融合长距离信道注意力与病理特征的肺结节分类

Classification of Pulmonary Nodule by Combining Long-Distance Channel Attention and Pathological Feature

针对现有深度学习网络普遍存在的长距离特征通道关联性缺失、网络自提取特征会湮没肺结节病理显性特征等问题,首先,将通道注意力和空间注意力结合,提出一种可以有效建立长距离特征通道关联性的注意力模块LCA(long-distance channel attention),让模型能获取肺部CT图像的全局显著特征,提高对肺结节的良恶性诊断精度.其次,将肺结节直径、纹理、钙化度等病理显性特征与其深度特征融合,以增强这些显性特征的重要度,提高模型的分类效果.最后,搭建一种特征提取网络DLCANet(dual-connected long-distance channel attention network)和一种分类器模型MARTM(multiple additive regression tree model).在数据集LIDC-LDRI和LUNA16上进行分类实验,与基准模型DPN(dual path network)相比,准确率提高了3.63%,假阳性率下降了8.66%,且整体性能优于目前主流模型.

Aiming at the common problems of existing deep learning networks,such as lack of correlation of long-distance feature channels and network self-extraction features obliterating the dominant pathological features of pulmonary nodules.Firstly,by combining channel attention and spatial attention,an attention module LCA(long-distance channel attention)that can effectively establish the correlation of long-distance feature channels is proposed.Secondly,the dominant pathological features such as pulmonary nodule diameter,texture and calcification are fused with their depth features to enhance the importance of these dominant features.Finally,a feature extraction network DLCANet(dual-connected long-distance channel attention network)and a classifier model MARTM(multiple additive regression tree model)are built.The classification experiments are carried out on the datasets LIDC-LDRI and LUNA16.Compared with the benchmark model DPN(dual path network),the accuracy rate is increased by 3.63%,the false positive rate is decreased by 8.66%,and the overall performance is better than those of current mainstream models.