A Pilot Research On CAD For Pulmonary Nodules In CT Images

At present,Computer Tomography(CT)is considered to be the most effective means for detecting lung diseases,and has important guiding significance in the clinical diagnosis and treatment.Pulmonary nodules is CT imaging manifestation of most of lung diseases,including lung cancer.Physicians give a presumptive diagnosis about benign and malignant nodules by reading CT images.The malignant nodules are lung cancer.But with the development of CT imaging technology,more and more images data need to be comprehend and diagnosed which increases the workload of physicians greatly and affects the quality of the diagnosis directly.Therefore,the computer-aided diagnosis system(CAD)is necessary and important to reduce the missed detection and misdiagnosis caused by work fatigue and subjective emotion.The CAD includes two aspects: computer-aided detection(CADe)and computer-aided diagnosis(CADx).This topic research focus on the CADx methodology of pulmonary nodules,and the main work are as follows:1?Designed two classifiers based on feature parameters of pulmonary nodules: BP neural network(BPNN)and support vector machine(SVM).And on the basis of SVM,we proposed an ensemble learning method for pulmonary nodules classification based on Bootstrap-heterogeneous SVM to improve the performance about diagnosis of pulmonary nodules under the unbalanced dataset.2?To overcome restrictions in methods of shallow machine learning and improve auxiliary diagnostic performance about benign and malignant pulmonary nodules in CT images,convolutional ceural cetwork(CNN)was carried on.And on the basis of single input CNN,we designed a CNN network structure called 3D-CNN,which combined Local Jet transformation and CNN for further improving of diagnostic performance.3?In order to ensure the accuracy of pulmonary nodules used in the extraction of feature parameters and CNN classifiers,remove the interference of false positive nodules,we proposed a semi-automatic CADe method based on fuzzy C-means clustering method(FCM)to achieve a precise segmentation of pulmonary nodules.4?On the basis of traditional intensity feature,shape feature and texture feature of pulmonary nodules,we also proposed a extraction method of texture feature parameters based on Local Jet transformation space.Next,kernel principal component analysis method was used to dimensional-reduce and optimize the extracted 259 parameters.The optimized parameters were used as dataset of BPNN?SVM and the ensemble learning based on Bootstrap-heterogeneous SVM.5?Transforming pulmonary nodule images into Local Jet component images.Then the original pulmonary images and all the Local Jet component images were normalized to the same size of 32×32.The former was used as the input data of CNN and the latter was used as the input data of 3D-CNN.In this study,we have 1305 pieces of pulmonary nodule images,including 299 pieces of benignancy and 1006 pieces of malignancy,of 164 patients as the sample database.With the ratio of about 3:1,the samples were divided into training set and test set.The highest classification accuracy of BPNN?SVM?ensemble learning method based on Bootstrap-heterogeneous SVM and 3D-CNN are 76.49%?80.06%?84.23% and 85.07% respectively,the value of auc that represents the comprehensive performance of classifiers are 0.75?0.62?0.76 and 0.78 respectively.The experimental results show that the classification performance of proposed Bootstrap-heterogeneous SVM based on feature parameters and the 3D-CNN based on pulmonary nodules images are superior to previous traditional classifiers.So,we could hold the opinion that CADx based on advanced machine learning method can effectively make a judgment about benign and malignant pulmonary nodules,provide physicians with a reliable and objective reference opinion and reduce the misdiagnosis rate caused by subjective emotion.