Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (13): 195-203.DOI: 10.3778/j.issn.1002-8331.2011-0211

• Graphics and Image Processing • Previous Articles     Next Articles

Improved U-Shaped Residual Network for Lung Nodule Detection

YUAN Jinli, ZHAO Linlin, GUO Zhitao, SU Yi, LU Chenggang   

  1. College of Electronic Information Engineering, Hebei University of Technology, Tianjin 300401, China
  • Online:2022-07-01 Published:2022-07-01

改进U型残差网络用于肺结节检测

袁金丽,赵琳琳,郭志涛,苏逸,卢成钢   

  1. 河北工业大学 电子信息工程学院,天津 300401

Abstract: Aiming at the low sensitivity of lung nodule detection in computed tomography(CT) images, and there are a large number of false positives, an improved U-shaped residual network is proposed for lung nodule detection. Firstly, the U-shaped structure of U-net network and residual learning method are used to construct deep-seated network, At the same time, self-calibrated convolutions  is introduced to enhance the information extraction ability of features and enhance the inter channel and local information, which is conducive to the detection of different shapes of nodules. Secondly, through the introduction of channel attention mechanism, the features in the feature extraction process are recalibrated to realize adaptive learning feature weight. In addition, DR loss is introduced as the classification loss function of this algorithm to solve the imbalance problem of positive and negative samples. The proposed algorithm is validated in LUNA16 data set, and the CPM score reaches 0.901, which improves the sensitivity of pulmonary nodule detection, and effectively reduces the average number of false positive results, which can effectively assist radiologists to detect pulmonary nodules.

Key words: image processing, pulmonary nodule detection, self-calibrated convolution, attention mechanism, feature extraction

摘要: 针对计算机断层扫描(CT)影像中肺结节检测灵敏度较低,且存在大量假阳性的问题,提出一种改进的U型残差网络用于肺结节检测。采取U-net网络的U型结构并利用残差学习方式构建深层次网络,同时引入自校正卷积增加特征的信息提取能力,进行通道间与局部信息增强,有利于检测不同形态的结节;通过引入的通道注意力机制,对特征提取过程中的特征进行重标定,实现自适应学习特征权重,进一步提高检测的准确率;引入DR loss作为该算法的分类损失函数,用于解决数据正负样本失衡问题。在LUNA16数据集对所提算法进行了验证,CPM得分达到0.901,提高了肺结节检测的灵敏度,而且有效降低了检测结果的平均假阳性个数,可有效辅助放射科医师对肺结节进行检测。

关键词: 图像处理, 肺结节检测, 自校正卷积, 注意力机制, 特征提取