Neuroinform DOI 10.1007/s12021-017-9325-1

ORIGINAL ARTICLE

Ensemble Neuron Tracer for 3D Neuron Reconstruction Ching-Wei Wang1,2 Hanchuan Peng4

· Yu-Ching Lee2,3 · Hilmil Pradana1,2 · Zhi Zhou4 ·

© Springer Science+Business Media New York 2017

Abstract Tracing of neuron paths is important in neuroscience. Recent studies have shown that it is possible to segment and reconstruct three-dimensional morphology of axons and dendrites using fully automatic neuron tracing methods. A specific tracer may be better than others for a specific dataset, but another tracer could perform better for some other datasets. Ensemble of learners is an effective way to improve learning accuracy in machine learning. We developed automatic ensemble neuron tracers, which consistently perform well on 57 datasets of 5 species collected from 7 laboratories worldwide. Quantitative evaluation based on the data generated by human annotators shows that the proposed ensemble tracers are valuable for 3D neuron tracing and can be widely applied to different datasets.

Electronic supplementary material The online version of this article (doi:10.1007/s12021-017-9325-1) contains supplementary material, which is available to authorized users.  Ching-Wei Wang

[email protected] 1

Graduate Institute of Biomedical Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan

2

NTUST Center of Computer Vision and Medical Imaging, National Taiwan University of Science and Technology, Taipei, Taiwan

3

Graduate Institute of Applied Science and Technology, National Taiwan University of Science and Technology, Taipei, Taiwan

4

Allen Institute for Brain Science, Seattle, WA 98103, USA

Keywords 3D neuron reconstruction · Ensemble neuron tracer

Introduction Understanding how the brain works by acquiring profound knowledge of the structure, function, and development of the nervous system is of crucial importance, enabling development of drugs and therapies in treating neurological and psychiatric disorders. As the three-dimensional shape of a neuron plays a major role in determining the brain connectivity and function (Chklovskii 2004; Kalisman et al. 2003; Krichmar et al. 2002; Mainen and Sejnowski 1996; Schaefer et al. 2003; Sholl 1956; Spruston 2008), 3D reconstruction of complex neuron morphology is important. In recent years, a number of studies have been conducted in neuron tracing (Abramoff et al. 2004; Brown et al. 2005; Collins 2007; Feng et al. 2014; Fiala 2005; Kim et al. 2016; Li et al. 2010; Lu 2009; Xiao and Peng 2013; Peng et al. 2015). Tracing of three-dimensional morphology of axons and dendrites is important in neuroscience, enabling scientists to distinguish and characterize neuron phenotypes, to model projection and potential connectivity patterns, and to simulate the electrophysiological behavior of neurons. A key component in 3D reconstruction of neuron morphology is how to extract data of neuronal morphology from microscopic imaging data to digital reconstruction. The steps can be summarized as follows (Meijering 2010). Firstly, data preprocessing includes noise reduction, deconvolution, shading correction, and mosaicking. Secondly, image filtering such as binarization and morphological filtering

