Research And System Design Of CT Image Segmentation Technology For Nasopharyngeal Carcinoma Based On Deep Learning

Nasopharyngeal cancer is a serious threat to people's health,and medical image segmentation plays a vital role in the radiotherapy plan for nasopharyngeal cancer.It can increase the accuracy of disease diagnosis and increase the probability of successful cancer treatment.However,Nasopharyngeal carcinoma(NPC)has many problems,such as various shapes,uneven density,blurred boundary and complex shape of lesions,which have brought great difficulties to the accuracy of CT image segmentation of nasopharyngeal carcinoma.With the rapid development of deep learning,researchers have applied deep learning to the field of medical image segmentation,and used deep learning algorithm to extract the best image features of medical images,thus improving the accuracy of medical image segmentation.Therefore,the thesis uses deep learning technology to study the CT image segmentation of nasopharyngeal carcinoma.The main research results are as follows:(1)For the difficulty of CT image of nasopharyngeal carcinoma and the convolutional neural network,such as time-consuming,large storage and small global feature extraction,a CT image segmentation method for nasopharyngeal carcinoma based on full convolutional network is proposed.This method makes full use of the network structure advantages of full convolution,up-sampling and skip structure,and adds a second path based on full convolution network.This dual-path structure can combine local features and larger context features.(2)In the full convolutional neural network,in order to solve the problem that downsampling the feature map reduces the spatial resolution of the image and the accuracy of the segmentation result,a three-dimensional convolutional neural network method which combining expansion convolution,residual connection and conditional random field is proposed.This method replaces the conventional convolution with expansion convolution,so that features can be extracted from image blocks with high spatial resolution.residual network is adopted to solve the problem of network optimization,and conditional random field is added after the output layer of 3D CNN network,which smoothness constraint information between pixels is added into pixel classification to improve the accuracy of nasopharyngeal carcinoma CT image segmentation.Experimental results show that the proposed 3D CNN model has good segmentation performance for nasopharyngeal carcinoma CT image segmentation.(3)The above two segmentation algorithms were implemented in anaconda by using the Tensorflow deep learning framework,and a CT image segmentation system for nasopharyngeal carcinoma was developed.Through the system,two segmentation method can be arbitrarily selected,which not only can obtain segmentation results with higher accuracy,but also simplify the manual manual operation complexity.The system has functions such as image import and export,image segmentation and 3D visualization.