Lung ROI Feature Extraction And Selection For Classification Of Benign And Malignant Nodules Based On Radiomics

As evryone knows,cancer has become one of the leading causes of morbidity and mortality all over the world.Diagnosing the benign and malignant tumors by biopsy,not only the patients need to have an invasive surgery,but also the doctor is not concinent to observe the tumor heterogeneity.By analyzing and quantification the descripitive features of radiographic image,the Radiomics matches the tumor phenotype with the evaluation of curative effect,and provides a good basis for the clinical diagnosis for lung cancer patients.In recent years,the deep learning methods have been widely used in medical image processing.Despite using the original radiographic image for deep learning can get a good recognition performance,we can't know the relationship between image features and classification results because of its closed way of learning,and the traditional shallow machine learning methods usually have some limitations.In this paper,the main work is as follows:(1)The traditional statistical methods are used to extract a total of 143 features which contain the geometric features,texture features and histogram features from the Regions of Interesting in CT images of Lung.The 143 features are as the original feature sets which used to research the classification model of benign and malignant nodules.(2)Inspired by Genomics in genes selection,the Recursive Feature Elimination is introduced in Relief feature selection algorithm,which named RFE-Relief feature selection algorithm.The RFE-Relief feature selection algorithm overcame the shortcomings that the Relief algorithm can't eliminate redundant features.A subset that 46 features with low dimension and the strong correlation between benign and malignant nodules was obtained.(3)A Deep Belief Network model has been constructed including three layer Restricted Boltzmann Machines,and a Softmax classifier has been mixed at the top of the model.A feature subset that contain 46 features is selected as input of Deep Belief Network.The purpose of training and fine-tuning of the Restricted Boltzmann Machines and Softmax are to classify benign and malignant nodules.The experimental results show that the final classification accuracy is 93.8%.Feature selection not only reduces the running time and efficiency of the algorithm,but also improves the accuracy of classification.So it can assist clinicians in the diagnosis of lung cancer.At the end of this paper,we discuss the number of hidden layer nodes and hidden layers of Deep Belief Network in the classification of benign and malignant nodules.And the results show that the Deep Belief Network model constructed in this paper can achieve a better classification task of the benign and malignant nodules.