Xiong Guanglei, Zhou Xiaobo, Degterev Alexei, Ji Liang, Wong Stephen T C
Bioinformatics Division, TNLIST and Department of Automation, Tsinghua University, Beijing, People's Republic of China.
Cytometry A. 2006 Jun;69(6):494-505. doi: 10.1002/cyto.a.20296.
To investigate the intricate nervous processes involved in many biological activities by computerized image analysis, accurate and reproducible labeling and measurement of neurites are prerequisite. We have developed an automated neurite analysis method to assist this task.
Our approach can be considered as automated with certain user interaction in setting initial parameters. Single and connected centerlines along neurites are extracted. The computerized method can also generate branching and end points. Owing to its multi-scale flexibility, both thick and thin neurites are simultaneously detected.
We employ the relative neurite length difference (defined as the difference between the lengths obtained by automated and manual analysis divided by the total length of the latter) and neurite centerline deviation (defined as the area of the regions enclosed by different paths between automated and manual analysis divided by the total length of the former) to evaluate the performance of our algorithm, which is of great interest in neurite analysis. The average of the relative length difference is about 0.02, while the average of the centerline deviation is about 2.8 pixels. The probabilities of the distributions being the same from the Kolmogorov-Smirnov (KS) test of the automatic and manual results are 99.79%. The KS test also shows no significant bias between different observers based on the proposed new validation scheme.
With the accurate and automated extraction of neurite centerlines and measurement of neurite lengths, the proposed method, which greatly reduces human labor and improves efficiency, can serve as a candidate tool for large-scale neurite analysis beyond the capability of manual tracing methods.
为了通过计算机图像分析研究许多生物活动中复杂的神经过程，准确且可重复的神经突标记和测量是前提条件。我们开发了一种自动神经突分析方法来辅助这项任务。
我们的方法可被视为在设置初始参数时有一定用户交互的自动化方法。沿着神经突提取单个和相连的中心线。该计算机化方法还可以生成分支和端点。由于其多尺度灵活性，粗细神经突都能同时被检测到。
我们采用相对神经突长度差异（定义为自动分析和手动分析得到的长度之差除以后者的总长度）和神经突中心线偏差（定义为自动分析和手动分析之间不同路径所围成区域的面积除以前者的总长度）来评估我们算法的性能，这在神经突分析中具有重要意义。相对长度差异的平均值约为0.02，而中心线偏差的平均值约为2.8像素。自动和手动结果的Kolmogorov-Smirnov（KS）检验表明分布相同的概率为99.79%。KS检验还表明基于所提出的新验证方案，不同观察者之间没有显著偏差。
通过准确且自动地提取神经突中心线和测量神经突长度，所提出的方法极大地减少了人力并提高了效率，可作为一种超越手动追踪方法能力的大规模神经突分析的候选工具。
