Research On Medical Image Segmentation Methods Based On U-Net And Clustering

Medical image segmentation is the core of medical image processing technology,whose accuracy and processing speed are directly related to the accuracy and efficiency of diagnosis.At present,there are two mainstream medical image segmentation methods: the traditional algorithm led by clustering algorithm and the deep learning algorithm led by the U-Net network.This dissertation studies the medical segmentation algorithms from these two directions respectively.The main research contents and innovations are given as follows:(1)To solve two problems of noise and data uncertainty commonly existing in medical image segmentation by fuzzy clustering algorithms,a medical image segmentation based on noise robust intuitionistic fuzzy c-means(NR-IFCM)clustering algorithm is proposed.In order to better deal with the noise in magnetic resonance image(MRI),the noise robust intuitionistic fuzzy set(NR-IFS)is applied to the NR-IFCM algorithm.NR-IFS uses the majority dominant suppression similarity function which combines neighborhood statistics and competitive learning.Compared with traditional intuitionistic fuzzy set(IFS),NR-IFS has better noise resistance.In addition,in order to overcome the problem of data uncertainty,a new intuitionistic fuzzy factor is proposed,which combines the local gray level and the spatial information.Simulation experiments are carried out,and the experimental results can verify the performances of the proposed medical image segmentation method based on NR-IFCM is more superior compared with several common fuzzy clustering algorithms.(2)In view of the lack of the ability to mine information from the perspective of full scale in medical image segmentation based on U-Net network,a new medical image segmentation method based on full scale connected U-Net(FSC-UNet)network is proposed.The full-scale residual connection can combine the high-level features and low-level features from different scale feature maps.Low-level feature maps capture a wealth of boundary information that can highlight the boundaries of organs,and high-level feature maps learn more about semantic information.The combination of high-level and low-level features can prevent feature loss during the upsampling process and improve the accuracy of medical image segmentation.Simulation experiments are carried out,and the experimental results show that the proposed medical image segmentation based on FSCUNet network has better segmentation performances compared with the classical U-Net network.(3)In view of the problem that the feature extraction ability of medical image segmentation based on U-Net network is sufficient and the medical image segmentation is only applicable to a single dataset,a medical image segmentation method based on densely connected U-Net(Dense UNet)network is proposed.First of all,in order to improve the feature learning ability,a dense block composed of dense connection layer,transition layer and residual connection is proposed.This dense block can also solve the overfitting problem caused by too small dataset and promote the dissemination of information in the network.Furthermore,a multi-feature fuse(MFF)block which combines the dense blocks of different levels is proposed.It is used to learn the features of different levels,improve the accuracy of feature extraction and make the network better distinguish the boundary.Finally,in order to make the proposed network suitable for more complex datasets,a new loss function which combines the cross-entropy loss function and the dice loss function is proposed to deal with the imbalance between the target and the background.Simulation experiments are carried out,and through qualitative and quantitative evaluation,the superiority of the proposed medical image segmentation based on Dense UNet are further verified.