Application Of Multi-scale Analysis In Deep Learning

Classification and identification of pulmonary nodules is a key technical link in computeraided detection systems for lung tumors.In order to improve the classification accuracy and the sensitivity of lung nodule while reduce the false positive rate.This paper gives a comprehensive studies of the sample expansion and preprocessing method based on the review of the current research of lung nodule classification and recognition technology.We also discuss the influence of convolution kernel construction on network performance,the structure design of convolutional neural network for lung nodule classification.The main research work of this paper includes the following aspects:(1)A sample preprocessing method based on multi-scale analysis is proposed.The proposed method combines high-frequency filtering and wavelet transform,which obtains extended samples through multi-channel sample extraction method.Noise reduction is performed in time domain and frequency domain respectively.Wavelet decomposition on different scales can effectively maintain the meaningful sample information,which improves the classification quality of the network.(2)Different convolution kernel scale construction methods are studied in the proposed network.The size of different convolution kernels has a large impact on the training model in different network.Base on the study the convolution kernel scale and the effect of different networks,we can obtain the best size of the convolution kernel,which can improve the classification accuracy of the network.(3)The number of convolutional layers and pooling layers in neural network is studied.LIDC-IDRI dataset is used in this paper,different combinations of convolution layers and pooling layers are applied in order to find the best one for the networkAdopt the above research results,Preprocessed the LUNA16 training data,And built Convolutional neural network,Trained the effects of multi-scale pre-processing,The generated network model can effectively detect lung nodules in the LUNA16 test set,Accuracy rate reached 89.6%.