基于改进U-Net++的CT影像肺结节分割算法  

Segmentation of Lung Nodules in CT Images Using Improved U-Net++

作　　者：黄鸿[1] 吕容飞 陶俊利 李远 张久权 HUANG Hong;LÜ Rongfei;TAO Junli;LI Yuan;ZHANG Jiuquan(Key Laboratory of Optoelectronic Technique System of the Ministry of Education,Chongqing University,Chongqing 400044,China;Department of Radiology,Chongqing University Cancer Hospital&Chongqing Cancer Institute&Chongqing Cancer Hospital,Chongqing 400030,China)

机构地区：[1]重庆大学光电技术与系统教育部重点实验室,重庆400044 [2]重庆大学附属肿瘤医院影像科,重庆400030

出　　处：《光子学报》2021年第2期65-75,共11页Acta Photonica Sinica

基　　金：国家自然科学基金(No.42071302);重庆市基础与前沿研究计划(No.cstc2018jcyjAX0093);重庆市留学人员回国创业创新支持计划(No.cx2019144);重庆市科卫联合项目医学科研项目(No.2019ZDXM007);2019年度中央高校基本科研业务费“医工融合项目”(No.2019CDYGYB008)。

摘　　要：卷积神经网络的语义分割模型未有效利用特征权重信息,导致在医学图像复杂场景中分割边界出现欠分割现象。针对该问题,基于融合自适应加权聚合策略提出一种改进的U-Net++网络,并将其应用于电子计算机断层扫描影像肺结节分割。该模型首先在卷积神经网络中提取出不同深度特征语义级别的信息,再结合权重聚合模块,自适应地学习各层特征的权重,然后将学习得到的权重加载到各个特征层上采样得到的分割图以得到最终的分割结果。在LIDC数据集和重庆大学附属肿瘤医院肺部电子计算机断层扫描数据集上进行了分割实验,所提方法的交叉比在两个数据集上分别可达到80.59%和87.40%、骰子系数分别可达到88.23%和90.83%。相比U-Net和U-Net++方法,该算法有效提升了图像分割性能。本文方法能在肿瘤微小细节上实现精确分割,较好地解决了肺结节向周围浸润性生长时出现欠分割的问题。Convolutional neural network-based semantic segmentation models do not effectively explore feature weight information,which will result in under-segmentation of segmentation boundaries in complex scenes of computed tomography images.To address this problem,an improved U-Net++model is proposed to explore adaptive weighted aggregation strategy based on U-Net++,and the improved UNet++model is applied to the segmentation of lung nodules in computed tomography images.In the convolutional neural network phase,the information from the different levels of deep features is extracted and combined with the weighted aggregation module,and thus the weights of features in each layer are adaptively learned.Then the learned weights are loaded on each feature layer and obtained a sampled segmentation map,and the final segmentation result can be obtained.Segmentation experiments are carried out on the lung cancer data sets of LIDC and Chongqing University Cancer Hospital.The intersection over union of the proposed improved U-Net++method on two datasets reach 80.59%and 87.40%,and the DICE of this method on two datasets could reach 88.23%and 90.83%,respectively.Compared with UNet and U-Net++,the proposed algorithm significantly improves the segmentation performance of lung nodules in computed tomography images.The experimental results show that improved U-Net++achieves accurate segmentation on tiny details of tumors,and it bring beneifits to solve the problem of under segmentation when lung nodules grow invasively to the surrounding.

关 键 词：计算机图象处理 分割算法 权重聚合 肺结节 CT影像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]