Research On Medical Image Fusion Algorithm Based On Neural Network

As an important basis for medical diagnosis,it is very important to ensure the richness and clarity of image information.However,the monomorphic medical image has a single presentation content,while the fusion image contains multi-modal image information,which is feasible in clinical practice and provides a more reliable basis for medical diagnosis.With the improvement of medical diagnosis,the fusion effect of medical image conforms to human visual perception,reduces the influence of prior knowledge on the fusion effect and enhances the expression of details.In view of the existence of artifacts in traditional medical image fusion and the need to manually set fusion rules and parameters by relying on prior knowledge,resulting in uncertain fusion effect and insufficient detailed expression,this paper studied and improved the following aspects:(1)A method of medical image fusion based on Laplacian pyramid and convolutional neural network is proposed.Firstly,the Laplace pyramid of the input region of the source image is decomposed.To improve the convolutional neural network,then the evaluation index based on empirical risk minimization replacement based on structural risk minimization,through standard convolution of step length is 2 layer for dimension reduction,with the improved through constant iterative convolution neural network,to generate the optimal weight figure to guide the fusion process,determine the optimal number of iterations through the experiment.Finally,the fusion image is generated by the inverse process of regional Laplacian pyramid.The simulation results show that the proposed algorithm not only achieves a good improvement in parameter adaptation,but also achieves a good fusion effect in subjective vision and objective evaluation index.(2)A method of medical image fusion based on improved generated against network is proposed,first of all,the two parts of generator and the discriminant network structure wasimproved,in the design of the generator in the network by using residual block and quick connection network structure,deepen the deep image information better capture,remove the sampling layer under the conventional network at the same time,to reduce the information loss in the process of image transmission,and change the batch normalized to normalized,to achieve the purpose of better keep the information of source image;Increase the depth of discriminator network to improve network performance.Then,the CT image and MR image are connected and input into the generator network to obtain the fusion image.The network parameters are continuously optimized through the loss function to train the most suitable model for medical image fusion and generate high-quality images.The simulation results show that the algorithm performs well in mutual information,information entropy and structure similarity,and the final fusion image is rich in texture and detail.At the same time,it avoids the influence of human factors on the stability of fusion effect,and there is no significant difference in running time,so it achieves effective and reliable practical application.(3)The medical image fusion system is constructed to make the image fusion algorithm feasible and operable.The system mainly consists of three modules: comparison of different reconstruction methods,comparison of LLP-CNN fusion method,and comparison of Res-GAN fusion method.The four excellent image decomposition and reconstruction algorithms are presented in the comparison module of different reconstruction methods.In the LLP-CNN fusion method comparison module,the medical image fusion algorithm based on Laplacian pyramid and convolutional neural network,the fusion effect diagram of its comparison algorithm and the calculation value of some evaluation indexes are presented.In the comparison module based on Res-GAN fusion method,the image fusion algorithm based on residual block and improved generated antagonistic network,the fusion effect diagram of its comparison algorithm and the calculation value of some evaluation indexes are presented.