Research On Organ Segmentation Algorithm Of Medical Image Based On Convolution Neural Network

This paper mainly focused on the technique of simultaneous segmentation of esophageal cancer radiotherapy target areas and surrounding organs at risk.In recent years,the convolution neural network is the main part of medical image segmentation/organ segmentation.Despite the encouraging results,there are two shortcomings.On the one hand,there are many literatures on single organ segmentation,but few literatures on multi-organ segmentation.On the other hand,many segmentation methods have better multi-objective segmentation effects on natural image segmentation data sets,but poor results on esophageal cancer medical image segmentation.In view of the above problems,this paper conducts research on organ segmentation of medical image,and the main work is as follows:(1)In order to accurately segment the organ region,this paper adds a normalization layer in u-net network to accelerate the convergence rate of the network and prevent the gradient from disappearing,thus improving the segmentation accuracy.In addition,in order to improve the segmentation speed of the convolutional neural network model,this paper skillfully uses1×1 convolution kernel and the appropriate number of convolutional channels to reduce the model parameters,thus improving the efficiency of organ segmentation.(2)In order to better apply the Mask R-CNN algorithm to esophageal cancer organ segmentation,this paper made four improvements on the Mask R-CNN algorithm,and proposed a new segmentation algorithm named MsMR-Net deep learning algorithm.The specific improvement is as follows: a)The ResNet network structure in the feature extraction network is improved by improving the Mask R-CNN algorithm,so that the trunk network can better obtain the information of esophageal cancer image organs;b)Carry out multi-scale ROI Align operation on ROI obtained from the regional recommendation network,and integrate multi-layer feature graph information.c)On the basis of the original Mask branch,the full convolutional network branch containing two 3*3 convolution layers was added,and the output Mask of the two branches was fused to obtain the final segmentation result.d)As this paper is aimed at image segmentation of esophageal cancer organs,it is very important to be sensitive to the segmentation results of boundary.Therefore,this paper proposes an improvement to the loss function by adding a boundary weighted loss function.(3)At the same time,this paper successfully transplanted the trained model to Windows by using OpenCV4.0(Open Source Computer Vision Library),and designed an image segmentation interface combining QT ? VTK(Visualization ToolKit,VTK)? ITK(InsightSegmentation and Registration Toolkit,ITK).The interface function includes the ability to read in a sequence of medical CT images of esophageal cancer DICOM(Digital Imaging and Communications in Medicine,DICOM),and by clicking the corresponding button,the target organ region segmted by the model trained by the algorithm in this paper can be displayed on each slice of the image read.It is very convenient to operate and can be used as a reference for doctors in the radiotherapy treatment of esophageal cancer.