ExplantAnalyzer: An advanced automated neurite outgrowth analysis evaluated by means of organotypic auditory neuron explant cultures

BACKGROUND
Neuronal outgrowth assays using organotypic explant cultures are commonly utilized to study neuroregenerative and -protective effects of drugs such as neurotrophins. While this approach offers higher organized tissue compared to single cell cultures and less experimental effort than in-vivo studies, quantitative evaluation of the neuronal network is often time consuming. Thus, we developed ExplantAnlayzer, a time-saving high-throughput evaluation method, yielding numerous metrics to objectively describe neuronal outgrowth.


NEW METHOD
Spiral ganglion explants were cultured in 24-well plates, mechanically fixed in a collagen matrix and immunolabeled against beta-III-tubulin. The explants were imaged using a fluorescent tile-scan microscope and resulting images were stitched. The evaluation was developed as an open-source MATLAB routine and involves several image processing steps, including adaptive thresholding. The neurite network was eventually converted to a graph to track neurites from their terminals back to the explant body.


COMPARISON WITH EXISTING METHOD(S)
We compared ExplantAnlayzer quantitatively and qualitatively to common existing methods, such as Sholl analyses and manual fiber tracing, using representative explant images. ExplantAnlayzer is able to achieve similar and as detailed results as manual tracing while decreasing manual interaction and required time dramatically.


RESULTS
After an initial setup phase, the explant images could be batch-processed altogether. Bright bundles as well as faint fibers were reliably detected. Several metrics describing the outgrowth morphology, including total outgrowth, neurite numbers and length estimations, as well as their growth directions, were computed.


CONCLUSIONS
ExplantAnalyzer is a time-saving and objective method for an in-depth evaluation of organotypic explant outgrowth.