Neuroinform

is applied. The third step is tree segmentation, which can be categorized into two types, i.e. global processing and local exploration. For global processing, methods include binarization, skeletonization, filling, prunning and point identification. On the other hand, for local exploration, methods include point identification, iterative model fitting, min-cost path searching and active contour fitting. The last step for segmentation process is spine segmentation, consisting of attached spine detection, detached spine detection, level-set segmentation, and heuristic sampling. Many recent studies have proposed automatic approaches to reconstruct the neuronal morphology from highresolution 3D images (Choromanska et al. 2012; Li et al. 2010; Peng et al. 2011; Peng et al. 2015; Peng et al. 2010; Yang et al. 2013; Xiao and Peng 2013; Wearne et al. 2005; Ming et al. 2013; Wang et al. 2011) , including ray casting (Wearne et al. 2005; Ming et al. 2013), pruning of overreconstructions (Peng et al. 2011; Xiao and Peng 2013), deformable curves (Peng et al. 2010; Wang et al. 2011), and others. In 2013, Ming et al.’s method (Ming et al. 2013) used a prediction and refinement strategy and they extended the rayburst sampling algorithm to a marching fashion for the rapid reconstruction of 3D neuronal morphology from light microscopy images, which achieves a reasonable balance between fast speed and acceptable accuracy. Xiao and Peng developed all-path-pruning 2.0 (APP2) for 3D neuron tracing (Xiao and Peng 2013) based on fast marching and hierarchical pruning in 2013. The method can generate more satisfactory results in most cases than several previous methods within a short amount of time. In 2011, Wang et al. proposed a broadly applicable algorithm for automated tracing of neuronal structures in 3D microscopy images (Wang et al. 2011). The method is based on 3D open-curve active contour (Snake). The open-snake tracing system provides multiple tracing modes to accommodate different image datasets. However, it is challenging to determine the optimal tracer for varying datasets. It is thus interesting to study robust, automated tracing methods that could be applied to 3D neuron image datasets from different sources. In machine learning, studies have shown that an ensemble learning algorithm could improve classifier performance and learning accuracy by combining multiple basic learners (Polikar 2006; Rokach 2010; Viola and Jones 2004). For instance, Breiman (1996) introduced a Bagging algorithm, which forms an ensemble by aggregating multiple classifiers, each of which is trained using a bootstrapped training set by data perturbation. Inspired by such an ensemble approach, we developed a fully automatic ensemble neuron

tracer by adopting the powerful ensemble framework with integration of data perturbation and model selection techniques and demonstrated its strength using quantitative evaluation. In evaluation, quantitative evaluations were conducted based on the data generated by human annotators, showing that the proposed ensemble tracer is valuable for 3D neuron tracing and can be widely applied to different datasets.

The Proposed Method - Ensemble Neuron Tracer We developed a fully automatic neuron tracer in an ensemble framework with integration of data perturbation and model selection concepts as in the ensemble machine learning algorithms. By regenerating multiple samples using the proposed data perturbation method, it enables us to generate multiple tracing results by base tracers, and then the ensemble framework combines and selects an optimal tracing result as the final output using the proposed model selection methods. The proposed ensemble neuron tracer contains two data perturbation approaches DP1 and DP2 for regenerating multiple data samples from a single data sample to increase the data variations, a smart data selector (denoted as SDS) for rejecting the generated data with the highest probability in containing large number of false positive signals, a combiner (denoted as C) for integrating multiple tracing results produced by base tracers (each is denoted as f (·)) and a model selector (MS) for producing the final output. We chose APP2 (Xiao and Peng 2013) as the base tracer f (·) as it is very fast and accurate based on pruning a dense initial reconstruction of a neuron. Each base tracer uses an independent data perturbation process, which consists of two steps DP1 and DP2 intermediated by a smart data selection (SDS) module. The combiner C merges multiple neuron reconstructions by fusing nearby compartments in several reconstructions. The final selector MS determines the optimal reconstruction by choosing the one with the longest neurite paths (the greatest number of compartments in our implementation) but shorter than an empirically predefined threshold (default: 11000 compartments). With respect to the ensemble machine learning algorithms such as bagging and boosting (Freund and Schapire 1999), as bagging is more robust to noisy data the proposed ensemble framework adopts a bagging-like strategy instead of a boosting approach, and individual base tracing results are generated in parallel instead of sequentially. A flowchart of the proposed ensemble neuron tracer is shown in Fig. 1, and the algorithm of the proposed ensemble neuron tracer is shown in Algorithm 1.

