solving the problems of detail loss and slow fusion speed of the current fusion method.