基于双通道并行网络的肺结节良恶性分类  

Classification of Benign and Malignant Pulmonary Nodules Based on Dual Channel Parallel Network

作　　者：罗益 李唯嘉 王梦瑶 曹秒[1] 李豪杰 LUO Yi;LI Weijia;WANG Mengyao;CAO Miao;LI Haojie(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学生命科学技术学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第4期130-135,共6页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技厅项目(20240401073YY)。

摘　　要：针对传统的肺结节良恶性分类方法中特征提取能力不足,提出一种结合残差网络和Swin-Transformer的双通道并行网络模型,并在特征融合处引入三重注意力机制有效提高肺结节良恶性分类的精度。该网络通过原始肺结节CT图像以及肺结节轮廓图像的辅助来提高分类精度。构建基于三重注意力的特征融合模块以连接两个网络,有利于发掘更多的肺结节图像特征。方法在LIDC-IDRI数据集上进行验证,AUC达到0.951 7,准确率达0.931 1。实验结果证明ResNet-Swin Transformer的分类精度相比ResNet和Swin Transformer等更高,可以辅助医生提高肺结节诊断率。In view of the lack of feature extraction ability in traditional classification methods for benign and malignant pulmonary nodules,a dual-channel parallel network model combining residual network and Swin-Transformer is proposed,and the accuracy of benign and malignant pulmonary nodules classification is effectively improved by adding a triple attention module.The network improves the classification accuracy with the help of the original CT image and the contour image of pulmonary nodules.Adding triple attention to connect the two networks is beneficial to explore the benign and malignant characteristics of pulmonary nodules.The proposed method was verified on the LIDC-IDRI dataset,and the AUC reached0.951 7,with an accurate rate reached 0.931 1.The experimental results show that the classification accuracy of ResNetSwin Transformer is higher than that of ResNet and Swin Transformer.The diagnostic rate of pulmonary nodules can be improved.

关 键 词：良恶性分类 ResNet Swin-Transformer 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]