Neuroinform

Fig. 1 The flowchart of the proposed ensemble neuron tracer

Algorithm 1 Ensemble neuron tracing algorithm Input : : a 3D neuron data. Auxiliary methods : 1 : Data Perturbation 1; 2: Data Perturbation 2; : : Smart Data Selection; ( ): a Base Tracer; : a Combiner; : Model Selection. Output : : an ensemble neuron tracing output. 1 input a neuron data 2 use data perturbation 1 to generate multiple data samples using to gener3 remove the abnormal data from ate to generate 4 use data perturbation 2 5 apply base tracer

( ) to

to generate base tracings

6 combine base tracings to generate base ensemble tracings by 7 select one ensemble neuron tracing as the final ensemble output from using

sample in some way. In implementation of data perturbation DP1 , three models (NModel, SModel and BModel) are used to apply on the input data d to produce D  :{di } where i = 1 . . . 3. 1. NModel d1 (x, y, z)

 =

1 , if d(x, y, z) > N 0 , otherwise



d(x, y, z) N = +k× #d(x, y, z)



(1)

d 2 (x, y, z) − #d(x, y, z)



d(x, y, z) #d(x, y, z)

2 (2)

where k is an input parameter and empirically defined as -0.2. 2. SModel  1 , if d(x, y, z) > S (3) d2 (x, y, z) = 0 , otherwise 

⎛

⎛

⎜ d(x, y, z) ⎜ ⎜ S = ×⎜ ⎝1−k × ⎝1− #d(x, y, z)



d 2 (x,y,z) #d(x,y,z) −

R

⎞⎞

d(x,y,z) 2 ⎟⎟ #d(x,y,z)



⎟⎟ ⎠⎠ (4)

Auxiliary Algorithm - Data Perturbation 1 (DP1 ) The idea of data perturbation is to produce multiple data samples from a single data sample by modifying the input

where R is equal to the maximum possible value of the standard deviation of the data range, and k is an input parameter and empirically defined as 0.5.

Neuroinform

Fig. 2 Smart data selection on three data samples with density equal to 0. Samples with zero density values are deemed as clean data and selected for subsequent data perturbation DP2 and tracing

3. BModel d3 (x, y, z) =

B =



1 , if d(x, y, z) > B 0 , otherwise

lmax + lmin 2

(5)

(6)

where lmax and lmin are the maximum and minimum values in d(x, y, z), respectively. Auxiliary Algorithm - Smart Data Selection (SDS) Given D  from DP1 , the smart data selection model searches for the data with the highest probability in containing large number of false positive signals and rejects the

abnormal data. Afterwards, the system continues on the second stage data perturbation DP2 . In detection of the data with the highest probability in containing large number of false positive signals, the smart data selection model first divides the data di into J blobs Bi,j and locates the farthest corner blob Bi,c to the blob Bi,s with the strongest signals. Bi,s = arg max



Bi,j (x, y, z)

j

Bi,c = arg max||Bi,j − Bi,s ||

(7)

j

where || · || is the Euclidean distance operation. Afterwards, compute the density value ei of Bi,c  SDS Bi,c where ei = #Bi,c . If ei is equal to 0, this indicates that all

Neuroinform

data in D  are clean and no sample is rejected; D  =D  . Otherwise, SDS chooses the data dw with the highest density value and rejects it. dw = arg maxei D 

=

D

di

− dw

(8)

Figure 2 shows three examples of clean data (ei = 0), and the smart data selection keeps all data in D  . Figure 3 shows other four examples in common situations where the smart data selection SDS rejects the data with the highest density.

 } D  :{da,z a=1...M,z=1...Z by different morphological operation models. In implementation of DP2 , three samples are created from one input sample (Z=3). The original copy is  = d  , and the kept unchanged as the first data sample; da,1 a rest samples are created by applying two different morphological operations to da . The first morphological operation, Dilation, aims to enlarge the boundaries of neurons by enhancing the structuring elements in 3D, and the second morphological operation, Closing, is designed to remove small holes and get better tracing results. DP2 produces the  and d  using a structuring element H . data da,2 a,3  da,2 = da ⊕ H

(9)

Auxiliary Algorithm - Data Perturbation 2 (DP2 ) The idea of data perturbation is to produce multiple data samples from a single data sample by modifying the input sample in some way. Given the data D  :{da }a=1...M from the data selection SDS, DP2 creates

 = (da ⊕ H )  H da,3

(10)

where H is a 3 × 3 × 3 cubic, and Hˆ is the reflection of the structuring element H .

