Research On Lung Nodule Detection Method Based On Multi-scale Deep Feature Fusion

In recent years,the morbidity and mortality of lung cancer have been increasing.Lung nodules are symptoms of early lung cancer.Accurate detection of lung nodules helps to better detect and diagnose lesions,which is also the key to early detection of lung cancer means.At present,the screening of pulmonary nodules is mainly achieved through the analysis and diagnosis of chest computed tomography images(CT images).With the rapid development of deep learning,convolutional neural networks are widely used for medical aided diagnosis of lung nodules.and achieved great success.Using deep learning methods to detect and classify pulmonary nodules can make the diagnosis of pulmonary nodules more accurate.Aiming at the problems of common convolutional neural network being easily affected by the size of lung nodules and incomplete feature extraction in CT image detection of lung nodules,this thesis proposes a lung nodule based on self-selective convolution kernel and multi-scale feature fusion algorithm.The nodule detection and classification method can effectively detect candidate lung nodules including tiny nodules,and classify the nodules from benign and malignant nodules.This article includes two parts: detection of lung nodules and classification of benign and malignant lung nodules.In the stage of lung nodule detection,this thesis designs a U-Net-shaped 3D convolutional neural network combined with a region extraction network structure to detect lung nodules.Aiming at the problems that lung nodules are not fixed in shape,different in size and easy to be confused with other tissues in CT images,this thesis designs a 3D SKRes Net convolution module,which can effectively select a convolution kernel of suitable size for feature extraction.Thus,lung nodules of different sizes are detected.In the classification stage of benign and malignant lung nodules,a residual network model based on multi-scale feature fusion is constructed.This method can adaptively adjust the receptive field according to the multi-scale convolution results of the input information,so as to distinguish different features more efficiently and solve the problem.The problem of difficulty in extracting features for tiny nodules,while ignoring and invalid features,provides better classification of lung nodules with different resolutions and sizes.In this thesis,the Luna16 dataset and LUNG-40 dataset are compared with several current mainstream deep learning methods.The lung nodule detection method performs well in sensitivity,FROC curve,and AR indicators.Based on multi-scale feature fusion The results of the classification method of benign and malignant lung nodules in the classification stage have high accuracy,sensitivity and specificity.The results show the effectiveness and superiority of the algorithm in this thesis.