Introduction
In organotypic cultures, primary ex-vivo tissue is used to address numerous scientific questions. This biological model system bridges the gap between cell-line culture and in-vivo studies (Al-Ali et al., 2017). While cell culture allows to investigate homogenous and determinate cells in large numbers, it often cannot model complex cellular interactions (Humpel, 2015). On the other end of the spectrum, in-vivo models offer the whole physiological context but need considerably more effort to be maintained and underlie restrictions related to animal experimentation. Therefore, organotypic explant cultures present a valuable balance of these factors, as differentiated tissue can be cultured in its local cellular environment while external conditions can be precisely controlled (Gähwiler, 1988).
One of the applications of organotypic cultures is the study of nerve regeneration and -growth. Tissues used for these studies include dorsal root ganglia (Gomes et al., 2016;Kim et al., 2021;Livni et al., 2019), spinal cord slices (Romeo-Guitart et al., 2020;Santos et al., 2016), the spiral ganglion (Sun et al., 2020;Szobota et al., 2019) and retinal ganglion cells (Binley et al., 2016;Schaeffer et al., 2020). One of their major aims is to find treatments to increase the number of outgrowing axons and dendrites and to elongate and guide fibers towards their targets. The presented approach was developed to investigate drugssuch as brain-derived neurotrophic factor (BDNF) or neurotrophin-3 (NT-3)to enhance axonal regeneration and sprouting of spiral ganglion neurons (SGNs). These bipolar neurons retract their peripheral axons when their sensory target, the hair cells, are damaged or absent, e.g. after exposure to acoustic trauma, an ototoxin or due to ageing, and survive as monopolar neurons over decades in humans (Glueckert et al., 2005). Cochlear implants are commonly used to replace the hair cell function by directly stimulating the auditory neurons. The spatial gap between the electrodes and the targeted neurons limits specific stimulation . Therefore, a desired strategy to overcome this gap is to induce neurite extension towards the electrode surface.
One of the major difficulties of neurite outgrowth assays using organotypic explants lies in the evaluation of the acquired images (Al-Ali et al., 2017). Several software packages and plugins are available to assess neurite outgrowth in general, but few are specifically dedicated to organotypic explants. Existing algorithms either require manual processing steps, are not well suited for organotypic cultures with high numbers of neurites or deliver only limited data on outgrowth morphology. These algorithms can be roughly divided into two groups: Sholl analysis-based methods and neurite tracing methods. Sholl analysis-based methods are modifications of a procedure first published by Sholl (1953). A pattern of concentric circles with incrementally increasing diameters is drawn around the soma of a neuron, or in this case an explant, and the intersections of each circle with the neurites are counted. The resulting Sholl profile illustrates the incremental increase and decrease of neurite intersections with growing distance from the soma or explant. Values such as the number of neurites exiting the explant, the distance from the explant with the highest number of fiber intersections or the radius of the largest circle crossing a nerve fiber are commonly evaluated. Implementations of this method are integrated in software packages such as ImageJ or are available as ImageJ plugins, specifically designed for explant evaluation, for example Neurite-J (Torres-Espín et al., 2014) or Neurite Length Index Kramer et al., 2017). As the Sholl analysis was initially developed to determine ramification of neurite networks, research questions such as pharmacological effects on neurite elongation or neuronal survival are difficult to answer. The second group of methods involves tracking of individual neurites. Therefore, they provide more detailed data on single fibers but often require significant manual interaction. Numerous commercial software packages, e.g. Neurolucida (MBF Bioscience), Amira (Thermo Fisher) and Imaris (Bitplane), as well as open-source software such as Simple Neurite Tracer (Arshadi et al., 2020), NeurphologyJ (Ho et al., 2011), NeuriteTracer (Pool et al., 2008), or NeuronJ (Meijering et al., 2004) are available. Other occasionally used methods include the measurement of the area covered by outgrown neurites (Cregg et al., 2009;Deister and Schmidt, 2006), the area within a polygon enclosing all neurites (Sun et al., 2016) or the immunolabeled area within certain rings around the explant (Gaublomme et al., 2013). Another elegant solution was provided by Weaver et al. (2003), using a ridge tracking algorithm.
Here we present ExplantAnalyzer, an easy-to-use MATLAB application to analyze neurite outgrowth in organotypic explants. The aim was to develop a procedure that can evaluate neuronal outgrowth in a fully automated manner and combine neurite tracing methods with user independent large scale outgrowth measurements. With this application, the manual workload and observer bias should be considerably reduced. The required files are available free of charge and can be found as Supplementary Data or on https://github.com/DominikSchmidbauer/Explan tAnlayzer. To install this package, the.zip file has to be extracted and all files need to be kept in the same folder. For validation of our tool and for comparison, we focused only on readily available, non-commercial tools which provide comparable metrics. Therefore, our method was compared to the manual tracing tool Simple Neurite Tracer (Arshadi et al., 2020) and the two Sholl plugins Neurite-J (Torres-Espín et al., 2014) and Neurite Length Index Kramer et al., 2017).
C57BL/6N (Charles River, Sulzfeld, Germany) mice of both sexes were used at postnatal days 6 and 7. All mice were bred at the animal facility in Innsbruck with a 12 h dark, 12 h light cycle and unlimited access to food and water. The animal studies conformed to the Austrian guidelines for the care and use of laboratory animals. After rapid decapitation, both inner ears were extracted under a stereo microscope and transferred to ice cold Hanks' balanced salt solution (14025-050, Thermo Fisher Scientific). With the aid of fine forceps (Dumont, Switzerland), the otic capsule was carefully extracted and the modiolus was separated from the residual bony labyrinth. Both the stria vascularis and the organ of Corti were torn off in a spiral movement from base to apex. Using micro scissors, the remaining modiolus containing the spiral ganglion was cut into three half turns. Each half turn was freed from bony debris and most of the cochlear nerve, leaving only the spiral ganglion. All half turns were subsequently cut into six equally sized explants per cochlea (Fig. 1). Each explant was then transferred to one prefilled culture well, utilizing a 100 µl pipette tip and a volume of ~10 µl. The plate was placed in the incubator and the explants were manually positioned into the center of the wells, using a sterile blunt hook, and subsequently cultured for 96 h at 37 • C and 5% CO 2 . To assess the ability of our new method to measure the outgrowth response to the neurotrophins BDNF and NT-3, 15 explants per condition were cultured. For one treatment group, three explants (from the apical, the middle and the basal turn) were taken from each of five inner ears.
Immediately before the culture was stopped, a gelling agent was prepared as previously described (Roy et al., 2010). Briefly, 50 µl 10x Basal Medium Eagle (B9638, Sigma-Aldrich) and 50 µl 2% sodium bicarbonate were added to 500 µl of collagen I from rat tails (354236, Corning), mixed and kept on ice. The culture medium was carefully removed with a Pasteur pipette and the culture plate was placed under a laminar flow cabinet for 2 min to let most of the remaining liquid evaporate. One drop of the gelling agent was then dropped directly on top of each explant, using a 21-gauge needle and a 1 ml syringe. This ensured that the explant was not flushed away but completely covered with the gelling agent. To let the gelling agent polymerize, the culture plate was placed into an incubator for 15 min at 37 • C. The collagen matrix secured the explant and the outgrown neurites mechanically and facilitated following procedures, as the risk of explant detachment from the bottom of the well was greatly minimized. Collagen was used, as it gels already at physiological temperatures and does not induce substantial background fluorescence. Subsequently, tissue was fixed with 4% formaldehyde for 1 h at room temperature (RT) and washed with phosphate buffered saline (PBS).

Immunohistochemistry
In order to saturate nonspecific binding sites and to permeabilize the tissue, 200 µl of blocking solution, composed of 7 ml PBS, 30 µl Triton X-100 (10789704001 Roche, Mannheim, Germany,) and 3 ml normal donkey serum (NDS, S30-100 ml, Merck Millipore, Burlington, USA) was added to each well for 2 h at RT. Rabbit polyclonal anti-beta-IIItubulin antibody (Tuj1, ab18207, Abcam, UK) was diluted 1:1000 in 0.1 M PBS, 0.3% Triton X-100% and 1% NDS. The explants were incubated with the antibody solution overnight at 4 • C and then for 2 h at RT. After three washes with PBS, Alexa Fluor 546 donkey anti-rabbit IgG antibody (A10040, Thermo Fisher Scientific) was diluted 1:1500 in 0.01 M PBS with 1% ml NDS and incubated for 2 h at RT. After thorough washing in PBS, one drop of Vectashield Antifade Mounting Medium with DAPI (H-1200, Vectorlabs, Burlingame, USA) was added on top of each explant.