Fig. 3 Smart data selection on four data samples with densities greater than zero. As the sample dw with the largest density value may contain the highest number of false positive signals, dw is rejected in SDS for subsequent data perturbation DP2 and tracing

Neuroinform

Fig. 4 Results of the 40 Drosophila neuron images by ENT(APP2)

Auxiliary-Combiner (C) The Combiner combines base tracings ta,z |a=1...M,z=1...Z to generate M base ensemble tracings Ta∗ =C(ta,z ) |a=1...M,z=1 . . . Z. For example, each base tracing result is represented by trees Ta,k , containing nodes N(Ta,k ) and edges E(Ta,k ); Ta,k =(Na,k , Ea,k ) means that Na,k and Ea,k are the node set and edge set of Ta,k , respectively. In merging a new tree T2 to an existing tree T1 , a classification method is built to decide whether the nodes in T2 are redundant or new ones to add. In classification of node types, a distance function d is developed to measure the shortest distance of a node nTi 2

to a node in T1 ; nTk 1 =arg min | nTi 2 (x, y, z) − nTj 1 (x, y, z) |, nTi 2 (x, y, z)

j T1 nk (x, y, z)

d=| − |. If d > r, then creates new nodes to add; otherwise, replace nTi 2 with nTk 1 where r is a user-input threshold. Auxiliary Algorithm - Model Selection (MS) MS select one ensemble neuron tracing as the final ensemble output t ∗ from all base ensemble tracings T ∗ :{ta | ta = C(ta,z )}a=1...M,z=1...Z . In the model selection, the number of nodes is the key element for selecting the ensemble

Neuroinform Fig. 5 10 Exemplar samples in testing flycircuit.org images. Blue rectangles highlight the misdetections by the benchmark approaches

neuron tracing output. MS first classify whether the input data isa very clean data or not based on the density value ei . If ei = 0, then the input data is determined as a clean data sample. Then, MS selects the base ensemble tracing model with the maximum number of the nodes; three examples are shown in Fig. 2. Otherwise, MS first rejects the base

ensemble model(s) with too many nodes as which tend(s) to contain a large number of false positives. Here, if the number of nodes is larger than 1 , the base ensemble models will be removed from the candidate base ensemble models Tc (1 is empirically defined as 11000). Then, MS selects the base ensemble tracing model with the maximum number

Table 1 Computational time (in seconds) on forty samples

95 % C.I. for mean Methods

N

Mean

Std. Deviation

Std. Error

Lower Bound

Upper Bound

APP2 MOST Simple tracing ENT(APP2)

40 40 40 40

1.31 1.41 59.9 3.55

0.07 0.1 43.95 0.24

0.01 0.02 6.95 0.04

1.29 1.37 45.85 3.47

1.34 1.44 73.96 3.63

Neuroinform Table 2 57 selected single neuron samples of five species (chick, frog, fruitfly, zebrafish and human)

Total

#datasets

Species

Data contributor

Country

1 5 1 4 23 12 4 3 4 57

frog Drosophila human human Drosophila Drosophila chick zebrafish zebrafish 5 species

Scrippts Institute George Mason University Allen Institute for Brain Science University of Cambridge Janelia Research Campus FlyCircuit.org University of Washington University of Washington University of Washington 7 data contributors

USA USA USA UK USA Taiwan USA USA USA 3 countries

of the nodes as output t from the candidate base ensemble models. Four examples are presented in Fig. 3.  TC =

 T∗ , ei = 0 T r = {t ∈ T ∗ : N(t) ≤ 1 } , otherwise

t ∗ = arg maxN(tj ) | tj ∈ T C

(11)

