基于改进的U-Net肺结节分割方法研究  

AN IMPROVED U-NET FOR LUNG NODULE SEGMENTATION

作　　者：苗语[1] 丰振航 杨华民[1] 蒋振刚[1] 师为礼[1] Miao Yu;Feng Zhenhang;Yang Huamin;Jiang Zhengang;Shi Weili(College of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,Jilin,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机应用与软件》2021年第12期213-219,共7页Computer Applications and Software

基　　金：吉林省科技发展计划项目(20170204031GX)。

摘　　要：由于肺部CT图像的特征信息复杂程度高,经典U型卷积网络对肺结节分割存在准确率较低和误分割等问题。针对这一问题,提出一种改进的U型卷积网络模型。该模型将U-Net网络和DenseNet网络融合,将解码器浅层特征连接至深层特征来增强特征的复用性。通过U-Net网络与卷积条件随机场(ConvCRF)的端到端结合训练来增强边缘特征,解决了边界模糊的问题。提出一种改进的focal loss损失函数,该函数提高了结节所占的权重,解决了正负样本不平衡的问题。在LUNA16数据集中作对比实验验证了模型的性能,分割精准度达到0.9374,敏感度为0.941,该结果证明了改进模型在肺结节分割中更优。Aimed at the complexity of lung CT image feature information and the problems of classical U-Net model such as low segmentation rate and mis-segmentation of lung nodules,an improved U-Net model is proposed.This model fused U-Net network and DenseNet network,and connected the shallow features of the decoder to the deep features to enhance the reusability of features.The edge features was enhanced through the combination of U-Net network and the end-to-end of the convolutional conditional random field(ConvCRF),which solved the problem of fuzzy boundary.An improved focal loss function was proposed,which increased the weight of nodules and solved the problem of imbalance between positive and negative samples.A comparative experiment was performed on the LUNA16 dataset to verify the performance of the model.The segmentation accuracy reached 0.9374 and the sensitivity was 0.941.The results prove that the improved model performs better in lung nodule segmentation.

关 键 词：肺结节分割 U型卷积网络 密集连接 损失函数 卷积条件随机场 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]