Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (22): 203-209.DOI: 10.3778/j.issn.1002-8331.2105-0052

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Automatic Segmentation Method of Spinal MR Images

YU Wentao, ZHANG Junhua, MEI Jianhua, LUO Xudong   

  1. 1.School of Information, Yunnan University, Kunming 650500, China
    2.School of Information, Yunnan Normal University, Kunming 650500, China
  • Online:2022-11-15 Published:2022-11-15

脊柱MR图像自动分割方法的研究

于文涛,张俊华,梅建华,罗旭东   

  1. 1.云南大学 信息学院,昆明 650500
    2.云南师范大学 信息学院,昆明 650500

Abstract: Accurate segmentation of spinal magnetic resonance(MR) images is the premise of spinal registration and 3D reconstruction. The traditional spinal MR image segmentation methods are complicated and low precision. To overcome the disadvantages of traditional methods, an automatic spinal MR image segmentation method based on deep learning is proposed. In this method, a symmetric channel convolutional neural network is designed to extract multi-scale image features. The residual connection solves the problem of network degradation in training and uses the jump connection layer to connect the features of the middle layer to reduce information loss. On this basis, the convolutional block attention mechanism is added to the proposed model to extract spatial and channel features more effectively. The experimental results show that the average DSC coefficient of this model on the test set is 0.8619, which is 15.34%, 7.08%, 5.79% and 3.1% higher than that of FCN, U-Net, DeeplabV3+ and UNet++ network models, respectively. The segmentation results can be applied in clinical practice to improve the segmentation accuracy of spinal MR images.

Key words: image segmentation, symmetric channel convolutional neural network, residual connection, deep learning, multi-scale feature

摘要: 脊柱磁共振(magnetic resonance,MR)图像精确分割是脊柱配准、三维重建等技术的前提。传统脊柱MR图像分割方法过程繁琐,精度低。为克服传统方法弊端,提出了一种基于深度学习的脊柱MR图像自动分割方法。该方法构建对称通道卷积神经网络提取多尺度图像特征,通过残差连接解决训练中网络退化问题,同时用跳跃连接层连接中间层特征减少信息丢失。在搭建的网络模型中加入卷积块注意力机制关注空间和通道中的有效特征。实验结果表明,该模型在测试集上的平均DSC系数为0.861?9,相比FCN、U-Net、DeeplabV3+和UNet++网络模型分别提高了15.34%、7.08%、5.79%、3.1%。该模型可应用于临床实践中提升脊柱MR图像的分割精度。

关键词: 图像分割, 对称通道卷积网络, 残差连接, 深度学习, 多尺度特征