Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (24): 239-246.DOI: 10.3778/j.issn.1002-8331.2107-0107

• Graphics and Image Processing • Previous Articles     Next Articles

U-Shaped Lung Nodule Segmentation Network with Bidirectional Enhancement Feature Structure

HUANG Xin, GUO Xiaomin   

  1. 1.College of Electronic Engineering and Automation, Guilin University of Electronic Technology, Guilin, Guangxi 541004, China
    2.Guangxi Key Laboratory of Automatic Testing Technology and Instruments, Guilin, Guangxi 541004, China
  • Online:2022-12-15 Published:2022-12-15

具有双向增强特征结构的U型肺结节分割网络

黄新,郭晓敏   

  1. 1.桂林电子科技大学 电子工程与自动化学院,广西 桂林 541004
    2.广西自动检测技术与仪器重点实验室,广西 桂林 541004

Abstract: Accurate and effective segmentation of lung nodules in CT images is the key to early diagnosis of lung cancer. However, the diversity of lung nodules and the complexity of the surrounding environment have brought huge challenges to the robustness of lung nodule segmentation. To improve the accuracy of lung nodule segmentation in CT images, Bi EFP-UNet lung nodule segmentation network is proposed. This structure uses an end-to-end deep learning method to solve the segmentation task of lung nodules. It integrates a bidirectional enhanced feature pyramid network between the encoder and decoder structures of the original U-Net network, strengthens the transmission and utilization of features, uses the Mish activation function to improve the segmentation efficiency, and eliminate the problem of the disappearance of the original U-Net network gradient. The experimental results on the lung nodule public data set LUNA16 show that the Dice similarity coefficient (DSC) of the Bi EFP-UNet network can reach 88.32%. Among them, the improvement brought by the Bi EFPN structure is 5.25 percentage points, and the Mish activation function brings the increase is 1.21 percentage points. Compared with the original U-Net network, the DSC of the Bi EFP-UNet network is increased by 6.46 percentage points, which can effectively solve the problem of poor segmentation performance for small target nodules and the disappearance of gradients of the original U-Net network.

Key words: CT, lung nodule segmentation, U-Net, Bi EFP-UNet, bidirectional enhanced feature pyramid network, Mish

摘要: 在CT影像中精准而有效地分割出肺部结节是肺癌早期诊断的关键。然而,肺结节形态的多样性以及周围环境的复杂性,都给肺结节分割的鲁棒性带来了巨大的挑战。为提高CT影像中肺结节分割的准确性,提出了Bi EFP-UNet(bidirectional enhanced feature pyramid UNet)肺结节分割网络。该结构采用端到端的深度学习方法来解决肺结节的分割任务,通过在原始U-Net网络的编码器和解码器结构之间集成一个双向增强型特征金字塔网络(bidirectional enhanced feature pyramid network,Bi EFPN),加强网络对特征的传递与利用;利用Mish激活函数提高分割效率,并消除原始U-Net网络梯度消失的问题。在肺结节公开数据集LUNA16上的实验结果表明,Bi EFP-UNet网络的Dice相似系数(DSC)可达88.32%,其中,Bi EFPN结构带来的提升为5.25个百分点,Mish激活函数带来的提升为1.21个百分点;与原始U-Net网络相比,Bi EFP-UNet网络的DSC提升了6.46个百分点,能有效解决原始U-Net网络对小目标结节分割性能差、梯度消失的问题。

关键词: CT, 肺结节分割, U-Net, Bi EFP-UNet, 双向增强型特征金字塔网络, Mish