The Research On Medical Image Segmentation Algorithm Based On Convolutional Neural Network And Transformer

Medical imaging is an image that reflects the internal structure of the human body and is one of the main cornerstones of modern medical diagnosis.In recent years,Convolu-tional Neural Networks(CNNs)have rapidly developed and been widely used in the field of medical image segmentation,with better performance compared to traditional machine learning methods.However,CNNs face the problem of capturing long-range dependencies.With the short information path of the Transformer,the visual Transformer has alleviated the limitations of CNNs to some extent.However,due to the various lesion morphologies,complex boundary structures,and the similarity between the lesions and the surrounding environment in different medical images,relying solely on the visual Transformer still can-not achieve accurate localization and segmentation of lesions.This thesis focuses on the research of medical image segmentation methods based on both CNN and Transformer and proposes two high-quality medical image segmentation algorithms to address the challenges in medical image segmentation.The main research contents are as follows:(1)In this thesis,we propose a Dual Branch Cascade Graph Network(DBCGN)to ad-dress the issues of irregular lesion shapes,blurred boundaries,and potential interference from a large amount of redundant information in medical images.During the feature extraction stage,our method utilizes a Transformer branch to establish cross-scale global context de-pendencies on the bottom three layers of feature maps from the convolutional neural network branch and integrates local and global contextual information.Moreover,to tackle the prob-lem of fuzzy and complex image boundaries,we propose a reverse graph inference method that refines the boundaries of segmentation predictions using rich features such as contours,edges,and textures in low-level features.We evaluate the effectiveness of our proposed DBCGN method on three skin lesion image datasets:ISIC2016,ISIC2017,and PH~2.The extensive experimental results demonstrate that DBCGN can effectively solve the problem of variable and diverse boundary structures of target lesions with good performance.(2)The existing medical image segmentation methods are mainly designed to segment lesions in a specific type of medical image and have poor generalization ability to solve other medical image segmentation problems.To address this issue,this thesis proposes a univer-sal dual-branch context-aware Transformer network(DPCTN)method to solve the common problems in 2D medical images.This method effectively captures local details and global contextual information and integrates them.Specifically,the method extracts the global con-text of the boundary region in the dual branch as guidance for refining and enriching the target boundary and region layer by layer.In addition,due to the problem of dimension collapse in the attention mechanism,a novel three-branch transposed self-attention method is proposed in this chapter,which calculates the attention across dimensions by exchanging dimensions,avoiding information loss during the pooling process.To verify the effective-ness and generalization ability of the proposed method,extensive experiments are conducted on three types of datasets including optical medical images,pathological medical images,and acoustic medical images,including colonoscopy polyp images(CVC-Clinic DB,Kvasir-SEG,CVC-Colon DB,ETIS,and CVC-300),skin lesion images(ISIC 2016 and ISIC 2017),gland segmentation(Glas)and breast tumor ultrasound images(BUSI)datasets.The experi-mental results show that the proposed method has good model performance and generaliza-tion ability.