| Nasopharyngeal carcinoma is one of the common head and neck malignant tumors in Guangdong,Zhejiang and other southern regions of China,with a high incidence.The nasopharynx is located in the center of the skull,and the surgical path is complex,so it is difficult to remove the whole tumor by surgical.The majority of nasopharyngeal carcinoma is poorly differentiated squamous carcinoma with high sensitivity to radiotherapy.So Radiotherapy is the preferred treatment for nasopharynx cancer.The accurate delineation of the gross target volume(GTV)and organs at risk(OAR)is crucial to radiotherapy for nasopharynx cancer.However,the delineation is challenging due to the morphological complexity of the target area and several surrounding OAR.Moreover,the efficiency and precision mainly depend on the clinical experience and knowledge of oncologists,which makes the delineation workload-heavy,time-consuming,and subjective.Nowadays,a variety of image segmentation techniques has been applied in Automated segmentation of nasopharyngeal carcinoma.A lot of data preprocessing work needs to be done manually,because the segmentation accuracy of direct segmentation mothod will be reduced due to the introduction of more useless pictures and backgrounds.secondly,The current auto-delineation networks are difficult to capture the unique features corresponding to the different target size,and the segmentation accuracy needs to be improved.In order to promote the application of automatic segmentation of nasopharyngeal carcinoma into clinical practice as soon as possible,this thesis proposes a whole system design scheme of radiotherapy from classification,detection to segmentation for nasopharyngeal carcinoma based on a two-dimensional convolutional neural network.The main work is divided into the following two parts:1.Locating the target region for segmentation of nasopharyngeal carcinoma based on convolutional neural networks.In order to solve the problem that too many useless pictures and backgrounds are introduced into the direct segmentation network,which leads to the decline of segmentation accuracy,This thesis proposes a concept of classification,detection and segmentation of the whole set of nasopharyngeal carcinoma CT images.Firstly,All of CT images of patients(from the top of the head to the chest)are used to find the target layer by the Densenet.Secondly we locate the target region for segmentation of nasopharyngeal carcinoma by the Efficientdet on the target layer.Finally we took the location as a reference,and make a minimum box containing all target as the input of the segmentation network.We can segment the target region more accurately by this way.2.Automatic segmentation network for nasopharyngeal carcinoma based on different receptive fields.In order to solve the problem the problem that The current auto-delineation networks are difficult to capture the unique features corresponding to the different target size,This thesis presented a novel segmentation network MA_net for nasopharyngeal carcinoma,creating an MACSPP(Multiple Atrous Convolution Spatial Pyramid Pooling)module combining the Dilated Convolution with the pyramid pooling module.This module could capture the unique features corresponding to the different target size without increasing in parameters and calculation,Meanwhile,the channel weight self-adjustment mechanism and concating the deep and shallow layers of information were introduced into the network structure to enhance the feature learning ability of the network.The results showed that the floating points,and the random access memory and the number of parameters of MA_net were less than those of Deeplabv3+,PSPNet,Unet++and U-Net,the four major segmentation networks.but the mean dice is 0.66%,0.91%,1.23%and 3.44%higher,respectively.To sum up,the algorithm proposed in this paper,"classification,detection and segmentation" can segment the target region more accurately.compared with the current auto-delineation networks,the proposed segmentation network MA_NET has higher segmentation accuracy of target GTV and seven organs at risk(the left and right eyes,the left and right lobes of brain,the left and right parotid glands and tongue).This scheme is expected to serve for future intelligent radiotherapy. |
PDF Full Text Request