| Nasopharyngeal carcinoma(NPC)is a head and neck malignancy with a high incidence,which is more common in southern China.Due to the particularity of anatomical location and pathological manifestations of nasopharyngeal carcinoma,radiotherapy has become an effective method for clinical treatment.In radiotherapy,the accurate delineation of the normal organs around the target area is a key step in radiotherapy planning,which is usually performed by the radiotherapy physician.However,the delineation of head and neck organs at risk is still very challenging Manual delineation is time-consuming,and the delineation of different doctors are often inconsistent,which affects the efficiency and effectiveness of treatment.The delineation of the head and neck remains challenging.In order to improve the efficiency and repeatability of delineation,automatic segmentation methods still need to be continuously explored.At present,the existing segmentation methods,such as the atlas-based and active contour theory,still have some limitations.With the continuous breakthrough of deep learning in computer vision,deep learning has been widely used in medical image processing.In order to effectively improve the accuracy and testing time of automatic segmentation of head and neck organs at risk in CT images,and to meet the actual clinical needs as much as possible,this paper constructs an automatic segmentation method based on the convolutional neural network,which is the main branch of deep learning.The main work is as follows:(1)Use a three-dimensional convolutional neural network and introduce a multi-scale cascade segmentation strategy.This paper draws on the idea of global positioning and local observation during the diagnosis and treatment of clinicians,it uses a three-dimensional convolutional neural network as the basic structure to use the information in the layer and inter-layer space of volume data(such as CT and MR images),which is different from Two-dimensional processing method.In training and application,the method based on 3D convolution greatly increases the data storage requirements and computational complexity at the same time,places extremely high requirements on hardware performance,and limits practical applications.In order to solve the above problem,this paper introduced multi-scale strategy,segmentation is divided into sub-models:localization and fine segmentation.After processing of the input image at different scales,the cascaded sub-model gradually removes redundant information from the image to obtain the final segmentation result.With limited hardware resources,end-to-end segmentation of the entire 3D image is achieved.Experiments show that the method presented in this paper has better segmentation efficiency and further improves segmentation accuracy without increasing hardware requirements.(2)Introduce SE(Squeeze-and-Excitation)module to improve the feature expression ability of the model.Based on CT 3D images,V-Net is used as the basic structure of the network to further explore in the paper.Among them,the information between layer and layer of volume data is extracted by three-dimensional convolution operation,and residual structure is introduced to promote gradient flow.To further improve the performance of the model,this article combines the SE module with the residual module.At the same time of feature learning,it introduces the modeling of the importance of the features learned in the network,and enhances or suppresses the features according to the correlation between the feature map and the segmentation task.In this paper,CT images of patients with nasopharyngeal carcinoma were used for experiment.The experimental results indicated that,for 22 head and neck organs at risk with large differences in the size and shape,the segmentation efficiency and precision of this method were improved.While not increasing hardware consumption,the average test time was reduced to less than 3 seconds,and the segmentation accuracy was improved by 9%,achieving better performance. |
PDF Full Text Request