Lung Nodule Detection via 3D U-Net and Contextual Convolutional Neural Network

EasyChair Preprint 463

Abstract

Lung cancer is the most common cause of cancer death worldwide. Lung nodule detection based on CT image is the most prevailing method for detecting lung cancer. In this paper, we propose a patch-based 3D U-Net and contextual convolutional neural network (CNN) to automatically segment and classify lung nodule and help the radiologists read CT images. Typically, lung nodule detection task could be divided into three stages, including lung segmentation, nodule detection or segmentation and false positive reduction. In lung segmentation stage, we use morphological methods to segment pulmonary parenchyma from raw CT images. To segment lung nodule, 3D U-Net is employed to extract suspicious nodule from preprocessed CT images. In order improve model accuracy, we use Generative Adversarial Network (GAN) to boost model training. To further enhance model performance, we use online sampling strategy to augment data and use 3D contextual CNN with Inception blocks to determine whether the volume is malignant nodule or not. Experimental results demonstrate that the proposed method could effectively detect the cancerous nodule from the CT scans.

Keyphrases: 3D Inception CNN, 3D U-net, Generative Adversarial Network, Lung nodule detection

@booklet{EasyChair:463,
  author    = {Chen Zhao and Jungang Han and Yang Jia and Fan Gou},
  title     = {Lung Nodule Detection via 3D U-Net and Contextual Convolutional Neural Network},
  doi       = {10.29007/bgkd},
  howpublished = {EasyChair Preprint 463},
  year      = {EasyChair, 2018}}