基于多维度融合的肺结节分类算法  

Lung nodule classification algorithm based on multi-dimensional fusion

作　　者：堵红群 李岳阳 崔方正 罗海驰[3] 顾中轩 DU Hongqun;LI Yueyang;CUI Fangzheng;LUO Haichi;GU Zhongxuan(Department of Medical Imaging,Affiliated Hospital of Jiangnan University,Wuxi 214122,China;School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi 214122,China;School of Internet of Things Engineering,Jiangnan University,Wuxi214122,China)

机构地区：[1]江南大学附属医院影像科,江苏无锡214122 [2]江南大学人工智能与计算机学院,江苏无锡214122 [3]江南大学物联网工程学院,江苏无锡214122

出　　处：《中国医学物理学杂志》2024年第11期1428-1436,共9页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(U1836218)。

摘　　要：采用多维度模型融合的方法,提出一种肺结节分类算法。在肺结节假阳性减少算法基础上进行优化,在多尺度特征融合模块得到特征之后引入高层特征增强软激活映射模块,以增强模型的分类能力;针对实际分类过程中各类结节数据不平衡的问题,引入平衡均方差损失来改进模型的训练效果;采用三维和二维模型融合方式进一步提升模型分类性能。在Private Lung数据集上进行的实验证明本研究提出的模型分类准确度达到93.8%,优于现有方法。A novel algorithm based on multi-dimensional fusion is proposed for classifying lung nodules.Based on the algorithm for reducing false positives of pulmonary nodules,the optimization is carried out by introducing a high-level feature enhancement soft activation mapping module after obtaining features by the multi-scale feature fusion module to improve the classification ability.To address the imbalance of different nodule data in the actual classification,a balanced mean square error loss is adopted to improve the training effect of the model.A three-dimensional and two-dimensional model fusion method is used to further improve the classification performance.The experiment conducted on a Private Lung dataset proves that the proposed model has a classification accuracy of 93.8%,outperforming the existing methods.

关 键 词：肺结节 计算辅助诊断 多尺度特征融合 软激活映射 平衡均方差损失 

分 类 号：R318[医药卫生—生物医学工程] TP391.4[医药卫生—基础医学]