Research On Key Technologies Of Lymph Node Zone Recognition And Lymph Node Segmentation

With the development of artificial intelligence technology,medical intelligence has played an important direction of the development of artificial intelligence.However,there are still many technical problems to be solved in the application of artificial intelligence technology to medical image analysis.Based on two medical image modalities of PET and CT,this thesis focuses on the study of thoracical lymph node zones and pathological lymph nodes,including the recognition of lymph node zones,the recognition and segmentation of thoracical pathological lymph nodes,and disease quantification of pathological lymph nodes.Specifically,the main research work and innovations are as follows:(1)This thesis presents a method of lymph node zone recognition based on AAR fuzzy model.It is difficult to directly model the lymph node zone,however,the organs which are easy to locate and identify in the chest cavity are selected as the auxiliary anchor points.Hence,the optimized hierarchical recognition structure of thoracic organs and corresponding lymph node regions is established by grid search algorithm.On the basis of determining the recognition structure,AAR fuzzy model is used to model the relative distance,scale and direction between the designated anchor organ and the corresponding lymph node region,so as to realize the automatic recognition of the lymph node region.The experimental results show that,compared with the single organ hierarchical recognition structure,the proposed method can effectively improve the performance of lymph node zone localization and recognition.(2)In this thesis,a new method based on shape prior is proposed.Aiming at the problem that lymph nodes are not obvious in CT and PET images and difficult to locate accurately,a spherical filter based on statistical analysis is designed to locate lymph nodes by using the shape prior knowledge of ellipsoid shape of lymph nodes,which reduces the false positive ratio of candidate lymph nodes and improves the accuracy of lymph node localization.Secondly,combined with the classification function of support vector machine,a three-level classification strategy of sphere,section and pixel and iteration was designed to increase the overall dimension of lymph node characteristics and improve the accuracy of lymph node recognition.(3)A quantitative evaluation method of disease based on semi-Gaussian function is proposed.In order to solve the problem that the boundary of the diseased lymph node in PET image is uncertain in the quantitative evaluation of disease,this method uses the characteristics of semi-Gaussian function weight gradient,and uses a fuzzy membership degree method to deal with the degree of lesion of lymph node and the pixel value near the lymph node boundary,and finally realizes the disease quantitative evaluation based on the pathological lymph node.The experimental results show that,compared with the direct calculation of the total lesion glycolysis(TLG)value of the disease evaluation index on PET images,the quantitative evaluation method based on semi-Gaussian function proposed in this thesis can significantly reduce the error in the quantitative evaluation process.(4)In this thesis,a new method of lymph node segmentation based on atrous convolution network is proposed.Firstly,aiming at the problem of the stability and efficiency of deep convolution network training caused by the imbalance of positive and negative samples,a sine and cosine loss function is designed,which can allocate small value weight of easy classification samples and large value weight of difficult classification samples in the training process,so as to improve the efficiency of network stability training and segmentation performance.The experimental results show that compared with other loss functions under the same network,the proposed sine cosine loss function can improve the training time and segmentation performance.Secondly,in order to solve the problem of low resolution of feature map caused by down sampling pooling operation in traditional convolution network,this thesis proposes a lymph node segmentation network based on void convolution,which introduces multiple atrous space pyramid pooling modules(ASPP)into Seg Net segmentation network,which enhances the ability of multi-scale feature learning of segmentation network and improves the network pairs The performance of the same scale lymph node segmentation.The experimental results show that compared with FCN,Segnet and Deep Labv3+ semantic segmentation network,the two kinds of node segmentation network based on atrous convolution can improve the performance of lymph node segmentation as a whole.(5)In this thesis,a new method of lymph node segmentation based on boundary loop attention mechanism is proposed.In order to solve the problem that the pixels near the boundary of lymph nodes are more difficult to be segmented,the concept of boundary loop is proposed.The pixels in the boundary ring are given more weight in the process of network training,which improves the attention of the network to the pixels in the lymph node boundary ring during the training process.At the same time,considering the fact that dice loss function can balance lymph nodes of different scales,this thesis also proposes a joint training method of dice loss function and other loss functions based on boundary loop.The experimental results show that the boundary loop attention loss function combined with dice loss function as regularization factor can effectively improve the overall stability of lymph node segmentation without sacrificing segmentation performance.