Research On Medical Image Fusion Model Based On Deep Learning

With the development of sensor technology,medical imaging equipment presents diversification.Different devices collect different information due to different imaging principles.In order to fully combine the advantages of various medical images to help doctors complete the diagnosis,researchers have fused medical images from different devices.The fused images can provide a more complete description of the morphological structure and metabolism of organs and help doctors complete the diagnosis and treatment plan of patients better.At present,the quality of medical image fusion directly depends on two parts: feature representation and fusion strategy.Traditional feature representation methods usually rely on new signal processing ideas,which limits the flexibility of algorithms.The selection process of fusion strategy is mixed with too much human intervention,which increases the uncertainty of the influence on the algorithm performance.With the rise of deep learning in computer vision,its powerful representation learning ability can reduce the influence of human factors on the fusion process.Deep learning medical image fusion model can simplify complex operations into module tasks,improve the quality of fusion images and enhance the scalability of the algorithm.In the article,relevant technical problems of medical image fusion model based on deep learning are studied.Four kinds of medical image fusion models based on deep learning are proposed to solve the problems of excessive noise,difficult processing of edge features,difficult description of details and difficult fusion of context features in medical image which are difficult to be solved by traditional fusion methods.The specific research contents are as follows:1.A residual dense network medical image fusion model based on intuitionistic fuzzy set is proposed.The intuitionistic fuzzy set and the feature multiplexing layer are constructed.The former can effectively reduce the noise information in medical images.The latter enables the shallow information to be directly used by the deep layer and enhances the saliency of the image.The first combination of cross entropy and structural similarity loss function effectively reduces the information loss in the process of image fusion.Experimental results show that the fused image obtained by the model can describe the color and texture information of organs completely.Compared with other algorithms,the model performs better.2.A Hahn-PCNN medical image fusion model based on convolutional neural network is proposed.The model obtains image feature maps through the depth of the convolution block with different convolution kernels and strides transformations.It selects the Hahn moment to guide the potential energy of the image block and combines with the pulse coupled neural network to realize the processing of the differential edge features,which solves the problem of difficult processing the edge features of medical images.The experimental results show that the model can better fuse the edge features of medical images from different devices,eliminate the artifacts of the fused images,reduce the loss of structure and color information and improve the quality of the fused images.3.A medical image fusion model of residual attention mechanism based on generative adversarial network is proposed.In the model,an image generator is constructed by combining the residual attention mechanism block and the concatenated detail texture block.The generator is used to describe the global dependency to improve the description of the image details.The dual-discriminator is designed to complement the color and structure information of the generated image.The experimental results show that the fused image obtained by the model can capture more details,which is helpful to assist doctors to understand the abnormal tissue conditions.The model also performs better than other comparison algorithms in both subjective and objective evaluation systems.4.An attention-multiscale network medical image fusion model based on context features is proposed.In the model,a hierarchical encoder and a cascade decoder are constructed to obtain image context features and transfer them between layers.The design of attention-multiscale network is to achieve the fusion of image context features.The experimental results show that the fused image obtained by the model can describe the important information of medical images more comprehensively.In the image quality evaluation system,the model also has outstanding performance.The main innovations are summarized as follows:(1)An intuitionistic fuzzy set preprocessing module is proposed to reduce the noise information of fused images.Combined with the designed residual dense network,the salient features of the image are obtained.The fusion strategy is constructed with the trace of the matrix as the core,which improves the quality of the fusion image.(2)The potential energy guided by the Hahn moment is selected to motivate the adaptive pulse coupled neural network,which solves the difficulties of processing edge features in medical image fusion and enriches the edge information of the fused image.(3)The residual attention mechanism module and the concatenated detail texture module are constructed to solve the problem of difficult description of details in medical images,while reducing the information loss caused by attention mechanism.The adversarial network is generated by the dual-discriminator to enhance the representation of tissue information in the fused image.(4)A hierarchical encoder and a cascade decoder are constructed to achieve the acquisition of medical image context features and inter-layer transmission of features.The attention branch and the residual multi-scale detail processing branch are designed to solve the problem of difficult fusion of medical image context features and enhance the description of organizational structure,texture details and metabolic information,so as to improve the application value of fused images.