Study On The Classification Of Benign And Malignant Pulmonary Nodules Based On Adaptive Gabor Filter And Deep Belief Network

The wide application of low-dose computed tomography(LDCT)in the early screening of primary lung cancer cause extremely large amounts of pulmonary CT images,which brings a huge workload for radiologists.In addition,LDCT has the characteristics of large number of images and low resolution,which leads to high missed diagnosis rate and misdiagnosis rate of early lung cancer screening.In response to this situation,people have developed the computer-aided diagnosis system(CAD)by using the advantages of computer in batch calculation and quantitative analysis to help clinicians to distinguish benign and malignant pulmonary nodules.With the development of CAD system,the study of geometric characteristics of pulmonary nodules has become more and more mature.When dealing with large data samples,the classification model constructed by traditional machine learning algorithm has high computational cost and poor performance of the model because of the simple structure.In this paper,we propose to extract the texture features of pulmonary nodules in frequency domain and construct a classification model using deep belief network to study the guiding significance of texture changes of pulmonary nodules for benign and malignant classification.Firstly,to fix the problem that the traditional Gabor filter with the fixed window method or uniform sampling method causes the difference of flexibility and data redundancy when extracts the texture features of pulmonary nodules.The adaptive Gabor filter which the size of filter window varies with the frequency is proposed in the study;Then,on the question that in order to reduce the model over-fitting phenomenon and improve the generalization ability of the model,it is necessary to reduce the dimension of the feature vector when building the classitication models by using the traditional machine learning algorithms,that causes the high time cost and the cumbersome process.The study uses the deep belief network with the structural function of RBM which has unsupervised nonlinear fitting data and the strong nonlinear mapping ability of complex data to establish the classification model.Finally,comparing with the traditional method of extracting texture features of pulmonary nodules in spatial domain and combining machine learning algorithm to construct classification model,the method of this study is used to research the guidance significance of texture change on pulmonary nodules in the classification of benign and malignant tumors.The accuracy of model detection is higher,the mis-diagnosis rate,the missed diagnosis rate and the time cost lower substantially.