A 3D Multi-scale Virtual Adversarial Network for False Positive Reduction in Pulmonary Nodule Detection

Abstract

Lung cancer is a high-mortality disease. Using Computer-based Automatic Detection (CADe) systems can detect pulmonary nodules to assist early diagnosis. An effective CADe system based on deep learning is often trained with sufficient labeled samples. However, radiologist cannot mark thousands of medical data. In this paper, we proposed a model based on 3D semi-supervised pulmonary nodule detection and false positive reduction. The framework consists of two stage: (1) Candidate proposals of pulmonary nodule; (2) False positive reduction. In the first stage, we employed a 3D Region Proposal Network (RPN) to effectively proposed candidate regions for pulmonary nodules. In the second stage, we proposed a 3D multi-scale semi-supervised Virtual Adversarial Training (VAT) network which make use of unlabeled data and utilize contextual information to reduce the false positive rate with more reliability. Experimental results on the public Lung Nodule Analysis 2016 (LUNA16) dataset showed that our model can perform pulmonary nodules detection and false positive reduction under the circumstance of insufficient labeled samples.