欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 专论：第12届中国计算机图形学大会 (CHINAGRAPH 2018 广州) • 上一篇    下一篇

基于三维卷积神经网络的肺结节假阳性筛查

  

  1. 1. 浙江工业大学计算机科学与技术学院，浙江 杭州 310023； 
    2. 浙江大学计算机科学与技术学院，浙江 杭州 310058； 
    3. 浙江大学睿医人工智能研究中心，浙江 杭州 310000
  • 出版日期:2019-06-30 发布日期:2019-08-02
  • 基金资助:
    国家自然科学基金项目(61801428，61672453)；浙江省自然科学基金项目(LY18F020034)；浙江大学教育基金会项目(K18-511120-004、 K17-511120-017)；之江实验室重大科研项目(2018DG0ZX01)

False Positive Reduction of Pulmonary Nodules Using 3D CNN

  1. 1. School of Computer Science and Technology, Zhejiang University of Technology, Hangzhou Zhejiang 310023, China; 
    2. School of Computer Science and Technology, Zhejiang University, Hangzhou Zhejiang 310058, China;
    3. Real Doctor Artificial Intelligence Research Center, Zhejiang University, Hangzhou Zhejiang 310000, China
  • Online:2019-06-30 Published:2019-08-02

PDF (PC)

159

摘要： 从 CT 影像中检测肺结节在肺癌的早期诊断中至关重要，而肺结节假阳性的筛查 是提高肺结节检测准确度的重要一步。为了从大量候选结节中快速准确地区分出真正的肺结节， 设计了一个 3D 卷积神经网络(CNN)筛查肺结节假阳性。提出了网络模型，通过恒等映射和残差 单元来加速模型训练，采用单连接路径重复利用特征并重组新特征。基于该模型的肺结节假阳 性筛查方法，与基于 2D CNN 的方法相比，不仅可以省略数据切片步骤，而且能够充分利用 CT 影像的空间信息; 与其他基于 3D CNN 的方法相比，具有参数量小、模型训练快的优点。该方 法在 LUNA16 数据集中的假阳性筛查中取得了较高的敏感度。

关键词: 3D CNN, 肺结节, 假阳性筛查

Abstract: Pulmonary Nodule Detection is the most promising way in early detection of pulmonary cancer. False positive reduction is one of the most crucial steps for improving the accuracy in automatic pulmonary nodule detection. For quickly and accurately discriminate true nodules from a large number of candidates, a 3D convolutional neural networks (CNN) is proposed for false positive reduction. In the proposed network, identity mapping and residual unit are adopted to accelerate network training. At the same time, single connected path is explored to get new features. Compared with 2D CNN based pulmonary nodule detection methods, the proposed method based on the proposed 3D CNN can take full advantage of space structure of CT. Compared with other 3D CNN based pulmonary nodule detection methods, the proposed method has fewer parameters that can make the training very fast. Experimental results on the public LUNA16 dataset demonstrate superior performance of the proposed method.

Key words:  3D CNN, pulmonary nodules, false positive reduction