(12)

tj

Experimental Results Data and Evaluation Approach The experiments were conducted in two parts with a test on 40 single neuron samples of a Drosophila brain dataset and a more comprehensive test with quantitative evaluation on 57 single neuron datasets of 5 species. For the 40 single neuron samples of the Drosophila brain, the

voxel size is 0.29 × 0.29 × 1 μm, the width and height are 297.58μm and the range of the depth is from 104 to 131μm. The image dimension of individual data ranges is from 1024 × 1024 × 104 to 1024 × 1024 × 131 pixels. The file size of individual data samples ranges from 100 MB to 140 MB. For the 57 single neuron samples corresponding to five different species, the file size ranges from 10 MB to 156 MB and the image dimension ranges from 512 × 512 × 41 pixels to 1882 × 1025 × 85 pixels. For quantitative evaluation, two measurement approaches were applied to measure the spatial distance between the tracing outcomes of individual methods and the gold standard data, including the Entire Structure Average (ESA) and the Different Structure Average (DSA). ESA represents the average shortest distance between two reconstructions’ compartments (Zhou et al. 2014); DSA is the average distance between corresponding compartments, which are apart from each other with a visible spatial separation (defined ≥ 2 voxels apart) (Zhou et al. 2014).

Table 3 ESA and DSA results of the 57 datasets

95 % Confidence Interval for Mean Mean Std. Deviation Std. Error Lower Bound Upper Bound ESA

ENT(APP2) APP2 Simple Tracing ENT(MOST) MOST DSA ENT(APP2) APP2 Simple Tracing ENT(MOST) MOST

8.31 19.18 68.81 8.84 15.7 11.66 25.03 69.27 14.38 20.87

10.08 26.94 84.03 12.14 30.68 12.46 29.79 80.97 21.13 33.63

1.34 3.57 11.13 1.61 4.06 1.65 3.95 10.72 2.8 4.45

5.63 12.03 46.51 5.61 7.56 8.35 17.12 47.78 8.78 11.94

10.98 26.33 91.11 12.06 23.84 14.96 32.93 90.75 19.99 29.79

Neuroinform

Fig. 6 ESA and DSA of the 57 datasets. Both ensemble tracers produced more similar reconstructions to the referenced standard manual reconstructions, achieving lower spatial distance to the respective referenced standard

Evaluation Results of 40 Drosophila Samples Similar to the ensemble learning methods where users have to specify a learning algorithm as the base learner, in the proposed ensemble framework, users also have to specify a tracing algorithm as the base tracer. In experiments, we chose the All-Path-Pruning 2 algorithm (APP2) (Xiao and Peng 2013) and Micro-Optical Sectioning Tomography (MOST) (Li et al. 2010) as the base tracer f (.) because these approaches are very fast and accurate based on pruning a dense initial reconstruction of a neuron. In evaluation, initially we collected 2000 samples from the FlyCircuit database (http://www.flycircuit.tw/) (Chiang 2011), and 40

samples of 3D confocal images of Drosophila single neurons were then randomly selected for a preliminary test to compare the proposed method with three other neuron tracers (Micro-Optical Sectioning Tomography (MOST) (Li et al. 2010), APP2 (Xiao and Peng 2013) and simple tracing (ST) (Yang et al. 2013)). The ensemble tracer is demonstrated to be able to reconstruct neurons (Fig. 4) while three recent automatic tracing methods perform poor (Fig. 5). Compared to the benchmark approaches using the evaluation approach - Percentage of Different Structure (PDS) (Peng et al. 2010), the ensemble tracer ENT(APP2) was able to reconstruct neurons with 23 % less error than APP2 and 53 % less error than ST or MOST on average.

