Semi-Supervised Cerebrovascular Segmentation by Hierarchical Convolutional Neural Network
Semi-Supervised Cerebrovascular Segmentation by Hierarchical Convolutional Neural Network
复制标题
分层卷积神经网络半监督脑血管分割
DOI：
10.1109/access.2018.2879521
复制
发表时间：
2018
期刊：
影响因子：
3.9
通讯作者：
Liang Jimin
中科院分区：
文献类型：
--
作者：
Zhao Fengjun;Chen Yibing;Chen Fei;He Xuelei;Cao Xin;Hou Yuqing;Yi Huangjian;He Xiaowei;Liang Jimin
由于脑血管的弯曲性和复杂性以及与背景相似的强度分布，磁共振血管成像（MRA）对脑血管的准确分割仍然具有挑战性。以往基于规则的方法在临床准确诊断时存在对复杂血管分割不足、依赖领域知识、缺乏量化估计等局限性。本文提出了一种基于层次卷积神经网络（H-CNN）的半监督脑血管分割方法，该方法将基于规则的方法中精细的模型/特征设计转化为MRA图像到脑血管的映射。首先，我们用中心线和估计的半径生成了脑血管的管级标记。其次，我们使用MRA图像和相应的管级标签构建并训练H-CNN。第三，采用基于部分标注体素级地真值定义的综合指数（CI）确定H-CNN的停止准则。将H-CNN与血管性、双高斯、最优定向通量、血管增强扩散、连续开关混合扩散、Mimics软件、卷积神经网络2D (CNN)2D和CNN3D进行了对比。H-CNN的平均灵敏度、准确度和CI分别为94.69%、97.85%和2.99%，优于其他方法。曲面改造也将H-CNN在脑血管分割方面的效果可视化。在仅给定管级标签的情况下，本文提出的H-CNN方法通过CNN的分层更新实现体素级血管分割。H-CNN在标签部分正确的情况下，有可能应用于脑血管疾病的准确诊断和其他医学图像分割。
Due to the tortuosity and the complexity of cerebral vasculature and the similar intensity distribution with the background, it remains challenging to accurately segment cerebral vessels from magnetic resonance angiography (MRA). The previous rule-based methods have limitations when applied to accurate clinical diagnosis, such as the under-segmentation on complex vessels, the dependence on domain knowledge, and the lack of quantification estimation. In this paper, we proposed a semi-supervised cerebrovascular segmentation method with a hierarchical convolutional neural network (H-CNN) that transfers the exquisite model/feature design in rule-based methods to solve the mapping from MRA images to cerebral vessels. First, we generated the tube-level labels of cerebral vessels with centerlines and estimated radii. Second, we constructed and trained an H-CNN with the MRA images and corresponding tube-level labels. Third, the stopping criterion of the proposed H-CNN was determined by the comprehensive index (CI) that was defined based on partially annotated voxel-level ground truth. The comparison of our H-CNN with the vesselness, bi-Gaussian, optimally oriented flux, vessel enhancing diffusion, hybrid diffusion with continuous switch, Mimics software, convolutional neutral network2D (CNN)2D, and CNN3D were conducted on six testing images. The mean sensitivity, accuracy, and the CI of our H-CNN are 94.69%, 97.85%, and 2.99%, respectively, outperforming the other methods. The curved planar reformation also visualized the performance of H-CNN for cerebrovascular segmentation. Given only the tube-level labels, the proposed H-CNN method accomplished the voxel-level vessel segmentation via the hierarchical update of CNN. The H-CNN can potentially to be applied for the accurate diagnosis of cerebrovascular diseases and other medical image segmentation with only partially correct labels.