| In recent years,medical image semantic segmentation has become the research focus of deep learning in the field of medical imaging.The essence of medical image semantic segmentation is categorizing medical images at pixel level and labeling pixels of specific organs or lesions in medical images.The effect of medical image semantic segmentation is often affected by long-distance dependence.The long-distance dependence in medical images is the interaction between the current pixel and the pixel of other organs or lesions.Traditional methods often use manual annotation to achieve pixel-level classification,which has a complex process and low robustness.The convolutional neural network in deep learning is more suitable for feature extraction,but it lacks the understanding of long-distance dependence.Because of the Vi T has the ability to effectively encode the long-distance dependency relationship in the data,this paper studies the method based on the Vi T to obtain the global features of the medical image and improve the impact of the long-distance dependency relationship on the medical image semantic segmentation.The specific research contents consists of three parts:Research on linear semantic segmentation algorithm based on Vi T.Firstly,this paper studies the traditional methods in the field of medical image semantic segmentation.According to the basic principle of image semantic segmentation algorithm,a semantic segmentation model based on Vi T is proposed,which is similar to FCN structure.This model extracts images’ features through Vi T to obtain global semantic information,then achieving medical image semantic segmentation.Experiments on multiple data sets show that this method is not only suitable for medical image semantic segmentation directly,but has advantages in acquiring and processing global semantic information.Research on semantic segmentation algorithm of U-shaped structure based on Vi T.By studying the common algorithms in medical image semantic segmentation algorithms,this paper applies the Vi T to a U-shaped network structure,which is more suitable for medical image semantic segmentation.The U-shaped structure makes full use of the shallow and deep features of the image,improving the accuracy of image semantic segmentation through the method of skip layer connection.Through experiments on several medical image data sets,it is proved that the model proposed in this chapter that applies Vi T to U-shaped structure can achieve good performance.Research on semantic segmentation algorithm of Vi T under multi-modality and multiview.By studying the features of multi-modality and multi-view images in medical images,this paper proposes a multi-modality fusion converter module and a multi-view fusion converter module for feature extraction and feature fusion of images in different modes and different angles of view in order to make full use of image features in different modes and different angles of view.Through experiments on multi-modality and multiview fundus image data sets,the feasibility of feature extraction and fusion of the two modules is proved.Combining the two structures proposed in the previous chapter with the model in this chapter proves that using the features of different modes and different angles of view can improve the semantic segmentation accuracy of the model. |
