基于多尺度特征互补和聚合约束的肺结节分类方法

Pulmonary Nodule Classification Network Based on Multi-Scale Feature Complementation and Aggregate Constraint

肺结节分类问题是早期肺癌检测与诊断的重要问题之一,针对现有的肺结节分类方法存在多尺度特征融合的信息冗余和缺乏判别性特征表示等问题,提出了一个基于多尺度特征互补与聚合约束(Multi-scale Feature Complementation and Aggregate Constraint, MFCAC)的肺结节分类方法,并提出了多尺度特征互补模块用于学习相邻尺度特征的差异信息,从而避免特征融合过程中的信息冗余;同时在网络特征层引入了聚合约束损失,实现对同类特征的聚集,提高网络判别性特征表示能力;将两个模块融入在编码器-解码器架构中形成MFCAC,共同作用实现高效分类。本文在LIDC-IDRI数据集上进行了对比实验,并通过消融实验分析了该方法中各组成部分的贡献和影响,结果表明,相较于对比算法,MFCAC在肺结节分类上具有更优的性能。

The classification of pulmonary nodule is one of the important issues in early detection and diagnosis of lung cancer.To address the problem of information redundancy in multi-scale feature fusion and lack of discriminative feature representation in existing lung nodule classification methods,a multi-scale feature complementation and aggregate constraint(MFCAC)pulmonary nodule classification network is proposed.A multi-scale feature complementation module is proposed to learn the difference information of adjacent scale features,thereby avoiding information redundancy in the feature fusion process.Meanwhile,aggregate constraint loss is introduced into the network feature layer to achieve aggregation of similar features and improve the discriminative feature representation ability of the network.The two modules are integrated into the encoder-decoder architecture to form MFCAC,which can achieve efficient classification.Comparative experiments are conducted on the LIDC-IDRI dataset,and ablation experiments are used to analyze the contributions and effects of each component in this method.The results show that MFCAC has better performance in lung nodule classification compared to the compared algorithms.