USING DEEP LEARNING NETWORKS FOR CLASSIFICATION OF LUNG CANCER NODULES IN CT IMAGES

Abstract

Purpose: One of the foremost common cancers around the world is lung cancer (LC) which evaluation of its incidence very important for more robust planning. Computerized tomography (CT) is important for the diagnosis of lung nodules in carcinoma. Recently, algorithms like deep learning have been considered as a promising method within all medical field, therefore, we try to using various deep learning networks for classification of lung cancer nodules in CT images. Methods: In this paper, open-source datasets, and multicenter datasets are used. Three CNN architectures (VGG16, VGG19, and Inceptionv3) were designed to detection lung nodules and classified them into two malignant or benign groups based on their pathologically and laboratory results. Results: The accuracy of these three CNN architectures in 10-fold training model were found to be 98.3%, 99.6%, and 99.5%, respectively. There was no difference in term of sensitivity and specificity between larger and smaller nodules. The model validation was checked by manually assessments of CT by doctors and compared with three-dimensional CNN results. The performance of the CNN model was better and accurate than manual assessment. Conclusion: Results showed that, of the CNN architectures, The VGG19 with an accuracy of 99.6% has the best performance among the three networks.