Image acquisition and stitching
Images were acquired with a DMi8 inverted fluorescence microscope (Leica, Wetzlar, Germany), equipped with a motorized stage and a 2048 × 2048 pixel ORCA-Flash4.0 (Hamamatsu Photonics, Hamamatsu, Japan) camera. A template for 24 well plates was used to acquire an overview of all wells with a 10x lens, followed by selecting a rectangular region of interest for each explant, enclosing all neurite extensions. The illumination was set to 100% for both channels, while the exposure time was 100 ms at 395 nm and 500 ms at 550 nm excitation wavelength (SPECTRA X light engine, Lumencor, Beaverton, USA). 16bit color depth was used to make very faint neurites distinguishable. The actual images were then taken with a 20x lens (HC PL FLUOTAR L 20x/0.40 DRY), enabled shading correction and a 20% overlapping tile scan. This overlap was necessary to guarantee precise stitching of neurites later on. This setup resulted in a pixel size of 0.328 µm.
The resulting.lif (Leica Image File Format) files contained each channel separately. An ImageJ script (Image Conversion Tools, Baecker, 2011, available as a supplement) was adapted to merge the channels into one RGB image, with the red channel representing the Tuj1 image and the blue channel the DAPI image. This step was necessary to precisely stitch both channels at the same time later on. The merged tile images were exported as 16-bit TIFF files. Those images were then imported into the freeware Image Composite Editor (2.0.3, Microsoft Corporation, Redmond, USA). The rectangular scanning region allowed us to reconstruct the tiles' positions. Camera motion was set to planar motion and the tile order was adjusted according to the microscopes tile scanning sequence. The final stitched image was exported as a 16-bit TIFF file again. The stitching process took approximately 30-60 s, depending on the size of the final image. This rather complex workflow resulted in the highest possible resolution and accuracy regarding correct neurite overlapping at tile borders, while keeping processing time reasonably fast. Any other workflow, resulting in a precisely stitched image of the whole outgrown explant, containing 16-bit DAPI and Tuj1 channels, is also suitable for the actual image processing described below.

