Research On Key Technologies Of Human Soft Tissue MR Image Matching

With the progress of computer and artificial intelligence technology,virtual surgery has achieved unprecedented development,in which the accuracy of human soft tissue deformation model is one of the decisive factors for the quality of virtual surgery.In order to improve the accuracy of the human soft tissue deformation model,it is necessary to accurately measure the parameters of human soft tissue materials in vivo.At present,the commonly used measurement methods in clinic can be divided into contact type and non-contact type.Contact instruments,such as skin elasticity tester,will interfere with the soft tissue in the natural state,and the collected parameters will have a large deviation,while non-contact methods all benefit from the development of medical imaging technology,such as ultrasound,MRI and so on.Because of the advantages of low radiation dose of MRI,clear imaging of soft tissue and not hindered by gas and bone,by matching two groups of images of human soft tissue recorded by MR image before and after deformation,the accurate displacement of pixels in the pre-deformation image on the deformed image can be calculated,and the deformation field can be measured,and then the deformation parameters of soft tissue can be obtained.However,there are three difficulties in matching feature points in soft tissue before and after deformation through MR image: 1)Few standard samples;2)Artifacts in MR image;3)Low accuracy of matching a large number of feature points in nonlinear soft tissue.In order to solve the above difficulties,the main work of this paper is as follows:（1）A method of MR medical image generation based on GAN is proposed.The main innovation of this method is that the MR image is generated by improving the generator and discriminator,and the structure information of multi-scale edge contours is integrated into the generator in order to better capture the edge details of MR images.At the same time,the multi-scale information discriminator based on Transformer is used to distinguish the generated image from the real MR image.The experimental results show that the improved generator and discriminator can effectively generate MR images similar to real images and provide an effective way to solve the problem of lack of medical image data.（2）A super-resolution image generation model which can be guided by reconstructed image network is designed.The main innovation of this model lies in the rational use of multi-task learning mode,the use of reconstruction network to guide super-resolution network for image generation,so that effective information between different tasks can be effectively used.In the part of reconstruction network,the inter-pixel position attention module is introduced into the output of each layer of the decoder,which makes the structure of the reconstructed image clear.At the same time,the task conversion module based on Transformer is used to generate super-resolution image.The experimental results show that,compared with the single-task super-resolution network,the image definition and detail accuracy generated by this method is higher.（3）An enhanced feature descriptor model is improved.The model introduces the graph neural network into the feature description,combines the idea of attention mechanism and the structure of the graph neural network,updates the node feature information by means of message transmission,and finally obtains the enhanced feature descriptor.Different from the traditional feature description method,this method pays attention not only to the local feature point information,but also to the target image feature point information.The experimental results show that this method can effectively improve the accuracy of feature point matching in two groups of MR images before and after human soft tissue deformation.