Research On Medical Image Fusion Algorithm Based On Neural Network And Dictionary Learning

Medical image fusion is a process of the different modal medical images are fused by source image decomposition and sub-image fusion.In today’s medical environment,medical image fusion has been widely used in clinical treatments,but the existing medical image fusion algorithms are difficult to extract structural information and functional information from the input anatomical images and functional images.In order to overcome this defect,three novel medical image fusion algorithms are proposed:(1)A fusion algorithm based on convolutional neural network(CNN)and weighted least squares method(WLS)is designed.Firstly,the CT and MRI images are decomposed into a basic layer and a series of detail layers by a hybrid multi-scale decomposition tool composed of Rolling guided filtering and Gaussian filtering,which can better retain the scale information and edge information of the source image.Secondly,the convolutional neural network is used to extract and fuse the basic layer,and the maximum absolute value rule and the weighted least squares optimization strategy are combined to fuse the detail layer,so as to obtain more perfect image detail information and have higher contrast.The simulation results show that the proposed algorithm has good fusion effect,and the fusion effect is more prominent in acute stroke,fatal stroke and meningioma.(2)A fusion algorithm based on structural tensor and Kirsch operator is designed.Firstly,the gray-level source medical images are decomposed into intensity layer and non-intensity layer by using structural tensor.Secondly,the intensity layer is fused by using the fuzzy-adaptive reduced pulse-coupled neural network(FARPCNN)model,and the non-intensity layer is fused by combining the Kirsch edge detection operator and the novel Sum-Modified-Laplacian(NSML).The fusion results of the proposed algorithm overcome the problem of color information loss in the functional image,and can extract the local fine geometric details in the source image as much as possible,including more prominent features and edge contour information.Compared with other comparative algorithms,the proposed algorithm has better advantages in visual evaluation and objective evaluation.(3)A fusion algorithm based on dictionary learning and Retinex enhancement method is proposed.Considering that the combination of low-rank dictionary and sparse dictionary can remove the noise and retain the texture information of the source images,dictionary learning is used as the decomposition tool of medical images,and the MRI image and PET(SPECT)image are decomposed into sparse components and low-rank components.The low-rank component and sparse component after Retinex enhancement are fused by using the pixel average rule.The simulation results show that the proposed algorithm has the stronger detail retention ability and the clearer contrast,which is more conducive to the clinical observation.