Image processing
The stitched images were processed in MATLAB (2019a, The Math-Works, Natick, USA) afterwards. In order to run the script, MATALB's Image Processing Toolbox has to be installed. There are two main scripts: the batch.m script and a function named ExplantAnalyzer.m, which is called by the batch script. The batch function enables the user to run the analysis fully automatically for all images at once. All relevant image processing parameters can be adjusted in the batch script according to individual needs and are then applied to each image.
The main idea was to generate a precise skeleton using different image processing techniques and then convert this skeleton to a graph (Fig. 2). A graph is a topological structure which models the relation and connection between nodes and edges (Fornito et al., 2016). It can be mathematically described by an adjacency matrix. This matrix contains all nodes (start-, end-or branch-points) and edges (connections between the nodes) as well as their weight, which is the length in this case. This transformation of the neurite network facilitates as well as accelerates further analysis such as finding the shortest path from a neurite terminal to the explant body. The upper row depicts preparation, culture, and immunohistochemistry of spiral ganglion explants. Each cochlea was dissected into 6 spiral ganglion explant pieces and cultured for 96 h. Hereafter, the medium was aspirated and the explants were immobilized in a collagen matrix to protect the tissue from mechanical damage during later procedures. Subsequently, the explants were fixed and immunohistochemically stained for beta-3-tubulin and nuclei counterstained with DAPI. The lower row shows the image acquisition process. The explants were tile scanned with a fluorescence microscope. The two channels of each tile were then merged into a RGB image and stitched, resulting in a precisely stitched image with continuous neurites. Scale bar: 1 mm.
In a first step, the.lif image was loaded and separated into its two channels. The DAPI channel representing all cell nuclei was used to determine the size and the boundary of the explant body without outgrown neurites. Therefore, the image was binarized by Otsu's thresholding method, only the largest object was kept and holes within the explant's area were filled. The explant was then dilated two times by a constant value. The amount of dilation was set so that the one-time dilated explant encloses the explant body tightly. For a detailed explanation please refer to the Image analysis parameter adjustment section. The smaller dilated explant area was later subtracted from the Tuj1 image to improve thresholding in the vicinity of the explant body. Otherwise, the adaptive threshold would have been increased by the bright explant body in this area and therefore some fibers would have been omitted. The boundary of the larger dilated explant area was later defined as the proximal reference for neurite length measurement.
Either the median intensity value of the Tuj1 channel or an arbitrary value can be chosen to be subtracted from the whole image to remove background. If the background is rather constant among the image set, a fixed value is recommended. Otherwise, the median intensity is able to equalize fluctuating backgrounds. A more detailed description is provided in the Image analysis parameter adjustment section. Subsequently, the image was high boost filtered to enhance the signal to background ratio and to increase the contrast of very faint fibers. A high boost filter amplifies the high frequency components, such as the transition from background to a neurite, without affecting low frequency components, such as the background or the explant body. A median filter was applied to reduce noise. To ensure that the bright explant body does not influence thresholding, the smaller dilated explant body was subtracted from the Tuj1 channel. Binarization of the image was performed by the adaptthresh and imbinarize functions. An adjustable window moved over the whole image and set a threshold value for each individual pixel according to the mean intensity values within the window. The size of the window had to be determined once but was then unchanged for all explants. Its size should be chosen in a way that it ideally contains only one single fiber or fasciculated bundle while still including a portion of the background. Consequently, bright fibers close to the explant's body as well as faint fibers in the periphery were correctly separated from the background. Only the largest connected area was kept and all smaller isolated areas were discarded, resulting in a binarized image of the neurites. The area of all fibers, the area of the convex hull enclosing all fibers and the neurite covered area were calculated as additional outgrowth measures. In a following step, the binarized neurites were dilated and again eroded to make their boundaries smoother. This closing operation leads to a much smoother neurite skeleton with less spurs in the following skeletonization step. Additionally, it closes small gaps in between the neurite network and The image was binarized and the largest area was then dilated one and two times by an arbitrary dilation value. Scale bar: 1 mm. Other rows: Processing of the Tuj1 stained fibers. After some basic filter operations, the 1x dilated explant area was subtracted from the image to minimize the influence of the bright explant body on the following adaptive thresholding. The binarized neurites were then skeletonized and the 2x times dilated explant was subtracted from the skeleton to generate open ends to be used as start-points. These start-points (orange +) were then subsequently detected and the skeleton was converted to a graph (1), using an adjacency matrix. All start-points were logically connected to a virtual center-point (C). Each edge was weighted by its actual length. The shortest path was then computed for each end-point. Note that some edges could be used multiple times by backtracked neurites (2). Finally, all unused segments were removed from the graph, resulting in a tree-like structure (3). Both the graph and the tree graph were stored for the evaluation later on.
improves continuity. For skeletonization, an additional minimum spur length threshold could be set. All spurs shorter than the threshold were removed. Subsequently, the adjacency matrix and lists of nodes and edges of the skeleton were computed by the Skel2Graph3D function (Kollmannsberger et al., 2017), which is also capable of processing 2D images. The algorithm detects all branch-points of the skeleton, defines them as nodes and finds all interconnecting edges between the nodes as well as their network structure. The adjacency matrix was then weighted by the Euclidean length of each edge. The Euclidean length is a measure for the actual length and takes diagonal pixels correctly into account. Therefore, the length of curved neurites was also measured precisely. Using the weighted adjacency matrix, a graph was generated. Subsequently, each intersection of the skeleton with the larger dilated explant boundary was defined as a start-point for neurite tracing. This step was necessary, as the size of the explant in the Tuj1 channel was larger than in the DAPI channel. All end-point nodes were then connected to a virtual common center-node. The edges of these connections were attributed with a length of zero and were therefore only logically connected to the center-node. This point served as a common target-point for the following step without altering neurite length measurements. Finally, the graph was reduced to a tree-like structure. This was done by finding the shortest path from each neurite terminal (end-point) to the center-point via the start-points, using a backtracking algorithm. A validation image of each explant was generated after image analysis. It contained the outlines of the explant and its two dilated boundaries, segmentation boundaries of the neurites, the skeleton and all end-points. This enabled a visual inspection of the segmentation quality. In case some background speckles were present and interfered with proper segmentation, these speckles were manually removed in Photoshop (CC 2019, Adobe, San Jose, USA). The edited image was then processed again. The results for each explant were stored in.mat MATLAB workspace files, termed identically as the file name of the image. This procedure allowed straightforward file handling and sorting afterwards.

Graph evaluation
In order to evaluate the results, the individual.mat files can be organized manually in folders or automatically in MATLAB. As this process is highly dependent on the research questions, we provide a simplified script (simple_statistics.m) to evaluate and gather all data within a folder. The size of the explant was stored within the.mat files and could be used directly. All distances from each end-point to the boundary of the explant body (start-points) were calculated by the shortestpathtree function, resulting in a shortest path length (SPL) for each end-point. Mean, median and maximal values of all SPLs were computed. The number of neurite terminals was calculated by simply counting the number of end-points. The number of nodes which were connected to more than two other nodes represent the number of branch-points. Additionally, three different measures describing the total outgrowth were computed: the total length of the whole skeleton/ graph (see 1, Fig. 2), the total length of the backtracked tree graph (3) and the sum of all SPLs (2). The latter one can contain some edges multiple times, especially close to the explant body. Finally, the average growing direction of each neurite regarding the upwards facing vertical image axis was determined. This outcome may serve to detect directed growth of neurites. In the end, all explants with a total outgrowth below an arbitrary threshold may be removed from the analysis. Explants exhibiting only little outgrowth might have been damaged during preparation or did only contain very few neurons and may therefore be excluded. Additionally, a Sholl analysis was implemented and can be activated by setting a Sholl step distance. The number of intersections as well as the pixel count of all intersections per ring were stored. Furthermore, a script (view_graph.m) to visualize the graph structure on top of the binarized neurites and to show individual SPLs is provided.
To evaluate the ability of ExplantAnalyzer to measure changes in the outgrowth morphology in presence or absence of the neurotrophins BDNF and NT-3, the total outgrowth, the number of end-points and the mean SPL of 15 explants per condition were compared in Prism 9 (GraphPad, San Diego, USA). The mean SPLs were statistically tested using an ANOVA test with Dunnett's post-hoc multi-comparison against control. As a normal distribution could not be assumed (Shapiro-Wilk test), the total outgrowth and the number of end-points were compared using the Kruskal-Wallis test with a Dunn's post-hoc multi-comparison against control.

