Research Of MRI Segmentation Based On U-shaped Deep Network

Magnetic resonance imaging(MRI)is a widely used medical imaging technique.It can obtain the structure information of tissues and organs inside the organism in a non-invasive way and provide great reference value for doctors to diagnose and treat diseases.Medical image segmentation technology can quickly extract a specific area of the original data,greatly improving the doctor's work efficiency.In the past,medical images were usually segmented using traditional image processing and machine learning methods,but these methods were often analyzed based on the shallow features of images and therefore their processing power was very limited.In recent years,the technology of artificial intelligence,represented by deep learning,has been greatly developed,and has achieved revolutionary achievements in many fields.Compared with traditional methods,deep learning network is more conducive to extract the deep abstract features of data,and the end-to-end learning mode enables deep network to directly process raw data,and truly realize data driven.Therefore,it is of great significance to study and design the depth network for medical image segmentation.In this paper,the PROMISE12 data set is used,and prostate MRI segmentation becomes a typical application of the method described.But the method is not limited to this dataset.For the MRI segmentation task,this paper studies the structure of 2D and 3D fully convolution networks(FCN),including the conversion of 2D network to 3D network.For 2D FCN,this paper first analyzes the current leading U-Net depth learning model.Through the introduction of the residual structure,dense network and other current ideas,several improvements are put forward.The improved network and the original U-Net is trained and tested on the same MRI dataset.For 3D FCN,it's realized by generalizing the improved U-Net to its 3D form,enabling the network to directly process 3D data,and further improving accuracy and efficiency.Finally,by summarizing and analyzing the above 2D and 3D FCN,the paper proposes a more generalized U shaped network structure,and the feature fusion method of the structure is analyzed.Through the experiment,the advantage of the general U network in the image segmentation application is proved.First of all,this paper studies the performance of 2D FCN in MRI segmentation task.Because 2D CNN can not directly process 3D data,so the 3D data needs to be stratified.Based on the analysis of FCN and U-Net networks,this paper proposes two 2-D FCN based on U-Net improvement named UR-1 and UR-K network.Based on U-Net,UR-1 introduced the residual structure and some other optimizations.However,UR-K further improves the input layer of the network on the basis of UR-1,and uses upper and lower K-layer data adjacent to the target layer as the data of the layer for processing,in order to extract related spatial structure information Which will help to segment the MRI data more accurately.In this paper,U-Net,UR-1 and UR-K network are respectively trained and tested on the same MRI data set.Experiments show that the results of UR-1 and UR-3 are better than U-Net and UR-3 is slightly superior to UR-1.Secondly,this paper studies the performance of 3D FCN in MRI segmentation task.This paper analyzes the transformation from 2D network to 3D network,extends UR-1 network to 3D structure and names it as 3D-UR network.3D CNN can extract more spatial information,and can directly handle the 3D data.Through a feedforward calculation,the 3D FCN can get the result of MRI segmentation.Compared with the 2D CNN,it has a higher efficiency.In this paper,the 3D-UR network is trained and tested on the same MRI dataset.Experiments show that the 3D CNN can extract more pattern features that are beneficial to decision-making,and the result is better than the UR-3 network.Finally,this paper analyzes the U-shaped network used in constructing 2D and 3D FCN.By simplifying and abstracting U-Net and adding the residual structure,a universal U-shaped network structure is proposed.This article focuses on the feature fusion architecture of universal U-shaped network,respectively the Skip and Shortcut connections.The U-shaped network's Skip connection is to stack the shallow and deep feature maps and then learn the fusion mode through the CNN.Shortcut connection is a residual structure,which is good for network convergence and network performance.By removing Skip from 3D-UR network,we get 3D-UR-RS network.And by both removing Skip and Shortcut,we get 3D-UR-RS network.In this paper,the two networks are trained on the same MRI data set.The experimental results show that Skip and Shortcut play a key role in the feature fusion of universal U-shaped networks.