Table 4 Paired T tests using 57 datasets using ESA Paired Differences

t

df

Sig. (2-tailed)

95 % Confidence Interval of the Difference Mean

Std. Deviation

Std. Error Mean

Lower

Upper

ENT(APP2)-APP2

−10.86900

25.88019

3.42791

−17.73594

−4.00206

−3.171

56

.002

ENT(APP2)-SimpleTracing

−60.50125

81.73255

10.82574

−82.18782

−38.81468

−5.589

56

.000

ENT(APP2)-MOST

−7.39503

25.80299

3.41769

−14.24148

−.54857

−2.164

56

.035

ENT(MOST)-MOST

−6.86895

27.92794

3.69915

−14.27923

.54133

−1.857

56

.069

Neuroinform Fig. 7 Exemplar samples of 57 datasets with the best ESA scores by the proposed ENT(APP2) method

Neuroinform Fig. 8 Exemplar samples of 57 datasets with the best ESA scores by the proposed ENT(MOST) method

Neuroinform

Computational Time For analysis of the computational complexity and execution time of the proposed methodology compared to the other methods, a computational time analysis is conducted using the 40 Drosophila samples. The file size of individual data samples ranges from 100 MB to 140 MB, and the image dimension of individual neuron samples ranges from 1024x1024x104 pixels to 1024x1024x131 pixels. In evaluation of computational time, a computer with intel Xeon two cores 2.6 GHz CPU, 32 GB RAM and fedora 20 operating system is used. Table 1 compares the computation time (in seconds) for the proposed method, namely ENT(APP2), and three benchmark approaches. The results show that simple tracing takes the longest computing time for processing a 3D sample, costing 59.9 seconds on average. APP2 and MOST take no more than 2 seconds to process one 3D sample. In comparison, although the presented ensemble neuron tracer combines three base tracers adapted from APP2, the ensemble costs no greater than three times of computational time APP2 takes, and on average the presented approach spends 3.55 seconds to process a 3D sample. Quantitative Evaluation on 57 Datasets of 5 Species We also tested the ensemble tracer against 57 selected single neuron samples of five species (chick, frog, fruitfly, zebrafish and human) (Table 2) from the BigNeuron initiative (http://bigneuron.org) (Peng et al. 2015) (In this study, we exclude the mouse data as the tracing systems of the proposed method and the benchmark methods tend to fail due to the out of memory problem). For quantitative evaluation, the average spatial distance (ESA) and the average distance between corresponding compartments (DSA) (Zhou et al. 2014) were computed automatically by comparing the referenced standard manual reconstructions produced in a BigNeuron annotation workshop and the tracing results by the proposed method and the benchmark approaches. Table 3 shows that both ensemble tracers produced more similar reconstructions to the referenced standard manual reconstructions, achieving lower spatial distance to the respective referenced standard. On average, ENT-APP2 generates the lowest ESA (8.3 voxels) and DSA (11.7 voxels), and ENT-MOST obtains the second lowest ESA (8.8 voxels) and DSA (14.4 voxels). In comparison, ESAs and DSAs of the benchmark approaches range between 15.7 to 68.8 voxels and between 20.9 to 69.3 voxels, respectively. Figure 6 presents all ESA and DSA results, showing that show that the ensemble tracers produce less error than the benchmark approaches.

Furthermore, paired-samples t-tests (Table 4) were conducted to compare ESAs of ensemble neuron tracers and single tracers. The results show that the ensemble framework can greatly improve the performance of a single tracer and the ensemble tracer (ENT(APP2)) is significantly better than the single neuron tracer APP2 (p = 0.002). Also, ENT(MOST) outperforms MOST as shown in Table 4 and Fig. 6. The tracing results of the 57 datasets with the best ESA scores by the proposed ensemble tracer ENT(APP2) are presented in Fig. 7, and the tracing results of the 57 datasets with the best ESA scores by the proposed ensemble tracer ENT(MOST) are presented in Fig. 8, respectively. For the limitation of the proposed tracer ENT(APP2), due to the sequential searching behavior, the ensemble tracer may miss parts of neuron when the signals are weakly connected. In comparison, as ENT(MOST) does not conduct sequential searching like others, it detects fragments of highly expressed paths to achieve lower ESAs for the entire neuron structure (Fig. 8).

