Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (24): 185-195.DOI: 10.3778/j.issn.1002-8331.2207-0027

• Graphics and Image Processing • Previous Articles     Next Articles

Lung Nodule Segmentation Model with Enhanced Edge Features

CHENG Zhaoxue, LI Yang, ZHOU Yan, LU Huimin   

  1. School of Computer Science and Engineering, Changchun University of Technology, Changchun 130102, China
  • Online:2023-12-15 Published:2023-12-15

增强边缘特征的肺结节分割模型

程照雪,李阳,周妍,鲁慧民   

  1. 长春工业大学 计算机科学与工程学院,长春 130102

Abstract: Aiming at the problem of low segmentation accuracy caused by blurred nodule edge pixels in lung nodule segmentation, a symmetric dual-branch structure segmentation model GEU-Net (Gabor edge U-Net) is proposed to enhance edge features. The main branch uses U-Net to capture the overall visual features of nodules as well as global contextual information, and the edge branch designs a Gabor convolutional module to extract nodule edge texture features, and fuses the features extracted by the main branch and the edge branch encoders through skip connections. In addition, a hybrid loss function that combines cross-entropy and Focal Loss is proposed to solve the problem of unbalanced positive and negative samples in the training process of the two branches. Nodule segmentation experiments are carried out on the public dataset LIDC-IDRI and LNDb, the experimental results show that the Dice coefficients of the GEU-Net segmentation model on the two datasets are 92.79% and 86.78%, respectively, the average intersection ratios are 87.53% and 79.09%, the recall rates are 94.35% and 87.43%, compared with baseline algorithm, the segmentation performance has been improved to a certain extent.

Key words: lung nodules segmentation, U-Net, Gabor convolution, edge feature enhancement

摘要: 针对在肺结节分割过程中结节边缘像素点模糊导致模型分割精度低的问题,提出一种用于增强边缘特征的对称双分支结构分割模型GEU-Net(Gabor edge U-Net)。其主干分支使用U-Net来捕获结节整体视觉特征以及全局上下文信息,边缘分支则设计Gabor卷积模块提取结节边缘纹理特征,并通过跳跃连接对主干分支与边缘分支编码器提取到的特征进行融合。此外,提出一种融合交叉熵和Focal Loss的混合损失函数,用于解决两个分支在训练过程中存在的正负类样本不均衡问题。在公开数据集LIDC-IDRI以及LNDb上进行了结节分割实验,实验结果表明,GEU-Net分割模型在两个数据集上的Dice系数分别为92.79%和86.78%、平均交并比分别为87.53%和79.09%、召回率则分别为94.35%和87.43%,与基线算法相比分割性能得到了一定程度的改善。

关键词: 肺结节分割, U-Net, Gabor卷积, 边缘特征增强