Classification model of benign and malignant nodules in lung CT images based on deep learning of residual network

Computer aided diagnosis of benign and malignant pulmonary nodules plays a significant role in timely treatment of lung cancer. Aiming at the current situation that the diagnosis accuracy of benign and malignant pulmonary nodules in computer aided diagnosis system is low and the misdiagnosis rate and wrong diagnosis rate are relatively high, a classification model of lung benign and malignant nodules based on residual network is proposed. Firstly, partial lung CT images (totally 10 402) from LIDC-IDRI are selected as a data set, the data are augmented by horizontal flipping the images and then the images are converted into single channel images. Afterwards, cropping and normalization processing is performed. Finally, the data are divided into training set and test set (7: 3), and used to train and test the designed residual network (ResNet-26). After training, the obtained test results are as follows: the accuracy rate of classification, sensitivity and specificity of benign and malignant pulmonary nodules are 97.53%, 97.91% and 97.18%, respectively; and the calculated AUC is 0.958. Comparison results show that in terms of accuracy, sensitivity and specificity specifications, the proposed method is superior to existing various other methods, the classification result can provide good assistance in diagnosis for the doctors. © 2020, Science Press. All right reserved.