Conclusion and Discussion As three-dimensional shape of a neuron plays a major role in determining the brain connectivity and understand how the brain works, 3D reconstruction of complex neuron morphology is important. Recent studies have shown that it is possible to segment and reconstruct 3D morphology of axons and dendrites using fully automatic neuron tracing methods. However, there is no agreement on which tracer produces the best tracing results. A specific tracer may be better than others for a specific dataset, but another tracer could perform better for some other dataset (Meijering 2010; Peng et al. 2015). In the machine learning field, the ensemble strategy such as bagging and boosting has been proved to be an effective way to improve learning accuracy and produce more robust models. In this paper, a robust and fully automatic ensemble neuron tracing method has been presented and demonstrated to be promising for reconstructing complex 3D neuron morphology from 3D image data. The proposed ensemble tracer can accommodate multiple neuron tracers and combine the strengths of different tracers to generate optimal tracing results. We have demonstrated the ensemble tracer with three simple tracers, and the method consistently performs well on 57 datasets of 5 species collected from 7 laboratories worldwide. The proposed ensemble neuron tracer is not limited to the selected base tracing approaches and can also include other tracers as base tracers. The experimental results demonstrated that the proposed ensemble tracers are valuable for 3D neuron tracing and can be widely applied

Neuroinform

to various datasets. In future work, we would like to test more data with additional synthetic noises and distortions for further investigation. In addition, as massive amounts of neurons have been digitized across multiple species, brain regions, and laboratories worldwide, and the variability introduced through different animal species, developmental stages, and brain locations has made systematic analysis a challenging task. The next steps to be taken towards a more effective 3D reconstruction of neurons would be developing a large community-generated database of single neuron morphologies with open-source tools for neuroscience and community-driven protocols intended to serve as the standard for digital reconstruction of single neurons, enabling effective modelling of whole brain neuronal structures.

Information Sharing Statement The software implementation of the presented method is developed in C++ and built as a plugin in the VAA3D (RRID:SCR 002609) framework (Peng et al. 2010) (v.3.055 version) with Qt-4.7.2 installed. The software implementation has been tested in various operating systems, including Fedora 20, Windows 7 64-bit version and Mac 10.10. The source code and the software with a user guide and a developer guide are made available in (https://figshare.com/s/ e32f50372e94014cf44b). Some testing data with associated gold standard data generated by human annotators are also made publicly available for scientific communities to use (https://figshare.com/s/3cd6c6238fa8c580030a).

Acknowledgments Ching-Wei Wang, Hilmil Pradana and Yu-Ching Lee were supported by the Ministry of Science and Technology of Taiwan, under a grant (MOST-105-2221-E-011-121-MY2). Zhi Zhou and Hanchuan Peng were supported by Allen Institute for Brain Science, Seattle, WA, USA.

References Abramoff, M.D. et al. (2004). Image processing with ImageJ. Biophotonics International, 11, 36–42. Breiman, L. (1996). Bagging predictors. International Journal Machine Learning, 24, 134140. Brown, K.M. et al. (2005). A cross-platform freeware tool for digital reconstruction of neuronal arborizations from image stacks. Neuroinformatics, 3, 343–359. Chiang, e.t.al. (2011). Three-dimensional reconstruction of brain-wide wiring networks in Drosophila at single-cell resolution. Current Biology, 21, 1–11. Chklovskii, D.B. (2004). Synaptic connectivity and neuronal morphology: two sides of the same coin. Neuron, 43, 609–617.

