Lung Nodule Diagnosis on 3D Computed Tomography Images Using Deep Convolutional Neural Networks

Abstract

This research studies a lung nodule classification problem using 3D computed tomography (CT) images for computer-aided diagnosis (CAD) system by applying deep convolutional neural networks (CNNs). Lung cancer is one of the most common cancers with a high mortality rate. There is an urgent need to diagnose lung nodules at an early stage to improve the survival rate of lung cancer patients. Through a CAD system, a complementary analysis result can be provided to radiologists to enhance the diagnosis process. The diagnosis performance of conventional CAD systems primarily relies on the results of previous analysis processes, such as nodule segmentation and feature extraction. In this research, the nodule segmentation and feature extraction in CAD system are eliminated by performing an end-to-end classification from raw 3D nodule CT patches using CNNs. Specifically, the state-of-artCNN models, e.g., VGG16, VGG19, ResNet50, DenseNet121, MobileNet, Xception, NASNetMobile, and NASNetLarge, are modified to 3D-CNN models and explored on a public CT lung dataset: Lung Image Database Consortium and image database resource initiative (LIDC-IDRI). Experimental results show that DenseNet121 and Xception achieve better results for lung nodule diagnosis regarding accuracy, sensitivity, specificity, precision, and area under the curve (AUC).