计算机工程 ›› 2019, Vol. 45 ›› Issue (4): 254-261.doi: 10.19678/j.issn.1000-3428.0051769

• 图形图像处理 • 上一篇    下一篇

基于多尺度特征结构的U-Net肺结节检测算法

朱辉,秦品乐   

  1. 中北大学 大数据学院,太原 030051
  • 收稿日期:2018-06-07 出版日期:2019-04-15 发布日期:2019-04-15
  • 作者简介:朱辉(1993—),男,硕士研究生,主研方向为医学图像处理、深度学习;秦品乐,副教授、博士。
  • 基金项目:

    山西省自然科学基金(2015011045)。

U-Net Pulmonary Nodule Detection Algorithm Based on Multi-scale Feature Structure

ZHU Hui,QIN Pinle   

  1. School of Data Science and Technlogy,North University of China,Taiyuan 030051,China
  • Received:2018-06-07 Online:2019-04-15 Published:2019-04-15

摘要:

针对肺结节低层特征在网络传输过程中的缺失问题,基于多尺度特征结构,提出一种改进的U-Net卷积神经网络肺结节检测算法。采用卷积操作与池化操作获取高层特征,通过密集网络使得特征信息在输入层和输出层之间高速流通,并结合扩张卷积生成多尺度特征,提高肺结节低层特征的利用率。实验结果表明,与传统U-Net卷积神经网络的肺结节检测算法相比,改进算法对于小型结节的检测准确率约提高20%,可实现更准确的肺部病灶区域定位。

关键词: 小目标检测, 卷积神经网络, 深度学习, 密集网络, 肺结节

Abstract:

Aiming at the defect problem in the low-level characteristics of pulmonary nodules in the process of network transmission,an improved U-Net convolution neural network algorithm based on multi-scale feature structure for pulmonary nodule detection is proposed.High-level features are acquired by convolution and pooling operations,high-speed flow of feature information between input and output layers is achieved by using dense network,and multi-scale features are generated combining with expanded convolution to improve the utilization of low-level features of pulmonary nodules.Experimental results show that,compared with the traditional U-Net convolution neural network detection algorithm of pulmonary nodules,the improved algorithm improves the detection accuracy of small nodules by nearly 20%,and achieves more accurate localization of pulmonary lesions.

Key words: small object detection, Convolutional Neural Network(CNN), deep learning, dense network, pulmonary nodule

中图分类号: