2.1.

Acquisition of Confocal Image Stacks

The datasets used in the present study were in part published previously in a different context.¹⁴ Images were acquired with an upright Zeiss LSM Pascal confocal microscope. Dendritic trees were visualized using a 40x water immersion objective lens (0.9NA, Zeiss). A 63x water immersion objective lens (0.95NA; Zeiss) at 4x scan zoom was used to image individual dendritic segments at high resolution. Up to 40 images were recorded per stack ($\mathrm{XY}:512\times 512\text{pixel}$, $0.11\mu \text{m}/\text{pixel}$ and $0.07\mu \text{m}/\text{pixel}$; $z$-steps: 3 μm and 0.5 μm respectively; Videos 1 and 2).

2.2.

Manual Quantifications

To verify the capability and correctness of the SpineLab tool to reconstruct neuronal dendritic branching and spines, four model cells were generated using NeuGen,¹⁵ transferred to a volume image using NeuRA2,¹⁶ and reconstructed with SpineLab, Neurolucida software 10.0,¹⁷ and Neuronstudio 0.9.92 (CNIC Mount Sinai School of Medicine, NY¹⁸) (Fig. 1). In addition, a dendritic segment containing spines was generated using the above-mentioned method. These datasets were analyzed by independent investigators (Gerlind Schuldt and Nadine Zahn).

Fig. 1

Validation of SpineLab reconstruction using model neurons: (a) volumetric projection of a model neuron generated with NeuGen ¹⁵ and NeuRA2;¹⁶ (b) reconstruction of the model neuron with SpineLab; (c) total dendritic branch length of model neurons reconstructed by SpineLab, Neurolucida, and Neuronstudio. Values in μm; SEM, standard error of the mean.

The dendritic tree of individual GFP-expressing dentate granule cells^19,20 was manually reconstructed using Neurolucida and Neuronstudio. Confocal image stacks of individual dendrites at higher magnification containing dendritic spines were assessed manually on 3-D image stacks of dendritic segments using the Zeiss LSM image browser to navigate through the stacks, as previously described.^14,21,22 All dendritic protrusions were counted as dendritic spines, regardless of their morphology. Images from control and denervated dentate granule cells^14,22 were analyzed blind to experimental condition to ensure unbiased observation. For each segment, a defined distance ($\approx 30\mu \text{m}$) from a dendritic branch point was analyzed and all spines were counted (red dots in Video 2). The spine density per μm was calculated based on these results. All results were compared with partially automatic reconstructions by the same independent investigators using SpineLab.

2.3.

SpineLab Software Tool

SpineLab is based on the morphology reconstruction software NeuRA2,¹⁶ which is an optimized and extended version of the award-winning Neuron Reconstruction Algorithm NeuRA.^11–13 SpineLab was developed for an automatic or partially automatic extraction of 3-D data, such as morphometric data of neurons, from image stacks suffering from reduced $z$-resolution. SpineLab combines simple noise reducing and segmentation operators, like median-filtering or global segmentation,²³ as well as the inertia based anisotropic diffusion filter, introduced in Refs. 11–13, together with the possibility to extract centerlines, using Telea’s Augmented Fast Marching Method.²⁴

SpineLab enables the user to modify the reconstruction manually via a highly developed graphical user interface. It was written in C++, using the Qt-GUI-Framework,²⁵ and is therefore available for Mac, Linux, and Windows systems. All methods integrated in SpineLab are implemented highly optimized, yielding a software tool, which requires few resources and runs reliably and fast on today’s computers.

2.4.

Pre-Processing of Three-Dimensional Image Stacks

In a first pre-processing step, SpineLab applies a median filter²³ with a small neighborhood size to all original images in order to reduce noise. Thereafter, the image stacks are projected along the $z$-axis using the maximum norm

2.5.

Generation of Neuronal Feature Skeletons

The projected image stacks [Fig. 2(a)] are segmented [Fig. 2(b), for details see Ref. 23] and the centerlines of the segmented regions are extracted using Telea’s Augmented Fast Marching Method²⁴ [Fig. 2(c)], which collapses the boundaries of the segmented image parts. The centerlines are subsequently transformed into the so called neuronal feature skeleton—a graph²⁶ consisting of nodes and edges linking the nodes [Fig. 2(d)]. To obtain the feature skeleton from the centerline, every pixel belonging to the centerline is considered to be a skeleton node. The nodes belonging to adjacent pixels are linked with edges. Nodes with one or three neighbors represent end points of spines and dendrites or branching points, respectively. Nodes with exactly two neighbors are important to approximate the curvature of dendrites or spines. However, a subset of these nodes can be removed if the angle between the two adjacent edges is large (i.e., $>135\mathrm{deg}$). This simplification yields a neuronal feature skeleton, which consists of a reasonable number of nodes [Fig. 2(e)]. This feature skeleton is 2-D, since the original image stack was projected along the $z$-axis. In a final step, the $z$-coordinate

Fig. 2

Reconstruction of GFP-labeled neurons using SpineLab: (a) through (d) Projected image stacks of a GFP-expressing dentate granule cell; (b) Segmentation using a threshold method in SpineLab’s graphical user interface; (c) Using Telea’s Augmented Fast Marching Method the centerlines of the segmented region are generated; (d) The centerlines of the regions of interest are selected by rolling over with the mouse. Scale bars: 10 μm; (e) To ease separation and to confirm correct reconstruction of individual neurons, SpineLab offers a real-time $z$-plane view, in which the investigator can correct the precise position of the centerline in all spatial directions. The blue squares denote the skeleton nodes whereas the green lines denote the skeleton edges. The eight labeled nodes are arbitrarily chosen to highlight the $z$-plane view. Scale bars 2 μm; (f) The dendrogram of the reconstructed neuron. Different colors indicate different branch orders; (g) Examples of values computed by SpineLab. The particular branches and angles are shown in Fig. 2(f). (Video 1 QuickTime, 2.8 MB) [URL: http://dx.doi.org/10.1117/1.JBO.17.7.076007.1].

2.6.

Analyzing the Neuronal Feature Skeleton

The neuronal feature skeleton can be used to extract additional information from imaging datasets—in particular spine length and spine volume. To obtain this information, the feature skeleton G is defined as a special graph with following attributes: (1) G is connected and contains no cycles (i.e., G is a tree²⁶); (2) G is weighted,²⁶ and the weights symbolize the real length in micrometers (i.e., every edge of the graph is attributed with its real length); (3) one path of G (usually the longest path in the graph) represents the dendrite; (4) branches connected to the dendrite represent single spines. Spines in $z$-direction were not included in this analysis.

The volume of single spines can only be calculated if they are oriented in the $x$- and $y$-axes. Spines are cut away from the dendrite on the image level and reconstructed via the Neuron Reconstruction Algorithm NeuRA^11–13 in the version NeuRA2,¹⁶ where the originally used Marching-Tetrahedron mesh generator²⁷ was replaced by the faster Marching-Cubes-Algorithm.²⁸ The volume of these spines can easily be calculated by using Gauss’ divergence theorem.²⁹

3.1.

SpineLab

One of the central problems in cell morphology reconstruction using transgenic mouse lines, which express fluorescent markers for visualization of neuronal morphology, is the fact that single- or multi-photon image stacks usually contain several cells, one of which should be reconstructed (Fig. 2, Video 1). This is one of the major problems automatic software-based reconstructions face. SpineLab is a lightweight (e.g., only 15 MB of hard disk space and low memory resources required) and easy to handle software for partially automatic cell reconstructions, which supports investigators with a set of automatic tools to identify and reconstruct neuronal morphology. All necessary parameters for noise reduction, segmentation, and centerline-extraction can be adjusted in real-time by the operator.

3.2.

Reconstruction of Dendritic Trees Using SpineLab

We have used SpineLab to reconstruct the morphology of the dendritic trees of individual cultured GFP-expressing neurons (Fig. 2). After applying a median filter,²³ the image stacks were projected along the $z$-axis using the maximum norm [Fig. 2(a); see Sec. 2] and segmented using a threshold method, which is reasonable, since the data resolution in $z$-direction is low compared with the horizontal resolution; every pixel of the projected image with a higher intensity than the selected threshold is colored orange and every pixel with a lower intensity than the selected threshold is colored black,²³ as shown in Fig. 2(b). The centerline of the segmented region [Fig. 2(c)] was then extracted by using Telea’s Augmented Fast Marching Method,²⁴ which generates centerlines from segmented 2-D images by collapsing the boundaries of the segmented parts. Parts, which belong to the cell of interest, can be selected by the investigator by defining regions of interest with the computer mouse [Fig. 2(d)]. In addition, SpineLab provides a real-time $z$-plane view supporting separation and precise position of centerlines, and therefore the identification and separation of individual dendritic branches belonging to different neurons [Fig. 2(e)]. The resulting 3-D neuronal feature skeleton is a tree in the sense of graph theory, which means it is a connected graph without cycles,²⁶ for which the root denotes the soma. This enables the visualization of the reconstructed dendritic tree as a dendrogram [Fig. 2(f)] and subsequent assessment of various morphometric parameters of interest. Starting at the soma, the lengths and orders of the dendritic branches and the angles between branches emerging from identified branching points can be automatically computed [Fig. 2(g)]. Using this approach, a trained investigator can reconstruct individual cultured dentate granule cells within $\approx 10\min$, i.e., an approximate total dendritic length of 1000 μm.

3.3.

Validation of Dendrite Reconstruction Using SpineLab

We validated the reliability and accuracy of SpineLab in dendritic reconstruction in two steps: First, model neurons with known dimensions were generated using NeuGen¹⁵ [Fig. 1(a)] and reconstructed with SpineLab [Fig. 1(b)], Neurolucida¹⁷ and Neuronstudio¹⁸ software. Second, we compared the results of SpineLab reconstructions of GFP-expressing neurons with the results of Neurolucida and Neuronstudio. We observed differences of $\le 1.5\%$ while reconstructing model cells [Fig. 1(c)] and differences of $\le 6\%$ between the reconstructions of GFP-labeled neurons done with SpineLab compared to Neuronstudio and Neurolucida software (Fig. 3). Taken together, we conclude that SpineLab is a reliable tool to determine dendritic branch lengths in 3-D image stacks, especially since SpineLab yielded accurate results in reconstructing the model neurons (which was also the case for Neurolucida and Neuronstudio-software).

3.4.

Reconstruction of Identified Dendritic Segments and Dendritic Spines

Next we assessed the ability of SpineLab to reconstruct and trace dendritic spines. In particular, we were interested in evaluating how many of the present dendritic spines could be automatically identified (Fig. 4). To obtain the neuronal feature skeleton from the projected image [Fig. 4(a)], the same approach was used as in the case of reconstructions of the entire dendritic tree of neurons: A centerline of the projected image was calculated and transformed to the neuronal feature skeleton, as described (see Sec. 2). The longest path of the resulting feature skeleton is marked as dendrite, which is correct for most images and can be changed by the operator if necessary [Fig. 4(c)]. The endpoints of the spines are detected reliably by the automatic approach of SpineLab, whereas the startpoints of the spines have to be corrected manually in most cases. The reconstruction of each dendritic segment can be saved as a spinogram [Fig. 4(d)].

3.5.

Spine Density Analysis Using SpineLab

Since dendritic spines, especially the neck of spines, are close to and in part beyond the resolution of confocal microscopy, the reconstruction algorithm described above did not detect all spines. To improve the automatic analysis, disconnected skeleton parts, which are often spine heads, were automatically connected to the dendrite by adding the shortest edge, which connects the dendrite and the spine head. Nevertheless, due to low resolution in $z$-direction, spines oriented in the $z$-axis, were detected only occasionally. We believe that this reflects the rather low detection rate of $31\pm 2.1\%$ ($n=10\text{segments}$) using the initial automatic approach in comparison to manual assessment of all spines in three-dimensions [Fig. 5(a); Video 2]. Regardless of these limitations, SpineLab significantly accelerated the analysis of spine densities since a considerable fraction of spines were readily identified and marked by the automatic approach of the software.

3.6.

Spine Density Changes Using SpineLab

It has been suggested in a consensus paper by several labs that only spines exceeding the dendrite laterally should be analyzed in order to minimize errors, due to the low $z$-resolution of the image stacks.³⁰ In fact, the reliability of SpineLab to detect the subset of lateral spines in dentate granule cells was considerably higher [$>80\%$; Fig. 4(c)], in comparison to detecting spines in all three dimensions (Fig. 5). In order to base the analysis of lateral spines on objective criteria, a function was included, which enables setting a lateral angle in which all spines will be considered. Using this specification, we reasoned that the reliability of assessment of changes in spine density by SpineLab should be particularly high if the same dendritic segment is visualized repeatedly. Thus SpineLab should be able to detect changes of spine density without the need of time-consuming and laborious post-hoc manual corrections. To test this hypothesis, dendritic segments of untreated controls and partially denervated dentate granule cells were analyzed.^14,22 In this earlier studies we were able to demonstrate that denervation leads to a reduction in spine densities of about 30% to 40% at seven days post-lesion. Indeed, the automatic pre-processing approach of SpineLab reliably detected these changes in spine density [$-37\pm 0.1\%$; $p=0.55$ in comparison to manual assessment; Fig. 5(b)]. Accordingly, we were able to confirm our previous result on denervation-induced changes in spine density using an objective computer-based approach, i.e., the SpineLab software.

Taken together, we conclude that while automatic spine detection of SpineLab fails to identify all dendritic spines and requires manual correction, it identifies a considerable portion of all spines, in particular lateral spines,³⁰ and thus appears suitable for detecting relative changes in spine density. This makes SpineLab an attractive tool for screening huge time-lapse imaging datasets for spine density changes in a rather short period of time, specifically compared with manual assessments. In case a difference is detected, a manual post-hoc analysis could be performed to verify these results.

3.7.

Spine Length and Spine Volume Analysis

One of the major advantages of SpineLab reconstruction is that the neuronal feature skeleton can be used to extract additional information from imaging datasets. In particular spine length and spine volume are considered important functional features of dendritic spines (for a recent review on spines and their function see Ref. 31). SpineLab automatically measures the length of spines and provides a function for volume calculation of selected spines. In Fig. 6(b) the surface reconstruction, on which the volume calculation is based on (for details see Sec. 2) is shown for some spines. Naturally, these extracted values are limited by the resolution of the confocal laser scanning microscope. However, their automated extraction again provides the opportunity to use this approach to assess relative changes in spine length and spine volume of large numbers of spines under various experimental conditions, as long as the same dendritic segment is imaged with the same parameters at the same microscope.

3.8.

Validation of Spine Reconstruction

To provide further evidence for the accuracy of SpineLab to detect dendritic spines, we generated and reconstructed a model dendritic segment studded with 57 equally distributed spines. The automatic approach detected the dendrite ($\text{error}\approx 1.2\%$) and $45(\approx 80\%)$ of the spines correctly. As expected, spines oriented in $z$-direction were not detected. After manual correction, all spines were reconstructed correctly with an average error of $\approx 0.3\%$ per spine. These results show that SpineLab is a suitable tool to accelerate detection and analysis of dendritic spines.