SAIC, Arlington, VA, United States.
Med Image Anal. 2012 Aug;16(6):1241-58. doi: 10.1016/j.media.2012.06.004. Epub 2012 Jun 26.
A graph of tissue vasculature is an essential requirement to model the exchange of gasses and nutriments between the blood and cells in the brain. Such a graph is derived from a vectorized representation of anatomical data, provides a map of all vessels as vertices and segments, and may include the location of nonvascular components, such as neuronal and glial somata. Yet vectorized data sets typically contain erroneous gaps, spurious endpoints, and spuriously merged strands. Current methods to correct such defects only address the issue of connecting gaps and further require manual tuning of parameters in a high dimensional algorithm. To address these shortcomings, we introduce a supervised machine learning method that (1) connects vessel gaps by "learned threshold relaxation"; (2) removes spurious segments by "learning to eliminate deletion candidate strands"; and (3) enforces consistency in the joint space of learned vascular graph corrections through "consistency learning." Human operators are only required to label individual objects they recognize in a training set and are not burdened with tuning parameters. The supervised learning procedure examines the geometry and topology of features in the neighborhood of each vessel segment under consideration. We demonstrate the effectiveness of these methods on four sets of microvascular data, each with >800(3) voxels, obtained with all optical histology of mouse tissue and vectorization by state-of-the-art techniques in image segmentation. Through statistically validated sampling and analysis in terms of precision recall curves, we find that learning with bagged boosted decision trees reduces equal-error error rates for threshold relaxation by 5-21% and strand elimination performance by 18-57%. We benchmark generalization performance across datasets; while improvements vary between data sets, learning always leads to a useful reduction in error rates. Overall, learning is shown to more than halve the total error rate, and therefore, human time spent manually correcting such vectorizations.
组织血管图是对大脑中血液和细胞之间的气体和营养交换进行建模的基本要求。这样的图是从解剖数据的矢量化表示中得出的，它提供了所有血管的顶点和线段图，并可能包括非血管成分（如神经元和神经胶质细胞体）的位置。然而，矢量化数据集通常包含错误的间隙、虚假的端点和错误合并的链。目前纠正这些缺陷的方法仅解决了连接间隙的问题，并且还需要在高维算法中手动调整参数。为了解决这些缺点，我们引入了一种监督机器学习方法，该方法 (1) 通过“学习的阈值松弛”连接血管间隙；(2) 通过“学习消除删除候选链”去除虚假的片段；以及 (3) 通过“一致性学习”在学习的血管图校正的联合空间中强制一致性。人类操作员仅需要标记他们在训练集中识别的单个对象，而无需调整参数。监督学习过程检查每个考虑中的血管段的邻域的几何形状和拓扑结构。我们在四组微血管数据上展示了这些方法的有效性，每组数据都有 >800(3) 个体素，这些数据是通过对小鼠组织进行全光学组织学和使用图像分割的最新技术进行矢量化获得的。通过基于统计学的验证抽样和精度召回曲线分析，我们发现使用袋装增强决策树进行学习可以将阈值松弛的等错误错误率降低 5-21%，并将链消除性能提高 18-57%。我们在数据集之间进行了泛化性能的基准测试；虽然改进因数据集而异，但学习始终可以降低错误率。总体而言，学习将总错误率降低了一半以上，因此，节省了手动纠正此类矢量化的人工时间。
