Research And Application Of U-Net Model In Deep Learning

Recently,with the rapid development of deep learning,convolutional neural network(CNN)is widely used in computer vision task because of its powerful feature extraction ability.The fully convolutional network(FCN)based on CNN is also a kind of deep learning method,which has been widely used in image segmentation task,but it has not achieved the expected results in some medical image segmentation tasks.Based on FCN,the deep learning U-Net model has better performance in biomedical image segmentation task.The deep learning method based on U-Net has become one of the mainstream methods in image segmentation task.However,due to the different image quality,the size and shape of target,the confusion of similar structure,data imbalance and others,the segmentation accuracy of U-Net needs to be improved for the fine-grained segmentation.Therefore,it has important theoretical and application value to study the theory of deep learning U-Net model and further improve the fine-grained segmentation performance of U-Net model.This paper discusses the theoretical knowledge of deep learning U-Net and proposes two improved U-Net models:Firstly,we propose deep recurrent U-Net model(DRU-Net).Inspired by residual learning model,we propose the DRU-Net combining with residual learning in the third chapter.It is designed with recurrent residual convolutional neural network(RRCNN)blocks to replace regular convolutional layers in the downsampling path and upsampling path.Compared with the classical U-Net model,DRU-Net model has a deeper network and stronger feature accumulation ability,which has certain advantages in fine-grained image segmentation.Secondly,we propose an enhanced residual U-Net(ERU-Net)model.For the task of fine-grained image segmentation,extracting more local information can improve the segmentation accuracy.So we propose an ERU-Net model with one downsampling path and three upsampling paths in the fourth chapter.The special structure of the model forms three U-paths,and the feature fusion among the three U-paths enables the model to capture more local information,which can achieve more accurate segmentation.We apply the two models to the fundus image segmentation task and compare the two models with the classical U-Net,ResU-Net and some existing methods on the public datasets.The numerical experiments results show that both DRU-Net and ERU-Net have a higher AUC value.Especially,in some challenging regions with low local contrast and similar structures,DRU-Net and ERU-Net can accurately segment lesions.In addition,the selection of loss function has a certain impact on the performance of the model.We analyze the function idea of loss function in U-Net model and make a comparison in the last part of the paper,so as to provide a reference for better selection of appropriate loss function in the future.