Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (14): 121-125.DOI: 10.3778/j.issn.1002-8331.2102-0209

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Research on Classification of Pulmonary Nodules by Three-Dimensional Multi-Scale Cross Fusion Network

YANG Jianli, ZHU Dejiang, SHAO Jiajun, LIU Xiuling   

  1. 1.College of Electronic and Information Engineering, Hebei University, Baoding, Hebei 071002, China
    2.Key Laboratory of Digital Medical Engineering of Hebei Province, Baoding, Hebei 071002, China
  • Online:2022-07-15 Published:2022-07-15

三维多尺度交叉融合网络肺结节分类研究

杨建利,朱德江,邵嘉俊,刘秀玲   

  1. 1.河北大学 电子信息工程学院,河北 保定 071002
    2.河北省数字医疗工程重点实验室,河北 保定 071002

Abstract: Accurate classification of benign and malignant pulmonary nodules on chest CT images is vital to the prevention and treatment of lung cancer. However, it is a challenging task due to the complex background of chest CT images and the unclear delineation of benign and malignant pulmonary nodules. A deep three-dimensional multi-scale cross fusion convolutional neural network is proposed to realize the accurate classification of benign and malignant pulmonary nodules. It uses dense connection structure to automatically extract multi-scale features of lung nodules. In order to reduce the loss of lung nodule-related information in the feature extraction process, a cross-fusion strategy is introduced for multi-scale features to obtain multi-scale feature groups, which enhances high and low-level semantic information expressing ability, and at the same time enhance the transmission and transfer of features in the network. This article uses the lung image federation data set for verification, the classification accuracy rate reaches 90.96%, and the AUC is 94.95%.

Key words: pulmonary nodule classification, dense connection network, multiscale cross fusion

摘要: 计算机断层扫描影像中良、恶性肺结节的准确分类对肺癌的预防和治疗至关重要。然而,由于计算机断层扫描影像中肺结节背景的复杂性,以及良、恶性肺结节判定之间存在的不确定性,使得良恶性肺结节的准确分类成为了一项极具挑战性的工作。提出了一种深度三维多尺度交叉融合卷积神经网络实现了良恶性肺结节的精确分类。使用密集连接结构自动提取肺结节多尺度特征,为了减少特征提取过程中肺结节相关信息的丢失,对多尺度特征引入了交叉融合策略得到多尺度特征组,增强了高、低层次语义信息的表达能力,同时增强特征在网络中的传递和转移。将提取的特征组分别连接至多个softmax分类器,模拟多位经验不同医生共同决策,实现了良、恶性肺结节的精确识别。使用肺图像联合会数据集进行验证,分类准确率达到了90.96%,AUC为94.95%。

关键词: 肺结节分类, 密集连接网络, 多尺度交叉融合