Comparison to other methods
To compare different approaches for organotypic explant outgrowth evaluation, other algorithms were tested as well. Ten explants with varying morphologies were chosen to test and compare different tools. These explants were imaged and stitched beforehand as previously described. Three other methods which were readily available and worked with our images were tested. All of these methods were available as ImageJ plugins. The Neurite Length Index Kramer et al., 2017) is based on a Sholl analysis, which counts intersections of expanding circular rings and binarized neurites. The tool itself does not include a thresholding step, so the thresholding was performed by manually finding a suitable value. Intersections between one and the next ring are multiplied by the ring distance and added up to representative value for the total outgrowth. The radius step size was set to 10 pixels. The same value is also used to estimate the number of fibers within a fiber bundle, e.g. a 40-pixel wide bundle would be counted as 4 intersections. Similarly, the Neurite-J plugin (Version 1.1, Torres-Espín et al., 2014) is also based on a Sholl analysis but with the concentric rings being shaped like the boundary of the explant body. The implemented guide leads through the entire processing step-by-step. The default step size of 25 µm was used. In order to compare our results to manual neurite tracings, Simple Neurite Tracer (SNT, Version 3.1.109, Arshadi et al., 2020) was chosen, as it offers a path finding algorithm between two points to facilitate tracing. Whenever fibers branched, another sub-path was added. As the individual neurite paths could often not be determined exactly, the visual morphology was reproduced. Consequently, loops occurred frequently because of overlapping neurite paths and impaired the ability of the plugin to compute metrics other than the total path length of the neurite network. Therefore, to enable counting the number of end-and branchpoints, a skeletonized representation of the neurite network was generated in SNT, filtered morphologically and subsequently, the values were computed.
The accuracy of ExplantAnalyzer's length measurements was determined by comparing the measurements of 20 individual neurites. These neurites were once manually traced with Simple Neurite Tracer and then compared to the respective quantifications of the new method.
All data was statistically tested in Prism using paired, two tailed ttests.

Culture and imaging outcome
While dissection and culture methods are well established, mechanical fixation in a collagen matrix before starting the immunohistochemistry greatly facilitated handling and preserved the fibers' morphology. Furthermore, the collagen did not increase autofluorescence background significantly with our excitation/emission spectrum. After image acquisition, stitching the tile images in Image Composite Editor proofed to be a reliable method to achieve necessary precision. Despite a low density of image features, especially in the periphery, the algorithm managed to stitch almost every neurite precisely, resulting in a continuous neurite network. A precisely stitched neurite network is an important prerequisite for the following image analysis.

Image analysis parameter adjustment
Several parameters had to be adjusted according to the average explant-and outgrowth-morphology. To facilitate parameter optimization, the batch script offers the possibility to display an overview window (via the setup switch). Its activation leads to an image panel with depictions of all major image processing steps ( Supplementary Fig. 1). The images, which expose all processing steps at once in a consecutive arrangement, may be enlarged and the region of interest placed to any area of the explant. This tool shall help to identify and fine-tune certain processing steps. We recommend using a small subset of 5-10 explant images that represent various outgrowth patterns and especially different fiber densities for parameter optimization.
The first parameter that had to be adapted for the Tuj1 channel was the background subtraction value. Either an arbitrary value could be set by interactive exploration, or the median value of the Tuj1 image was used if the background subtraction value was set to zero. If the background illumination varies considerably in between a set of images, the median might be the right choice. In some cases, very faint fibers could not be separated from background noise, leading to highly ramified skeleton branches. Consequently, the background subtraction value should be slightly increased. The explant dilation value used for the DAPI channel should be chosen in a way that the boundary of the onetime dilated explant encloses the actual Tuj1-positive area of the explant tightly (Fig. 3A). For our explants, a dilation value of 25 µm was suitable to cover the explant bodies sufficiently. The value of the center element of the high boost kernel should be chosen, so that faintly stained fibers are clearly enhanced against the background while more intensely stained fiber bundles should not be completely overexposed ( Fig. 3B and  C). A value of 20 proved to be a good compromise if the images are exposed in a way that the full dynamic range of intensity values is being used without clipping bright values. The size of the median filter is less critical as this filter should only remove pixel noise. The neighborhood size determines the size of the window for adaptive thresholding. It has to be as small as possible while it must not cover thick fascicles completely. Otherwise, the fascicles would not be segmented correctly.
If the window is too large, faint fibers surrounded by bright fibers will not be detected. For our dataset, 65 µm window size was just large enough to cover the thickest bundles, although very few faint fibers remained undetected (Fig. 3D and E). Additionally, the sensitivity for adaptive thresholding can be changed if necessary but the standard value of 0.5 should work well for most cases. The neurite smooth size defines the size of the structuring element for the closing operation of the binarized neurites. A size of 1.5 µm yielded suitable results. The last value to be set was the length of spurs removed by the skeletonization algorithm. This value can be chosen arbitrarily, but the smaller the value the increasingly more neurite endings remain. 6 µm was the cutoff value which, in average, produced end-point counts closest to manual tracing (see Section 3.5 Comparison to other methods).

Image processing outcome
For a sample dataset of spiral ganglion neuron explants, average processing time was approximately 2.4 min per explant on a 2014 built Z800 workstation (HP Inc., Palo Alto, USA). Most fibers were precisely segmented, both in the immediate vicinity of the explant, as well as in the periphery (Fig. 4A). Complex and overlapping structures were also recognized well, although the course of single fibers cannot always be determined (Fig. 4B). Segmentation errors occurred mostly due to low and interrupted fluorescence signals (Fig. 4D) or densely clustered fibers in the close vicinity of the explant body (Fig. 4C). Fig. 4E shows the final graph and tree structures as an overlay on top of the binarized neurites. The neurite branches are correctly connected and end-points can be tracked back to the virtual center-node. This presentation is included in the overview panel for parameter optimization (see Supplementary Fig.  1) and can also be quickly generated by the view_graph.m script. It allows a comparison of the whole graph with the backtracked tree-like structure and reveals that not all segments were used by the backtracking algorithm. Those segments are represented as blue lines in Fig. 4E and F and were not used because they would pose as a detour for the algorithm. With increasing complexity of the outgrown neurite network, more interjacent segments are being omitted (Fig. 4F). The spur removal length turned out to be an influential factor regarding the number of end-points, as the spur lengths are approximately decreasing in an exponential manner (Supplementary Fig. 2). This means that a slight change, especially around the value of 5 µm, leads to a large change in end-point numbers. Nevertheless, the spur removal parameter is crucial to eliminate artificial terminals and should therefore be tuned carefully.

BDNF and NT-3 outgrowth assay
Both BDNF and NT-3 are known to promote neuronal survival and to enhance neurite outgrowth. Therefore, we compared neurotrophin treated cultures to unsupplemented ones with the aim of assessing the ability of ExplantAnalyzer to detect changes in the outgrowth morphology. 15 explants per condition, consisting of three explants (one from the apical, the middle and the basal turn) of each of the five inner ears were cultured for 96 h with either 50 ng/ml BDNF or NT-3, or without neurotrophic supplementation. Afterwards, all explants were processed as described. The mean total outgrowth was 36.417 mm for the control group while both BDNF and NT-3 supplementation led to a significant increase (91.383 mm, P = 0.0452 and 87.548 mm, P = 0.0146 respectively). Likewise, the number of neurite endings was considerably increased by BDNF (mean 557.13, P = 0.0084) and NT-3 (mean 606.33, P = 0.0014) over control Blue lines represent the whole graph while orange lines mark the backtracked tree graph. Note that the orange and blue lines only illustrate the topological connections and not the actual lengths of the corresponding segments. Dashed lines are the virtual connections to the common center-node. Green crosses denote start-points, red crosses mark end-points and blue Xs denote branch-points. Note that some paths were not considered for measuring the SPLs because they would act as a detour for finding the shortest paths. Scale bars: 100 µm. (E) This depiction of a rather simple radially outgrown explant demonstrates that the algorithm reliably converts neurites to a graph. (F) The more reticulated and complex the outgrown neurite network becomes, the more interjacent segments are being omitted by the backtracking algorithm. (mean 209.07). However, only a slight enhancement in the mean SPLs were detected. The mean SPL of the control group was 0.58060 mm and BNDF or NT-3 supplementation resulted in 0.66213 mm (P = 0.2767) and 0.65987 mm (P = 0.5950) respectively.

Comparison to other methods
Three other methods were tested as described above. Ten morphologically diverse explants (see Supplementary Fig. 3) were chosen to quantify and compare key metrics. One explant (Supplementary Fig. 3I) could not be processed in Neurite-J due to an error. Fig. 6 depicts the methods and their outcomes visually, while Fig. 7 compares the resulting data. Each method was only compared if the respective data was generated by this method. The total processing time (Fig. 7A) was fastest for the here presented tool (36.7 s in average), followed by the two Sholl-based methods at roughly similar times (Neurite-J: 115.7 s, NLI 141.8 s). Manual tracing took the longest (1.39 h), depending greatly on the care and precision taken to trace individual fibers. While the calculation time for one explant using the Sholl methods mainly depends on the overall image dimensions, it is strongly increasing with the complexity of the neurite network when using the other two methods. The total outgrowth (Fig. 7B) was slightly but significantly lower in ExplantAnalyzer than with manual tracing in SNT (Mean of 21.099 mm versus 25.596 mm, P = 0.0196). Especially towards explants with higher outgrowth the difference of the two methods' results increases. 20 individual neurites were evaluated to verify the accuracy of length measurements (Fig. 7C). The lengths of these neurites were significantly longer in SNT than in ExplantAnalyzer (P < 0.0001), but the mean difference was 51.68 µm, which matches the 2x dilation value of 50 µm quite closely. The number of end-points (Fig. 7D) were not significantly different (Means: ExplantAnalyzer 168.9, SNT 178.4, P = 0.3598) but exhibit a higher inconsistency in between the datasets. The evaluation of the start-points (Fig. 7E) by ExplantAnalyzer (Mean 33.8) resulted in higher numbers than by SNT (Mean 28.5 P = 0.0022). No differences in the number of branch-points (Fig. 7F) could be found (P = 0.1524). The explant body area (Fig. 7G) was compared to the results of Neurite-J as this plugin was the only one which provided such a measure. The area was significantly (P = 0.0006) higher when measured by Neurite-J (Mean 0.06700 mm 2 ) than by ExplantAnalyzer (0.08730 mm 2 ). Fig. 7H displays a comparison of the Sholl profiles of all ten explants. The Sholl profiles generated by the NLI plugin were computed with a different methodology, resulting in much higher intersection counts. Therefore, only one exemplary profile is displayed to visualize the difference.

Discussion
Here we present a fully automated evaluation method for the quantification of neurite outgrowth in organotypic explant cultures. The aim of this tool was not to achieve highest accuracy in fiber tracking, but to provide an objective method to measure multiple morphological parameters with the least amount of manual interaction and modest computing time. The outgrowth performance of organotypic explants strongly depends on the dissection process, especially on the volume and therefore the amount of contained neurons and moreover on the type and location of the explanted tissue. For example, tonotopic gradients of nerve growth factor receptors occur from the low frequency apical regions to the high frequency basal portions of the cochlear coil and even revert during postnatal development (Adamson et al., 2002;Davis, 2003). Additionally, the outgrowth pattern depends on the radial position within the spiral ganglion (Druckenbrod et al., 2020). Furthermore, compared to the in-vivo situation, only a fraction of the neurons survives the explantation process due to surgical trauma (Sobkowicz et al., 1993). Consequently, high variation results thereof and a considerable number of specimens is needed for conclusive results. Preparing and culturing organotypic explants can already be a time-consuming process but evaluating those explants often even exceeds that time. The here presented fully automated process can drastically reduce that effort, as only a few interactions at the beginning are needed to process all explants at once. Furthermore, all crucial parameters are set beforehand and are equally applied to all images, which eliminates any subjective investigator bias.
ExplantAnalyzer is not able to track back single fibers. Fibers usually fasciculate along their path, most notably in the vicinity of the explant body, but also cross each other or bifurcate (e.g. see Figs. 3A and D & 4B and C). Hence, a single fiber could not be tracked precisely and individually. This goal is also hardly achievable in a two-dimensional, widefield imaged culture, which was chosen to minimize image acquisition time and to avoid further processing steps. Therefore, neurite lengths (or rather SPLs) are mainly underestimated by the algorithm, as a single neurite could have deviated from the shortest path. An overestimation of the length is unlikely and only possible if segments leading to a shorter path are disconnected due to inconsistent staining or image artefacts. Since there is a huge variability of resprouted nerve fiber diameters that is even not constant along its length, we did not consider the fiber thickness to assess the possible number of neurites in a fascicle. Nerve fibers that terminate within a fascicle will not be included in our measurements. Considerably more effort using high throughput confocal imaging would be needed to enable three-dimensional growth and to increase image resolution to a point where single fibers can be more easily distinguished. Together with the already mentioned variability, such an approach appears to be impractical. Therefore, the assumption was made that neurites in explant cultures grow mostly radially and consequently, the shortest path is likely close to the actual path taken by a neurite. This procedure is similar to how manual estimation of neurite lengths from organotypic explants would be done. As the associated soma of a neurite cannot be determined within the explant body using widefield fluorescence microscopy, manual tracing usually starts at the fiber end-points to ensure that all terminals are included. Then, the fiber is back-traced towards the explant body (Deister and Schmidt, 2006), resulting in a path that is probably quite close to the one found by the shortest path algorithm. Therefore, both procedures lead to similar results. Another drawback of the presented method is that a distinction  Supplementary Fig. 3). As the NLI Sholl profiles were computed differently, only one profile is shown in H I. Asterisks denote significance levels of paired, two-tailed t-tests: * P ≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.001, ****P ≤ 0.0001, ns P > 0.05. between central axons and dendritic processes cannot be made. For example, the two processes, emerging from the bipolar cell body, can be clearly distinguished by diameter in histological sections. Peripheral axons are roughly half as thick in diameter as central axons. But since SGNs are not uniform in their morphology and their processes vary considerably in diameter and are often fasciculated, it is hardly possible to distinguish them reliably in organotypic explants.
The neurotrophins BDNF and NT-3 play an important role during the development of the inner ear as they promote survival and are absolutely necessary for successful innervation (Fritzsch et al., 1997). They are also powerful drugs to enhance outgrowth in SGN explant cultures (Frick et al., 2020). Using these neurotrophins, we demonstrated that ExplantAnalyzer is able to detect their positive effect on the total outgrowth of treated explant cultures (Fig. 5). Furthermore, we could show that the increased total outgrowth is more likely a result of a higher number of outgrowing neurites than of longer neurites. This supports the hypothesis that BNDF and NT-3 act mostly as promotors for neuronal survival and prevent cell death induced by the explantation trauma.
