Research On Detection Method Of Pulmonary Nodules Based On Improved Residual Structure

Lung cancer is a common disease with high incidence.Generally,lung nodules formed by the irregular growth of certain cells in the lung are early signs of lung cancer,so the detection of pulmonary nodules is particularly important in the prevention and detection of lung cancer.With the development of medical image processing and artificial intelligence,deep learning has made many research results in pulmonary nodule detection.However,the lack of feature extraction receptive fields,high model computational complexity,and overfitting in the pulmonary nodule detection model affect the results of pulmonary nodule detection.Aiming at the above problems,this paper makes in-depth research on pulmonary nodule detection,and proposes a pulmonary nodule detection method based on improved residual structure.Main tasks as follows:Firstly,an improved residual structure 3D Res-I(3D Residual Improved)based on rectangular convolution kernels,group convolution,and pre-activation operations was proposed to address the problems of large model calculations and over-fitting caused by complex network structures in pulmonary nodule detection methods.Combining this structure with the improved U-Net structure,as the feature extraction part of the Faster R-CNN network,a pulmonary nodule detection model based on 3D Res-I structure was obtained.The model reduces the amount of calculation of the model and reduces the complexity of the model by using group convolution.By increasing the pre-activation operation,the regularization of the model is improved and the over-fitting phenomenon is alleviated.At the same time,the rectangular convolution kernel is used to expand the receptive field of the convolution operation on the premise that the calculation amount of the model is slightly increased,effectively taking into account the global and local characteristics of pulmonary nodules.The experimental results on the LUNA16 date set show that this model improves the detection accuracy of pulmonary nodule,reduces the average number of false positives and the size of the generated model.At the same time,the validity of the proposed model was further verified by visualizing the detection results of pulmonary nodules.Secondly,in order to further reduce the complexity and calculation of the model,this paper proposes an improved residual structure based on depth separable convolution 3D Res-DSC(3D Residual Depth Separable Convolution).This structure uses depth separable convolution to replace the group convolution in 3D Res-I.Applying this structure to pulmonary nodule detection,a pulmonary nodule detection model based 3D Res-DSC structure was obtained.The model separates regions and channels through depth separable convolutions,which greatly reduces the number of parameters and model calculations.The experimental results on the LUNA16 data set show that the pulmonary nodule detection module based on 3D Res-DSC structure further reduce the computational burden and complexity of the pulmonary nodule detection model on the premise of ensuring the detection accuracy,and reduce the model size by nearly 10 times.This enables a lightweight development of the model,which facilitates rapid iteration of the model.