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

Medical images reflect the internal structure of the human body and are the basis for modern medical examinations.Medical image segmentation is an important step in the field of biomedical imaging,and its main feature is to divide the image into several regions based on the similarity or difference between regions.Because conventional machine learning techniques rely on feature engineering for image segmentation,the expressiveness of the obtained features is limited,so the segmentation results are quite restricted.In recent years,medical image segmentation methods based on deep learning have received much attention,among which methods based on convolutional neural networks have excellent feature recognition and extraction capabilities,and their performance in the field of medical image segmentation far exceeds that of traditional machine learning methods.This article focuses on medical image segmentation and points out that the existing medical image segmentation methods in research still have performance indicators such as accuracy and stability that are not high enough,and the combination with clinical application is seriously insufficient,especially CT brain hemorrhage images and pituitary MR sequences.This article aims to achieve high-quality medical image segmentation by using deep learning-based medical image segmentation algorithms as the technical core and focusing on the demand for medical diagnosis and assistance.It conducts in-depth research and proposes some effective theories and methods.The main research contents of this article are as follows:(1)In view of the problems that the deep learning CT image intracerebral hemorrhage segmentation method has single data set parameters and low generalization,a study on the measurement of intracerebral hemorrhage with mixed precision training based on deep learning is carried out,and a coarse-to-fine deep learning strategy is proposed,,which combines the advantages of 2D and 3D models to segment intracerebral hemorrhage in three stages: skull dissection,intracerebral hemorrhage location and intracerebral hemorrhage fine segmentation.The three stages share the same convolution neural network(called s ICHNet)for the segmentation of intracerebral hemorrhage.Using the proposed model and strategy,training and verification are carried out on private data set and public data set(CQ500)with various layer thickness parameters.The experimental results have proven that the proposed model exhibits excellent generalization and robustness,and its performance is better than that of a single2 D or 3D segmentation model.Finally,the visualization of hematoma follow-up volume and the changes of hematoma volume in different treatment methods of intracerebral hemorrhage were explored,so as to guide clinical decision more accurately.(2)In order to solve the problem that it is difficult to segment small targets such as adenohypophysis,the research of adenohypophysis segmentation based on MR images is carried out based on deep learning and radiomics.A strategy of adenohypophysis localization and adenohypophysis fine segmentation is proposed,and a fusion model of Transformer and CNN(PIT-Former)which can capture local details and long-distance dependence is designed for fine adenohypophysis segmentation.The experimental results prove the advanced performance of PIT-Former in the task of adenohypophysis segmentation.In addition,we also compared the adenohypophysial image radiomics features of automatic segmentation with those of manual segmentation,and the experiment proved that the radiomics features of the above two methods are highly consistent.Finally,adenohypophysial radiomics features of automatic segmentation were used to make regression prediction of hormone levels,and the potential of imaging features to replace hormone levels was preliminarily explored.