Research On Medical Image Segmentation Algorithm Based On Vision Transformer

Medical image segmentation aims at various types of patient images to label the target organs or tissues of diagnostic research.It can effectively help doctors to make accurate diagnosis and treatment.In recent years,many methods have constructed deep learning models based on convolution for medical image segmentation.However,due to the inherent limitations of convolution operation,its receptive field is usually limited.and the dependence of global information cannot be modeled.The visual Transformer has obvious advantages in the modeling of global information.and it has been widely studied and made significant progress in the field of natural language processing.In this context,this paper introduces Transformer into the traditional deep learning model based on convolution and conducts in-depth research.The specific work and achievements are as follows:(1)In this paper.we propose a medical image segmentation method based on dense multi-scale convolution and Transformer.First.the image is encoded using multiple residual convolution and efficient Transformer.Secondly,a novel muiti-scale feature fusion scheme is used to fuse the features in different stages of coding and pass them to the decoding part.Finally,a combined weighted loss function is introduced to guide the training optimization of the model from multiple perspectives of pixel.local region and global feature map.(2)In this paper.we propose a novel feature integration network with hierarchical Transformer to achieve accurate medical image segmentation under the guidance o f explicit label and implicit boundary feature information.Firstly,a new hybrid attention module is used to extract feature information of different scales step by step.Secondly,the cross-fusion module is used to aggregate the features of adjacent layers and to supplement the semantic information of low-level features.Then.depth boundary supervision and label supervision are carried out in parallel in different decoding branches.Finally,the multi-path fusion module is used to integrate the semantic expressions of different global decoding branches to obtain the final accurate segmentation results.(3)In this paper,we construct a medical image segmentation system using the proposed method.Firstly,the proposed model is trained on three public datasets,such as abdominal multiple organs,myocardial ventricle and prostate,so that it can accurately segment each organ.Then,the trained model is embedded into a human-computer interaction interface,so that it can perform automatic segmentation operations for various types of images under artificial control.The system can achieve accurate and fast image segmentation for up to ten different human organs(such as ventricle,prostate,liver,etc.).