摘要
腰椎解剖区域自动分割在脊柱影像自动分析流程中发挥着重要作用。尽管经典的卷积神经网络能够捕捉影像全局特征,其局部先验和权重共享的特性限制了长距离建模的能力。为了解决以上问题,本文提出了一种用于腰椎解剖区域分割的Swin Transformer融合网络,将Swin Transformer网络和多尺度空洞卷积融合作为编码器来得到全局和局部特征的层次化表达。设计了特征耦合模块,在通道和空间2个维度将来自Transformer模块和卷积模块的特征进行耦合,提高了模型的局部和长距离建模能力。为了解决开源数据缺乏的问题,提出了带有体素级标注的、包含663个腰椎椎骨计算断层成像的数据集。在此数据集上的实验表明提出的模型分割精度超过了典型医学图像分割方法,本文模型的骰子系数、Hausdorff距离和平均表面距离分别为88.24%、14.48和0.997。消融实验进一步验证了所提出模块的有效性。
Automated segmentation of the lumbar spine anatomical region plays a crucial role in the automated analysis pipeline of spinal images.Although classical convolutional neural networks can capture global image features,their inherent local priors and weight-sharing characteristics limit their ability to model long-range dependencies.To address these issues,a Swin Transformer hybrid network is proposed for the segmentation of the lumbar anatomical region.Firstly,the Swin Transformer hybrid network and multi-scale dilated convolution are combined as an encoder to achieve the hierarchical representation of global and local features.Additionally,a feature coupling module is designed,which couples the features of the Transformer and CNN in the channel and spatial dimensions,enhancing the model′s local and long-distance modeling capabilities.Dealing with data scarcity problems,a dataset composed of 663 lumbar vertebrae CT images with voxel-level labeled annotations is proposed.Experiments on this dataset show that the segmentation accuracy of the proposed model surpasses that of typical medical image segmentation methods.Specifically,the dice coefficient,the Hausdorff distance,and the average surface distance of the proposed model are 88.24%,14.48,and 0.997,respectively.Ablation experiments further verify the effectiveness of the proposed modules.
作者
张英迪
史泽林
王欢
崔少千
张磊
刘嘉琛
单修祺
刘云鹏
赵恩波
ZHANG Yingdi;SHI Zelin;WANG Huan;CUI Shaoqian;ZHANG Lei;LIU Jiachen;SHAN Xiuqi;LIU Yunpeng;ZHAO Enbo(Key Laboratory of Opto-Electronic Information Processing,Chinese Academy of Sciences,Shenyang 110016,China;Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;University of Chinese Academy of Sciences,Beijing 100049,China;ShengJing Hospital of China Medical University,Shenyang 110004,China)
出处
《信息与控制》
北大核心
2025年第3期390-400,共11页
Information and Control
基金
辽宁省自然科学基金项目(2022-KF-12-09)。
