A Transfer Learning Method for CT Image Classification of Pulmonary Nodules

Abstract

Listen

Translate article

A pulmonary nodule classification method of Computer Tomography (CT) images based on transfer learning of deep convolutional neural network (CNN) is proposed. Lung CT images with labels are quite limited compared with the large scale image database such as ImageNet. It is easy to produce over-fitting problem when using the limited data to train the deep CNN for classification task. In this paper, in order to overcome this difficulty, the deep CNNs GoogleNet and ResNet are pre-trained on the large scale database ImageNet. The fully connected layers and the classifiers of the pre-trained networks are replaced to complete the classification of CT images of pulmonary nodules. A sub set of the Lung Image Database Consortium image collection (LIDC-IDRI) is used to fine-tune the network and validate the classification accuracy. This is the process of transfer learning. It solves the problem of the deficiency of lung CT images as labeled training data for CNNs. By the knowledge obtained from the pre-trained CNNs which have been trained on ImageNet, the network is easier to converge and the training time is greatly reduced. The classification accuracy of Pulmonary Nodules can be reached up to 71.88% by using the proposed method.