We compared our new tool to three existing methods (Table 1). SNT (Arshadi et al., 2020) was mainly chosen as a reference to validate our algorithm rather than as a practical alternative. Manual tracing makes the process laborious and consequently too time consuming for larger studies (Fig. 7A). Furthermore, it is subjected to the investigator's assessment and therefore less reproducible. Moreover, this particular plugin became laggier the more segments were traced, making it difficult to process large explants. The advantage of manual tracing is primarily the possibility to identify very faint and interrupted fibers by eye and to estimate and interpolate the most probable path, especially where fibers are crossing each other. Two plugins explicitly developed for analyzing organotypic explant outgrowth morphology were also compared to our approach. Neurite-J (Torres-Espín et al., 2014) and Neurite Length Index Kramer et al., 2017) are both based on a Sholl analysis, which is a well-established method but only produces limited data on outgrowth morphology. A major advantage of these methods is its robustness for topologically unconnected neurites; hence image quality and stitching accuracy are less critical. While ExplantAnalyzer requires all neurites to be connected to the explant body and therefore omits all unconnected segments, a Sholl analysis is able to incorporate these segments. On the other hand, Sholl based methods are more susceptible for speckles as every intersection with a foreground object is counted. Therefore, a rather clean background is necessary at the expense of time-consuming manual interaction. Background problems usually increase with a prolonged culture duration, since debris from degenerated cells or migrated glial cells from the explant body "pollute" the background (Sobkowicz et al., 1993). All three existing methods have in common, that their outcome greatly depends on the binarization threshold or how much the contrast is increased in case of SNT. A lower threshold includes more neurite segments in the Sholl analysis at the cost of more background speckles and therefore higher intersection counts. Similarly, a higher contrast in SNT results in more visible fibers for the human eye.
The comparison (Fig. 7) revealed in general a high consistency of the data produced by the three methods, considering the mechanics of each approach. The total outgrowth measured by SNT was approximately 21% higher. Two factors contribute to this fact: The measurements in ExplantAnalyzer start further away from the explant body. This was necessary to ensure proper segmentation but, on the downside, reduced the measured total outgrowth and SPLs. Furthermore, few faintly stained fibers as well as densely clustered fibers, which could have been traced by eye and hand, were not recognized by the algorithm. However, faint fibers seem to be most likely degenerated and not functional anymore and can therefore be omitted. Delicate filopodia, characteristic for the growth cone of a regenerating fiber, may also not be captured but are usually too short (~7 µm, Anderson et al., 2006;Sun et al., 2016) to influence the outcome considerably. As already mentioned, the mean SPLs were also shorter if measured by ExplantAnalyzer due to the dilated start-line. However, the average difference matched well with the 2x dilation value. The number of end-points showed a bit more inconsistency. As pointed our earlier, the spur removal length is a quite sensitive value for the number of end-points and small differences can have a big effect on end-point counts. On the other hand, manual tracing is subjective and prone to fatigue related errors. Especially tracing of short neurite endings highly depends on the investigator's attention and therefore, manually traced end-points do not provide an objective measure. The difference in start-point numbers may be explained again by the dilated start-line as usually the number of fibers increases in the vicinity of the explant body towards the periphery. The shift in explant sizes results from the fact that Neurite-J uses only a morphologically processed image of the Tuj1 channel to estimate the area, while the ExplantAnalyzer utilizes the DAPI channel. If an explant features thick fiber bundles exiting the explant body, it is difficult to estimate the boundary between the explant body and outgrown fibers by only using the Tuj1 channel. The Sholl profiles exhibit in general similar curves. In Table 1 Comparison of four different methods for analyzing neurite outgrowth. Four methods were compared regarding their output metrics. Indicated values can be measured by either tracking each neurite terminal back to the explant body (*) or by assessing the skeleton using morphological methods (**). (***) Further measures can be deducted from Sholl profiles, such as the radius of the largest circle with an intersection, the radius of the circle with the highest number of intersections and the number of fibers exiting the explant body. Advantages and disadvantages represent the experiences and opinions of the authors after familiarizing and using the tools with several explants. most cases, the curve of SNT revealed more intersections than the one of ExplantAnalyzer. Like the higher total outgrowth, this can be attributed to the ability to distinguish more fibers by eye. Additionally, the chosen threshold (Neurite-J and NLI), the placement of the central node for the Sholl analysis (SNT) and non-circular Sholl rings (Neurite-J) may influence the shape of the Sholl curves and make them less comparable. As the NLI algorithm counts intersections with thick fiber bundles multiple times, the curves are, especially close to the explant body, considerably higher and therefore not directly comparable.

Conclusion
The methodology presented here helps to accelerate high-throughput organotypic outgrowth assays. It offers a wide range of metrics to describe outgrowth morphology while limiting examiner interaction to an initial setup phase. As a consequence, the evaluation process is objective and enables inter-laboratory comparisons of results. For semiquantitative evaluations, Sholl based analyses like the NLI may be sufficient, as they produce a simple measure to describe the effect of an external factor on neurite outgrowth. If more dataand especially metric dataare needed on the morphology, ExplantAnalyzer demonstrates its power by producing data for comparative studies.

Declarations of interests
None.