基于Transformer和CNN交错混合的肺结节分割网络  

Interlace mixed net of lung nodule segmentation based on Transformer and CNN

作　　者：吴骏[1,2] 侯宪哲 王健 肖志涛[2] 王雯[2,4] WU Jun;HOU Xianzhe;WANG Jian;XIAO Zhitao;WANG Wen(School of Electronics and Information Engineering,Tiangong University,Tianjin 300387,China;Tianjin Key Laboratory of Optoelectronic Detection Technology and Systems,Tiangong University,Tianjin 300387,China;Department of Electronic Teaching and Research,93756 Units,Tianjin 300131,China;School of Life Sciences,Tiangong University,Tianjin 300387,China)

机构地区：[1]天津工业大学电子与信息工程学院,天津300387 [2]天津工业大学天津市光电检测技术与系统重点实验室,天津300387 [3]93756部队教研部电子教研室,天津300131 [4]天津工业大学生命科学学院,天津300387

出　　处：《天津工业大学学报》2024年第1期74-81,共8页Journal of Tiangong University

基　　金：天津市自然科学基金资助项目(21JCZXJC00170);京津冀基础研究合作专项(H2021202008)。

摘　　要：针对肺结节尺寸多样、形状异质化高等问题,提出基于Transformer和卷积神经网络(CNN)交错混合(IMTC)的肺结节分割网络,该网络是一个对称的层次连接网络,具有很强的多尺度特征提取能力。该网络通过集成2种方案分别解决肺结节多尺寸与形状异质化问题:(1)采用感知注意力模块(inception attention module,IAM),通过并联多个不同大小的卷积核来增加浅层网络的感受野组合,以此捕获更为丰富的浅层特征;(2)为获取更具表示能力的高级语义特征,利用由Transformer和CNN组成的基本骨干网络交错提取结节特征,使得全局特征与局部特征充分融合,从而提高结节特征表示的泛化能力和鲁棒性。实验结果表明:本文模型可以准确分割直径较小以及边缘复杂的肺结节,在LUNA16公开数据集上分割性能良好,Dice和IOU分别达到86.15%和76.10%。Aiming at the problems of multi-size and high heterogeneity of lung nodules,an interlace mixed network based Transformer and convolutional neural network interlace mixed(IMTC) is proposed.The network is a symmetrical hierarchical connection network with strong multi-scale feature extraction capabilities.It integrates two new schemes to solves the promblems of multi-size and shape heterogeneity.(1) Inception attention module(IAM) is proposed to capture richer shallow features by paralleling multiple convolution kernels of different sizes to increase the combination of receptive fields.(2) In order to extract deeper semantic features with more expressive ability,the basic backbone network composed of Transformer and CNN is used to extract nodule features alternately,so that the global features and local features are fully integrated,and then the generalization ability and robustness of nodule feature representation are improved.The experimental results show that the model in this paper can accurately segment nodules with small scale and complex margin,and has good segmentation performance on the LUNA16 public dataset,and the Dice and IOU reach 86.15% and 76.10%,respectively.

关 键 词：肺结节 TRANSFORMER 卷积神经网络(CNN) 感知注意力模块(IAM) 交错混合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]