摘要
肺结节的准确分割,对结节恶性肿瘤风险分析及预测起到重要作用。在肺结节自动分割领域,结节分割的准确度普遍不高,对结节的空间信息利用不完全。针对这一情况,基于改进的U-Net提出肺结节分割的3D网络模型。为了综合考虑局部信息与全局信息,采用多尺度的特征提取方法。为了高效地适应不同尺度的特征映射,为浅层特征的映射提供指导信息,所提出的模型在U-Net的基础上添加了全局注意力上采样模块(Global Attention Upsample,GAU)。针对空间信息在编码器自上而下的结构中会逐步丢失,为了最大化利用空间特征信息,所提出模型方法采用一种空间注意力(Gateunit)机制来逐步抑制不同特征尺度的不相关背景信息,并能将注意力集中在前景之中。分割阶段在LIDC-IDRI数据集上的肺实质尺度Dice系数可达到82.51%。与其他模型相比,取得了有竞争力的结果。
Accurate segmentation of pulmonary nodules plays an important role in the risk analysis and prediction of malignant nodular tumors.In the field of automatic pulmonary nodule segmentation,the accuracy of nodule segmentation is generally not high,and the spatial information of nodule is not fully utilized.To solve this problem,a 3D network model for pulmonary nodule segmentation is proposed based on the improved U-net.In order to comprehensively consider local information and global information,a multi-scale feature extraction method is adopted.In order to adapt to different scale feature mapping efficiently and provide guidance information for shallow feature mapping,the Global Attention Upsampling(GAU)module is added to the model based on U-Net.In view of the gradual loss of spatial information in the top-down structure of the encoder,in order to maximize the utilization of spatial feature information,the proposed model adopts a spatial Gateunit to gradually suppress irrelevant background information of different feature scales and focus attention on the foreground.The Dice coefficient on the LDC-IDRI data set can reach 82.5%in the segmentation stage,and the results are competitive compared with other models.
作者
夏家权
XIA Jiaquan(School of Advance Manufacturing,Fuzhou University,Fuzhou 350108,China)
出处
《电视技术》
2023年第8期63-67,74,共6页
Video Engineering