Choromanska, A. et al. (2012). Automatic reconstruction of neural morphologies with multi-scale tracking. Neural Circuit, 6. doi:10.3389/fncir.2012.00025. Collins, T.J. (2007). ImageJ for microscopy. Biotechniques, 43, 25–30. Feng, L. et al. (2014). neuTube 1.0: A New Design for Efficient Neuron Reconstruction Software Based on The SWC Format eNeuro, doi:10.1523/ENEURO.0049-14. Fiala, J.C. (2005). Reconstruct: A free editor for serial section microscopy. Journal of Microscopy, 218, 52–61. Freund, Y., & Schapire, R. (1999). A Short Introduction to Boosting. Journal of Japanese Society for Artificial Intelligence, 14(5), 771– 780. Kalisman, N. et al. (2003). Deriving physical connectivity from neuronal morphology. Biological Cybernetics, 88, 210–218. Krichmar, J.L. et al. (2002). Effects of dendritic morphology on CA3 pyramidal cell electrophysiology: a simulation study. Brain Research, 941, 11–28. Kim, E.J. et al. (2016). Neuroscience: Optimized tracing of neural circuits. Nature Methods, 13, 470. Li, A. et al. (2010). Micro-optical sectioning tomography to obtain a high-resolution atlas of the mouse brain. Science, 330, 1404– 1408. Lu, J. (2009). Semi-automated reconstruction of neural processes from large numbers of fluorescence images. PLoS One, 4, e5655. Mainen, Z.F., & Sejnowski, T.J. (1996). Influence of dendritic structure on firing pattern in model neocortical neurons. Nature, 382, 363–366. Meijering, E. (2010). Neuron Tracing in Perspective. Cytometry, 77, 693–704. Ming, X., Li, A., Wu, J., Yan, C., Ding, W., Gong, H., Zeng, S., & Liu, Q. (2013). Rapid reconstruction of 3D neuronal morphology from light microscopy images with augmented rayburst sampling. PloS one, 8, e84557. Peng, H., Ruan, R., Long, F., Simpson, J.H., & Myers, E.W. (2010). V3D enables real-time 3D visualization and quantitative analysis of large-scale biological image data sets. Nature Biotechnology, 28, 348–353. Peng, H., Ruan, Z., Atasoy, D., & Sternson, S. (2010). Automatic reconstruction of 3D neuron structures using a graph-augmented deformable model. Bioinformatics, 26, i38–i46. Peng, H., Long, F., & Myers, G. (2011). Automatics 3D neuron tracing using all-path pruning. Bioinformatics, 27, i239–i247. Peng, H. et al. (2015). BigNeuron: large-scale 3D neuron reconstruction from optical microscopy images. Neuron, 87, 252– 256. Polikar, R. (2006). Ensemble based systems in decision making. IEEE Circuits and Systems Magazine, 6(3), 21–45. Rokach, L. (2010). Ensemble-based classifiers. Artificial Intelligence Review, 33, 1V394. Schaefer, A.T. et al. (2003). Coincidence detection in pyramidal neurons is tuned by their dendritic branching pattern. Journal of Neurophysiology, 89, 3143–3154. Sholl, D.A. (1956). The measurable parameters of the cerebral cortex and their significance in its organization. Progress in Neurobiology, 2, 324–33. Spruston, N. (2008). Pyramidal neurons: dendritic structure and synaptic integration. Nature Reviews Neuroscience, 4, 206– 221. Viola, P., & Jones, M. (2004). Robust real-time face detection. International Journal of Computer Vision, 57(2), 137V154. Wang, Y., Narayanaswamy, A., Tsai, C.-L., & Roysam, B. (2011). A broadly applicable 3-D neuron tracing method based on opencurve snake. Neuroinformatics, 9, 193–217.

Neuroinform Wearne, S., Rodriguez, A., Ehlenberger, D., Rocher, A., Henderson, S., & Hof, P. (2005). New techniques for imaging, digitization and analysis of three-dimensional neural morphology on multiple scales. Neuroscience, 136, 661–680. Xiao, H., & Peng, H. (2013). APP2: automatic tracing of 3D neuron morphology based on hierarchical pruning of a gray-weighted image distance-tree. Bioinformatics, 29, 1448–1454.

Yang, J., Gonzalez-Bellido, P.T., & Peng, H. (2013). A distance-field based automatic neuron tracing method. BMC Bioinformatics, 14. doi:10.1186/1471-2105-14-93. Zhou, Z., Sorensen, S., Zeng, H., Hawrylycz, M., & Peng, H. (2014). Adaptive image enhancement for tracing 3D morphologies of neurons and brain vasculatures. Neuroinform, 13, 153–166.