基于多尺度特征融合的肺结节良恶性分类方法

1)河北工业大学河北省大数据计算重点实验室, 天津 300401; 2)河北工业大学人工智能与数据科学学院, 天津 300401; 3)河北工业大学电子信息工程学院, 天津 300401; 4)北华航天工业学院信息技术中心, 河北廊坊 065000

人工智能; 深度学习; 特征提取; 特征融合; 通道注意力; 肺结节分类; SE-ResNeXt; 电子计算机断层扫描图像

Classification of benign and malignant pulmonary nodules based on multi-scale feature fusion
GU Junhua1, 2, SUN Zheran1, 3, WANG Feng1, 2, QI Yongjun1, 4, and ZHANG Yajuan1, 2

1)Hebei Province Key Laboratory of Big Data Computing, Hebei University of Technology, Tianjin 300401, P.R.China2)School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, P.R.China3)School of Electronics and Information Engineering, Hebei University of Technology, Tianjin 300401, P.R.China4)Information Technology Center North China Institute of Aerospace Engineering, Langfang 065000, Hebei Province, P.R.China

artificial intelligence; deep learning; feature extraction; feature fusion; channel attention; pulmonary nodule classification; SE-ResNeXt; computed tomography(CT)image

DOI: 10.3724/SP.J.1249.2020.04417

备注

为解决肺结节分类问题中肺电子计算机断层扫描(computed tomography, CT)图像特征提取不全面和随着卷积网络深度的加深易导致的梯度消失问题,提出一种基于多尺度特征融合网络(multi-scale feature fusion network, MSFFNet)的肺结节良恶性自动分类模型.使用多尺度卷积操作对输入的肺结节CT图像分别进行不同范围的特征提取和特征的融合拼接,解决特征提取不全面的问题; 引入SE-ResNeXt模块,充分利用通道注意力机制,有效解决特征信息丢失的问题; 输出肺结节良恶性的分类结果.在大型公开可用的肺图像联合数据库(lung image database consortium, LIDC-IDRI)上进行实验,MSFFNet模型的分类准确率达97.2%,特异性和敏感性分别为96.14%和98.62%,优于SE-ResNeXt等方法的分类效果.

In order to solve the problem of incomplete feature extraction in lung computed tomography(CT)image and the gradient disappearance caused by network degradation with the deepening of convolution network, an automatic classification model of pulmonary nodule malignancy based on multi-scale feature fusion network(MSFFNet)is proposed. The multi-scale convolution operation is used to extract and fuse the features of different ranges of input CT images. The SE-ResNeXt module is introduced to make full use of the channel attention mechanism to effectively solve the problem of feature information loss. Finally, the classification results of benign and malignant pulmonary nodules are obtained. The accuracy of MSFFNet classification model achieves 97.2%, and the specificity and sensitivity achieve 96.14% and 98.62%, respectively, which are better than those of SE-ResNeXt.

·