Research On Medical Image Segmentation Algorithm Based On Deep Residual Network

Medical image segmentation has wide application and research value in medical research,clinical diagnosis,pathological analysis,surgical planning and computer-assisted surgery.Quantitative measurement and analysis of relevant imaging indicators before and after the treatment of patients is helpful for the diagnosis and selection of treatment plans.The traditional medical image segmentation annotation task is performed by the clinician through a pixel-level manual annotation of each medical image.However,manual delineation of medical image is tedious and time-consuming.Moreover,radiologist infection annotation is a highly subjective task,often influenced by personal bias and clinical experience.Therefore,the automatic segmentation of medical images is paid more and more attention.Traditional medical image segmentation methods require human participation to obtain good feature representation.In contrast,the segmentation network based on convolutional neural network segmentation network can automatically learn the feature representation,so the CNN-based segmentation model is widely used in many segmentation tasks.However,due to the complex texture of medical images,large-scale variation and low contrast with normal tissue,fully extracting useful features and accurately segmenting the region of interest is still a challenging task.(1)A segmentation network based on multi-scale feature fusion and attention mechanism is proposed to realize the automated segmentation of COVID-19 infected areas.Because of the problems of high variability of texture,size and position of infection in COVID-19 segmentation problem,firstly,the residual network as the backbone network is used for feature extraction,and the global context aggregation strategy is used to obtain the rough segmentation results.Secondly,the Multi-scale Feature Fusion module is added to the network bottleneck,using void convolution and Multi-kernel pooling to enhance the ability of the network segmentation.Finally,a Reverse Attention module with cascading structure is designed,and the detail features of the complementary area are used to enhance the contrast between the background and target.The accuracy,specificity and sensitivity of the proposed method in the public COVID-19 CT segmentation test set reached 0.714,0.700,0.958.The missed areas were significantly reduced,and the segmentation ability of delicate lesions was improved considerably.(2)We propose a segmentation network based on edge supervision and global context aggregation further to improve the segmentation accuracy of COVID-19 infected areas.In view of the problems of low contrast between infected areas and normal tissues,an edge supervision module is added to the feature extraction part,and the generated edge supervision features are used to integrate and enhance the contrast between lesions and normal tissues and solve the problem of edge blur.In addition,the context aggregation module is introduced to aggregate advanced and underlying features and generate global information,further improving the global perception capability of the network.Experiments show that the proposed method is outperformed on the public COVID-19 CT segmentation dataset.(3)We propose a segmentation network combining Transformer and CNN for the automated segmentation task of skin diseases and polyps.Continuous down-sampling operations bring about the redundancy of networks and loss of local details,and they also have significant limitations in remote mining relationships.Instead,Transformers show great potential in global context modeling.This thesis proposes a bidirectional cascade segmentation network combining Transformer and CNN for medical images to improve the efficiency of network modeling in the global environment while maintaining control over the underlying details.Furthermore,a new feature integration technique,termed as Two-stream Cascade Feature Aggregation module,is constructed to integrate multi-level features from the two branches effectively.Moreover,a Multi-scale Expansion-Aware module is designed,which can extract high-level features containing more abundant contextual information and further improve the perceptual capability of the network.Extensive experiments on the skin dataset and polyp dataset show that the proposed network performs well and accurately calculates the segmentation of skin disease and polyp.