Research On Medical Image Segmentation Method Based On Dual Attention Mechanism

The segmentation of anatomy and histopathology is the primary basis for clinical medical diagnosis.Medical image segmentation is the most complex and critical task in the field of medical image processing and analysis.Its purpose is to separate the regions of special significance from the images with large noise and inconspicuous texture of the lesion area,and extract the features of relevant regions efficiently However,due to the high complexity of medical images itself,which is characterized by the lack of simple linear transformation,large noise transformation,and uncertain pixel gray classification.As a result,the task of edge texture segmentation and fine-grained feature extraction is limited.In recent years,the network model based on deep learning technology has far surpassed the traditional image segmentation technology and has excellent performance on various image tasks.However,a large number of image segmentation models only rely on expanding the network depth for feature extraction,which makes the architecture of the model complex.Model training usually needs a lot of iterations and lower generalization ability than otherTo solve these problems,we propose a composite network model CA-Net(Composite Attention Network)architecture based on dual attention mechanism.It contains a 3D backbone network encoder for feature extraction,a 3D decoder added to the channel attention block to generate semantic consistent pixel regions in the spatial range,and we also introduce a boundary attention network to better integrate the edge information of low-dimensional feature maps.At the same time,we add dense connections between the three modules.On the one hand,in order to prevent the loss of important features,on the other hand,to better integrate features of multiple scales This paper is mainly aimed at the 3D medical image segmentation task,and improves the segmentation accuracy by composite attention mechanisms.The main research contents are as follows(1)Summary of medical image characteristics and the background of medical image segmentation.Firstly,we describe the development of image segmentation methods in recent years,and introduce the advantages and disadvantages between of traditional segmentation and deep learning-based segmentation techniques(2)In view of the complexity of medical image,we propose the image preprocessing techniques and image enhancement methods for 3D medical images.The image enhancement methods based on traditional data enhancement and generation countermeasure network to explore the influence of the two methods on the experimental results(3)In order to solve the problem that 2D models not suitable for 3D medical images,we propose an efficient medical image segmentation architecture.The model architecture consists of three parts.The first part is a 3D encoder with fewer learning parameters to extract the features of the image.In the second part,a 3D decoder with channel attention mechanism is designed to generate features of spatial semantic consistency.In the third part,a spatial edge attention mechanism network is proposed,which takes the initial stage features of the encoder as the input to better extract the edge features of the image(4)In order to prevent the loss of feature information in the process of network information transmission,we use dense connections between the three modules to improve the segmentation effect of the network.(5)Finally,this paper conducts relevant experiments on three different types of medical images to prove the effectiveness and generalization ability of the proposed model.