Research On Context Based Pulmonary Nodule Detection Method

The early form of lung cancer is often presented as lung nodules,and if the nodules can be diagnosed in early screening will significantly reduce the incidence of lung cancer.Computer-aided detection systems,which can efficiently and rapidly complete lung nodule screening,have gradually become a hot topic of interest with the increasing development of computer technology and artificial intelligence.In the field of lung nodule detection,the core of computer-aided detection system is the lung nodule detection method.Advanced and superior detection methods will provide important references for physicians and significantly reduce the rate of missed and false detections of pulmonary nodules.Although many lung nodule detection methods have been proposed,there are still problems such as low sensitivity and missed detection when dealing with small nodules,complex nodules and non-significant nodules.In this thesis,we discuss the problems and shortcomings of current lung nodule detection methods and propose a series of lung nodule detection methods focusing on contextual mechanisms,which are as follows:(1)Due to the characteristics of pulmonary nodules often presenting multi-morphology,multi-scale and low differentiation between nodules and surrounding tissues,current lung nodule detection methods often lead to missed detection.To address this problem,this thesis proposes a lung nodule detection method based on 3D multiscale and semantic contextual heterogeneity.Firstly,a multiscale context fusion module is constructed using 3D dilated convolution with different dilated rates to extract multiscale fusion features of lung nodules to enhance the sensitivity to nodules of different morphologies and scales.Secondly,to address the problem of noise and redundant information in multi-scale fusion features,a channel semantic contextual heterogeneity model is proposed to suppress redundant noise and reduce useless information interference,thus making the potential discriminative features of indistinguishable nodules more significant and finally enhancing the ability of nodule detection and localization.(2)The 3D convolutional network-based lung nodule detection method suffers from the problems of tendency to overfit,excessive computational effort and model complexity.In addition,it is difficult to obtain long-range contextual correlation due to the local nature of convolutional operations,which ignores the global information of lung nodules in the whole lung tissue and easily misses the detection of non-significant nodules.To address the above problems,this thesis proposes a lung nodule detection method based on spatio-temporal mixed convolution and multi-axis contexts.Firstly,a spatio-temporal mixed convolutional network is composed of 2D spatio-temporal separation convolution and 3D convolution in the feature extraction network,and the 3D spatio-temporal feature learning is performed by combining the sequence frame dimension;Secondly,a multi-axis context mechanism is proposed to capture the long range information of lung nodule features from the spatial level and the spatio-temporal level by designing a spatial axial attention mechanism and a spatio-temporal axial attention mechanism,respectively,to compensate for the lack of global information in the network.The proposed method effectively reduces the computational effort of the network while ensuring the performance of the detection network.(3)Existing methods often ignore the distribution discrepancy and correlation between intra-frame and inter-frame features of 3D lung nodules when capturing the 3D spatial context of lung nodules.In addition,most of the existing loss functions take the way of bounding box regression to optimize the network,which fails to truly reflect the morphological features of 3D lung nodules and the relative distance between nodules.Based on this situation,this paper proposes a lung nodule detection method based on multiscale contextual correlation and discrepancy.Firstly,a multi-scale feature decomposition network MFD is constructed to obtain multi-scale spatial features of lung nodules;then a local global attention mechanism is constructed to differentially encode the spatial feature context of lung nodules using multi-head dot product attention mechanism and window attention.Then an implicit context decoder is used to decode the encoded context information in 3D space to generate 3D spatial context explicit with more discriminative and accurate location information.The multi-scale feature fusion module is used to obtain 3D lung nodule features that fuse multi-scale spatial context information for detection.Finally,a loss function based on the generalized spherical representation Io U is proposed to optimize the training process of the network and to impel the network to learn more robust lung nodule feature representations.