Research On Medical Image Segmentation Based On Multi-level Feature Fusion

With the increasing attention of the public to medical health and the development of digital imaging technology,digital medical imaging has become a critical reference data for medical experts to diagnose the degree of illness.In order to better analyze the patient's condition qualitatively and quantitatively,it is essential to extract pathological features from medical images.Because image segmentation technology can observe and process image features from the pixel level,and can better grasp the shape,texture and other characteristic information of the image,medical image segmentation has gradually become an auxiliary means for experts in diagnosis and treatment.However,medical images often have a large number of instance objects,blurred boundaries between target objects,low contrast,and a large number of similar noise structures around the target objects.These problems make it difficult and challenging to achieve accurate and efficient segmentation of medical images.This thesis focuses on the problems and challenges faced by medical image segmentation,and conducts research based on deep learning and multi-level feature fusion.The research content includes: the segmentation algorithm based on a symmetric network and deep bottleneck structures;the segmentation algorithm based on multi-level feature fusion,group convolution and feature channel;and the segmentation optimization algorithm based on dense conditional random field.The specific main content is as follows:(1)In view of the poor generalization ability and low accuracy of traditional medical image segmentation methods,we propose a segmentation algorithm based on a symmetric fully convolutional residual network.Specifically,based on the U-Net neural network,three novel deep bottleneck structures are proposed to replace each convolution layer in the U-Net network,which can not only increase the number of layers in the network,strengthens the extraction of features,but also reduces the amount of network parameters,speed up the network training.At the same time,the jump connection structure is used to promote the integration of multi-level features to enhance the reuse and spread of features.And through experiments,the algorithm can accurately segment the CBCT tooth image.(2)Aiming at the problem that there are many similar noise structures in medical images and the low contrast makes it difficult to accurately segment,we propose a multilevel feature fusion algorithm based on grouped convolution and feature channels.The algorithm uses two different multi-level feature fusion methods,namely point-by-point addition of feature pixels and feature channel concat.Fully combine the contextual feature information of each level of the image,use grouped convolution to combine the features of different subspaces,and use the different intensities of each feature channel to calibrate and filter the feature information to maximize the ability to extract feature information of the target object.And through experiments,the algorithm can segment CBCT tooth images and gland cell images more accurately and efficiently.(3)Aiming at the problem of blurred contours and unclear boundaries of the target object in the segmentation results of general medical image segmentation methods,a segmentation optimization algorithm based on dense conditional random field is proposed.We apply the dense conditional random field directly to the segmentation probability map generated by the neural network.By using the structural prediction ability of the dense conditional random field for local features,the target contour can be accurately located and its boundary can be refined.And proved its practical value through experiments.