Research On Segmentation Algorithm Of Eyeball Ultrasound Image Based On Semantic Embedding And Attention Mechanism

At present,with the rapid development of deep learning and medical equipment,all kinds of medical image data and scale present a spurt of growth,in the face of the consumption of the massive image data processing,which need a lot of manpower material resources,in order to save resources of artificial processing,research scholars for different kinds of images into deep learning approach to deal with huge amounts of medical images.With the growing maturity of deep learning,semantic segmentation of ultrasound images has become one of the important steps in the field of machine vision for medical device assisted diagnosis.However,ultrasound image has disadvantages such as noise interference,low contrast,blurred edges,which lead to inaccurate segmentation results of eyeball ultrasound image.In order to achieve more accurate segmentation,using the deep learning method to represent in the ultrasound images of eyeball form and learning,classification and projections for each pixel,so as to more accurately segment the eye region and background region,different from traditional image segmentation algorithm,based on deep learning methods of the semantic segmentation algorithm on the segmentation accuracy and speed are greatly improved.In this paper,based on image processing and deep learning,the semantic segmentation algorithm of eyeball ultrasound images is studied.The research content involves deformable convolution,multi-layer feature fusion with semantic embedding,and semantic segmentation algorithm of attention mechanism.The specific work is as follows:(1)Unet segmentation algorithm based on deformable convolution and semantic embedding branch.First of all,eyeball ultrasound images have disadvantages such as noise interference,regional blurring and gray similarity.Therefore,the encoderdecoder network architecture based on Unet uses deformable convolution to replace the traditional convolution operation to better adapt to the geometric shape of the eyeball region and improve the feature learning ability of the convolutional neural network.In view of the problem of semantic difference between several layers caused by the use of skip connection in segmentation network,the eyeball region in ultrasound image cannot be segmented accurately.Therefore,semantic embedding branch is introduced to fuse the feature information between different layers to provide more aeccurate features for the decoder feature recovery.Secondly,the data set of eyeball ultrasound images is randomly divided into training set and test set,and the appropriate parameters of the convolutional neural network are selected for multiple network training and predicting.Finally,compared with the popular segmentation network,the prediction results are analyzed to further verify the effectiveness and robustness of the proposed algorithm.(2)Semantic embedding segmentation algorithm of attention mechanism.Firstly,based on attention mechanism constructe the segmentation network of eyeball ultrasound image,in the feature map extraction stage,the semantic embedding branch is introduced into to realize the fusion between different layer characteristics,further enrich the feature map of the detail information and semantic information,through the attention mechanism highlight significant features of eyeball ultrasound images,which improve the network’s understanding of the eye area characteristics and expression and reduce the false segmentation of eyeball ultrasound image;Secondly,the segmentation network was trained several times and ablation experiment was carried out on the training set of eyeball ultrasound images to optimize the segmentation capacity of the constructed module.Finally,in order to verify the effectiveness of the proposed algorithm in this paper,comparative experiments were carried out on the test set of eyeball ultrasound images,and qualitative comparisons were made with the several popular segmentation methods.