计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (3): 493-501.DOI: 10.3778/j.issn.1673-9418.1901052

• 图形图像 • 上一篇    下一篇

三维卷积网络在脑海马体分割中的应用

刘辰,肖志勇,吴鑫鑫   

  1. 江南大学 物联网工程学院,江苏 无锡 214122
  • 出版日期:2020-03-01 发布日期:2020-03-13

Application of Three-Dimensional Convolution Network in Brain Hippocampus Segmentation

LIU Chen, XIAO Zhiyong, WU Xinxin   

  1. School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2020-03-01 Published:2020-03-13

摘要:

为了提高海马体分割的精确性和鲁棒性,提出一种新型的三维卷积网络Dilated-3DUnet。该网络中卷积层的通道数采用“金字塔”分布的方式,有效缩小了参数的规模。此外,使用三维空洞卷积作为级联卷积操作,不仅有效地结合了脑磁共振成像(MRI)的深层特征和浅层特征,而且在不改变参数个数的情况下,扩大了卷积层的感受野,获取了多尺度信息,能够更好地捕捉MRI图像的浅层特征,从而提高了分割精度。在ADNI数据集上进行实验,以相似性系数、灵敏度、阳性预测率为评价指标,准确率分别达到了89.32%、88.72%和90.05%。实验表明,Dilated-3DUnet充分利用了脑MRI图像的三维空间信息,具有更强的泛化能力和更好的特征表达能力,从而大大提升了分割精度。

关键词: 海马体分割, 脑磁共振成像(MRI), 卷积网络, 空洞卷积

Abstract:

In order to improve the accuracy and robustness of hippocampus segmentation, a new three-dimensional convolutional network named Dilated-3DUnet is proposed. The number of channels in the convolution layer of the network adopts the “Pyramid” distribution method, which effectively reduces the size of the parameters. In addition,using 3D dilated convolution as cascading convolution operation not only effectively combines the deep and shallow features of brain MRI (magnetic resonance imaging) images, but also expands the receptive field of convolution without changing the number of parameters. Multi-scale information is obtained, which can better capture the shallow features of the brain MRI image, so as to improve the segmentation accuracy. Experiments are carried out on the    ADNI dataset, using dice similarity coefficient, sensitivity and predictive positivity value as evaluation indexes, and the accuracy reaches 89.32%, 88.72% and 90.05%, respectively. Experiments show that Dilated-3DUnet makes full use of the three-dimensional spatial information of brain MRI images, which has stronger generalization ability and better feature expression ability, thus greatly improving the segmentation accuracy.

Key words: hippocampus segmentation, brain magnetic resonance imaging (MRI), convolutional network, dilated convolution