Research On Lung Nodule Classification Algorithm Based On Convolutional Neural Network

Lung cancer has become the most important one of the malignant diseases that lead to the death of the patients.Because of the high incidence of the disease and the lower cure rate after diagnosis,the survival rate of the patients has been at a low level.CT technology is widely used to provide accurate help for the diagnosis of lung lesions.However,there are 3 main reasons for the high number of patients with lung cancer:(1)the large number of CT images in the lung,and the experienced reading doctor is not enough to cope with the rapid growth of the cases.(2)there are more vessels and tissues in the lungs,which are not easily distinguishable from the lung lesions.(3)the traditional method of machine learning is to extract the characteristics of the low level of the focus by artificially.Different experts have different extraction methods,and the different features have great influence on the classification results.The computer aided diagnosis system for the CT scan image of the lung can effectively improve the doctor's diagnostic rate,reduce the doctor's burden,and provide the patient with more objective diagnosis results.The computer-aided diagnosis system is mainly composed of two parts:the extraction of the region of interest and the classification of benign and malignant nodules.The extraction of region of interest is the basis of subsequent classification.By dividing the lung parenchyma and extracting the region of interest,the interference of unnecessary tissues and organs can be removed,and the amount of data can be reduced and the accuracy of the classification is improved.In addition,the ability to identify lung nodules accurately is a criterion to evaluate the performance of a classifier.Two key points of computer aided diagnosis system are studied in this paper:extraction of interested regions and classification of pulmonary nodules in CT images.1.In this paper,in the traditional segmentation method,there are too many manual segmentation or only segmentation of lung parenchyma while the lung parenchyma still remains a large number of tissues.A segmentation algorithm for ROI is proposed.First,we use a histogram equalization method to enhance the original lung CT image.Then the lung parenchyma was segmented by iterative threshold method and morphology.Finally,a morphological and method is used to extract ROI from lung parenchyma images.The methods used in this paper mainly include the following two innovations:(1)The histogram equalization method is used to improve the clarity and discernibility of the image,which is beneficial to the segmentation of the lung parenchyma in the next step.(2)Using morphological method to de noise the lung parenchyma and reduce the irrelevant noise.2.We use a convolutional neural network based Convolutional neural network(CNN)CT image classification model.First,a network of 9 levels is designed,including four volumes,four pools and a full connection layer a.After comparison,TanH was selected as activation function,and 0.0001 as learning rate.The main innovation of this part is to apply the deep learning method in the classification of pulmonary nodules,avoiding the artificial extraction features,and the classification results are more objective and accurate.3.The experiments on the LIDC dataset show that the accuracy,sensitivity and specificity of the custom 9 tier CNN classification are 89.0%,90.9%and 87.2%respectively.The pre-processed data sets are improved more than the original data set in various classification indexes,and the best results are compared with other classification methods.