Method for Identifying Benign and Malignant Pulmonary Nodules Combing Deep Convolutional Neural Network and Hand-Crafted Features

Here, we present a method for identifying benign and malignant pulmonary nodules that combines convolutional neural network(CNN)learning features and conventional hand-crafted features. First, the pulmonary nodules area is segmented from computed tomography (CT) images, and traditional machine learning methods are used to extract the image features of the nodule area. Then, the CNN features of network learning are extracted, using the intercepted pulmonary nodules to train the 3D-Inception-ResNet model, and the 2 kinds of features are combined, the random forest (RF) model is used for feature selection. Finally, support vector machine (SVM) and RF classifier are used to identify benign and malignant pulmonary nodules. The 1036 pulmonary nodules in the LIDC-IDRI database are used for experimental verification. Classification accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve ( AUC) of the proposed method can reach 94. 980 0, 90.02%, 97.03%, and 97.43%, respectively. The proposed method can accurately distinguish benign and malignant lung nodules, more effectively than most existing mainstream methods, as shown by the experimental results.