Research On The Method Of Lung Nodule Computer-aided Diagnosis Based On Attention Mechanism

Worldwide,lung cancer is becoming the main cause of cancer deaths.Most lung cancer cases are caused by malignant nodules.Early diagnosis of lung cancer needs to find pulmonary nodules in chest CT(Computed Tomography)images.In recent years,computer-aided diagnosis has become a research hotspot in the field of artificial intelligence and medical imaging,which has practical application value in intelligent medicine and fast diagnosis.In this thesis,we research on computer-aided diagnosis methods based on chest CT images,including lung nodule detection,lung nodule segmentation,benign-malignant lung nodule classification.The specific research contents of this thesis are as follows:Firstly,for the task of lung nodule detection,we propose the lung nodule detection algorithm based on spatial context attention.The Non-local backbone network is designed to extract the features of chest CT images.The 3D RPN(Region Proposal Network)is used to get the spatial locations of lung nodules in an end-to-end way.Combined with false positive reduction network,the refined detection results are obtained.In the part of loss function,the OHEM(Online Hard Example Mining)strategy is used to address the sample imbalance problem in the dataset.Experimental results verify the effectiveness of the proposed algorithm.Secondly,according to the three-dimensional characteristic of chest CT images,we propose a lung nodule segmentation network based on dense attention.The U-Net network structure is utilized as the basic framework,we use the 3D Resnet-34 as the encoder,the dense attention module is designed in the decoder to enhance the feature expression ability of the network,and the PVP(Pyramid Volumetric Pooling)module is designed at the end of the network to obtain multi-scale context information of the network.The experimental results show that the proposed network improves the segmentation accuracy of lung nodules.Finally,aiming at the parameter redundancy problem in benign-malignant lung nodule classification networks,a benign-malignant lung nodule classification network based on channel and spatial attention is proposed.The algorithm designs the search cell to construct the search network.The differentiable architecture search method is used to optimize the search network on the lung nodule dataset.And the important connections in the search cell are retained according to the experimental accuracy after optimization,then we obtain the optimized cell.We design the channel and spatial attention models,which are used to construct the benign-malignant lung nodule classification network combined with the optimized cell.Experimental results show that the proposed network achieves the state-of-the-art classification accuracy.Based on above research contents,this thesis designs and implements the lung nodule computer-aided diagnosis system.The system integrates the lung nodule detection,lung nodule segmentation,benign-malignant lung nodule classification algorithms proposed in this thesis,which has been verified by testing each function.