Study On Pulmonary Nodule Malignancy Classification Based On Multiple Features

Lung cancer is a widespread disease,which has one of the highest morbidity and mortality among all the cancers around the world.Studies have shown that accurate early diagnosis can effectively improve the success rate of treatment and the survival rate of patients.In the diagnosis of lung cancer,radiographers usually use CT(Computed Tomography)images to screen lung nodules and diagnose their lesion types.After the lung nodule is detected,it is benign and malignant diagnosis that is a very important step.However,limited by the radiographers’ knowledge and experience level,manually nodule detecting is prone to high misdiagnosis rate.For this reason,computer-aided diagnosis(CAD)systems have been extensively developed to help doctors process and analyze patients’ CT image,so as to realize automatic diagnosis and nodule malignancy classification.In recent years,with the development of deep learning,CAD systems have further improved the efficiency and accuracy of clinical diagnosis of pulmonary nodules through the combination of deep neural networks.Based on the public lung cancer CT image data set LIDC-IDRI(Lung Image Database Consortium and Image Database Resource Initiative),in order to improve the accuracy of nodule malignancy classification models and explore the effect of various features,this paper has undertaken such research works as followed:1.It has designed an approach of dataset preprocessing and nodule image segmentation for LIDC-IDRI.2.For exploring the utility of traditional image features(such as structural features,texture features,etc.)in nodule malignancy classification task,four traditional image features of nodules were extracted,and Support Vector Machines(SVM)was applied to classify benign and malignant nodules.Finally,the classification results were analyzed for explore the advantages and disadvantages of various traditional image features.3.A novel model based on multilevel features and bilinear pooling is proposed,which employs two paralleled streams for feature extraction.For each feature extraction stream,multi-level features are extracted through a series of convolutional neural network(CNN)-based convolutional layers.Both fine and semantic features from the two CNN streams are paired,and then multiplied using outer product and pooled.The bilinearly pooled features are put into a specified SVM classifier.The proposed bilinear model is designed to fuse similar features to intensify the distinguishability of the nuances in the benign and malignant pulmonary nodules.4.On the basis of the multilevel feature bilinear pooling model,a weighted voting mechanism based on Gaussian distribution is added to merge the nodule classification results and further improve the classification accuracy.Through the above research work,the proposed classification model has achieves accuracy of 94.97%,which is a state-of-the-art performance in this research field at present.