Research On Aided Diagnosis Methods For Multi-Pathological Types Of Pulmonary Nodules Based On 3D Multi-Resolution Attention Capsule Network

To identify the benign and malignant pathological types of pulmonary nodules is very important for the formulation of accurate and effective diagnosis and treatment plan.The aided diagnosis methods based on deep learning can reduce the dependence on individual experience and working status of doctors,and improve the diagnostic efficiency.However,there are few pathologically confirmed lung CT images in the existing public datasets,which can not guarantee the accuracy and credibility of sample labeling.To better meet the clinical needs,according to the lung CT images confirmed by gold standard collected from the cooperative hospital and research institute,aided diagnosis methods for multi-pathological types of pulmonary nodules based on 3D multi-resolution attention capsule network is studied.According to the needs of aided diagnosis methods and the performance analysis of deep learning methods,the capsule network,which has more advantages than convolutional neural network,is selected as the infrastructure to construct the aided diagnosis models.The classical capsule network model is improved from two aspects:network structure and dynamic routing algorithm.In the improved design of network structure,in order to fully extract and learn the morphological features and hierarchical information of pulmonary nodules in three-dimensional space,and adapt to the dimensionality of CT images,3D capsule network structures are designed based on the 2D capsule network structure.In the improved design of dynamic routing algorithm,the update rule of the log prior probabilities in original dynamic routing algorithm can not limit the amplitude of update increment,which can easily lead to the long-term inactivation of vector neurons.A solution to ensure the limited update amplitude is proposed by using the cosine similarity between predicted vectors and iterative output vector neurons as the update increment of the log prior probabilities.By visual tracking of parameters change during training,the inactivation phenomenon of numerous vector neurons caused by original dynamic routing algorithm is verified,and the effectiveness of improved dynamic routing algorithm is verified.Compared with the classical deep learning models such as AlexNet,ResNet-18 and ResNet-50,and the multi-modal fusion models based on image features and serum biomarkers,the effectiveness of the aided diagnosis model constructed by 3D Capsule Network with Improved Dynamic Routing Algorithm(3D CapsNet-IDRA)is verified.On the basis of 3D CapsNet-IDRA,in order to explore the influence of interpolation disturbance introduced by multi-resolution method on classification performance of multi-input and single-input aided diagnosis models,cross-path multi-resolution attention mechanisms are proposed to suppress and enhance disturbance information.These two types of mechanisms are based on the feature map of the path in which the real resolution input is located to measure the deviation of the feature map from the corresponding position of the other paths.Then the deviation is used to calculate the soft attention distribution of each element in the feature map in space,and then to exert influence in the form of multiplication coefficient.In this way,3D multi-resolution attention capsule networks(3D MRA-CapsNets)with multi-input are constructed by embedding multi-resolution attention modules on the basis of 3D CapsNets-IDRA.Through the comparative analysis of 3D MRA-CapsNets and 3D CapsNets-IDRA,the different effects of interpolation disturbance introduced by multi-resolution method on the classification performance of multi-input and single-input models are demonstrated respectively.Similarly,compared with the classical deep learning models and the multi-modal fusion models,the good performance of aided diagnosis models based on 3D MRA-CapsNet is further verified,and their performance are better than that of 3D CapsNet-IDRA.