Research Of Pulmonary Nodules Classification Method Based On CT Image Sequences

In recent years,lung cancer is one of the leading causes of death worldwide.Early diagnosis and treatment of lung cancer can effectively improve the 5-year survival rate of lung cancer,and lung nodules are potential manifestations of lung cancer.CT imaging is an effective screening tool for detecting lung nodules and non-small cell carcinoma.However,due to the complex structure,changeable morphology and uncertainties over time of pulmonary nodules in early stage of lung cancer,it is difficult for experienced radiologists to diagnose malignant nodules and benign nodules in CT images of some cases.At the same time,with the maturity of medical imaging technology and the explosive growth of medical imaging data,people have proposed to use computer-aided diagnosis system to help physicians improve the efficiency and accuracy of diagnosis of lung nodules.In the benign and malignant diagnosis of pulmonary nodules,the extraction of features that can fully and effectively characterize pulmonary nodules is a critical step,followed by appropriate classification techniques to achieve benign and malignant classification of pulmonary nodules.This paper considers the clinical three-dimensional characteristics of pulmonary nodules,and the two-dimensional traditional visual features of nodules are difficult to effectively characterize the spatial nature of nodules.At the same time,single features cannot fully characterize the nodules,resulting in unsatisfactory diagnostic classification accuracy.A method for the diagnosis of benign and malignant pulmonary nodules based on three-dimensional invariant features and a classification method for lung nodules combined with three-dimensional depth and visual features on computed tomography images were proposed.The main research contents and innovations of this paper are as follows:(1)Aiming at the problem that the two-dimensional features extracted from traditional diagnosis of benign and malignant pulmonary nodules could not effectively characterize the heterogeneity of pulmonary nodules,which resulted in low diagnostic accuracy,a new diagnosis method of benign and malignant pulmonary nodules based on three-dimensional scale invariant features was proposed.Firstly,the scale-invariant feature points on the pulmonary nodule sequence image were calculated.According to the gradient magnitude,the orientation information and the gray-scale difference histogram in the three-dimensional neighborhood of each feature point,the three-dimensional scale invariant feature descriptor of the feature point was established.Then the fuzzy C-means clustering algorithm based on distance optimization was used to establish the visual dictionary,and the classical statistical Bag of Word model was used to establish the characteristic representation of each pulmonary nodule.Finally,the support vector machine was used to achieve the benign and malignant diagnosis of pulmonary nodules.The experimental results show that the method can effectively classify pulmonary nodules with high accuracy,sensitivity,specificity and area under the ROC curve.At the same time,it is proved that the three-dimensional features of pulmonary nodules are more distinguishable than the two-dimensional features,and the signs of uncertain lung nodules with malignancy of "3" are more similar to benign nodules.(2)Although traditional visual features are widely used,they only describe a perspective of nodule heterogeneity.It is difficult for a single traditional visual feature to fully and effectively characterize pulmonary nodules.The deep learning model has achieved good results in image classification tasks.However,it is unrealistic to generate large-scale labeled data with pathological experiments or manual labeling by radiologists.The lack of training data may lead to the unsatisfactory effect of deep learning technology.At the same time,considering the three-dimensional characteristics of clinical pulmonary nodules,this paper proposes an algorithm to classify benign and malignant pulmonary nodules by combining three-dimensional deep and visual features.The algorithm extracts deep features based on three-dimensional convolutional neural networks and visual features based on three-dimensional scale invariant transformation texture descriptors and shape descriptors based on three-dimensional shape index.Multiple kernel Adaboost classifiers were trained for each type of feature,and the results of the three classifiers were combined at the decision level to distinguish lung nodules.Compared with the four most advanced nodule classification methods on LIDC-IDRI dataset,the experimental results show that this method can improve the performance of pulmonary nodule classification by combining three-dimensional depth features with visual features in decision-making level.