A platform for stereological quantitative analysis of the brain-wide distribution of type-specific neurons

Quantifying the distribution of specific neurons throughout the whole brain is crucial for understanding physiological actions, pathological alterations and pharmacological treatments. However, the precise cell number and density of specific neurons in the entire brain remain unknown because of a lack of suitable research tools. Here, we propose a pipeline to automatically acquire and analyse the brain-wide distribution of type-specific neurons in a mouse brain. We employed a Brain-wide Positioning System to collect high-throughput anatomical information with the co-localized cytoarchitecture of the whole brain at subcellular resolution and utilized the NeuroGPS algorithm to locate and count cells in the whole brain. We evaluated the data continuity of the 3D dataset and the accuracy of stereological cell counting in 3D. To apply this pipeline, we acquired and quantified the brain-wide distributions and somatic morphology of somatostatin-expressing neurons in transgenic mouse brains. The results indicated that this whole-brain cell counting pipeline has the potential to become a routine tool for cell type neuroscience studies.


Results
Validation of the SOM-Cre mouse line. We first confirmed the accuracy of using the SOM-IRES-Cre:Ai3-EYFP transgenic mouse (Fig. 1a) to genetically target SOM neurons. We performed immunostaining using an antibody against SOM in sections from the primary motor area (MOp), the hippocampal formation (HPF) and the caudoputamen (CP) and collected images using a Zeiss 710 confocal microscope with a 20× objective (Fig. 1b). The overlapping images of fluorescently labelled somas and immunostained molecules showed that the majority of the enhanced yellow fluorescent protein (EYFP)-labelled neurons were SOM-positive (SOM+), consistent with previous studies 32, 39 . These results indicated that SOM-IRES-Cre driver lines reliably target SOM neurons and can be used to study the distribution of all SOM neurons in the whole brain.
Whole-brain optical imaging. We acquired the whole-brain datasets from three SOM-IRES-Cre:Ai3-EYFP mice, including EYFP-labelled SOM neurons and PI-stained anatomical reference, in coronal planes (Fig. 2). Figure 2a and b show typical SOM distribution and co-localized cytoarchitecture in the hippocampal coronal plane, respectively. The PI-stained landmarks enabled us to easily identify the distinct laminar distribution of SOM neurons with different cell densities in different layers in a typical six-layer structure in the cortex in Fig. 2c. The overlap of GFP-and PI-channels suggested that a voxel resolution of 0.32 × 0.32 × 2 μm enables orientation of the neuronal cell bodies based on counterstained nuclei at single cell resolution (inset in Fig. 2c). The high resolution and data continuity of BPS facilitated brain reconstruction along the sagittal direction by resampling original coronal images at 0.32 × 0.32 × 2 μm. SOM cells were primarily distributed in layers IV and V, with a small number of cells detected in layers II/III and VIa 40,41 . Figure 2d shows the largest sagittal plane of a SOM-IRES-Cre:Ai3-EYFP mouse brain including almost all major brain regions. This result also demonstrated the data integrity of the whole-brain dataset and the variable density of SOM neuron distribution in different brain regions. Enlarged views of white square boxes in Fig. 2d illustrated that BPS enabled imaging of the entire brain in high enough detail to distinguish individual cell bodies in brain regions with variable cell densities (insets in Fig. 2d). These results demonstrate the ability of BPS to visualize single neurons for cell localization and counting in entire brains.
Stereological cell counting. Accurately and automatically segmenting, locating and counting cell bodies is a prerequisite for quantifying the whole-brain cell distribution. To validate the accuracy of stereological cell counting, we compared NeuroGPS cell detection with manual identification (Fig. 3). Figure 3a depicts cell localization in a 100 × 100 × 100 μm data cube in the superior colliculus (SC) from the SOM-IRES-Cre:Ai3-EYFP mouse brain dataset. A total of 59 soma centres were automatically identified (indicated by red dots) using NeuroGPS, whereas 57 neurons were manually distinguished. Two background signals were mistakenly identified as neurons (indicated by yellow arrows in Fig. 3a). The recall and precision of automatic localization were calculated as 100% and 96.7%, respectively, for the overlap of automatically identified and manually detected cell bodies. We also generated a maximum intensity projection of this data cube to simulate traditional cell quantification in histological tissue sections (Fig. 3b). Due to no axial resolving power, some disjunctive but aligned cells in the axial direction of the data cube (indicated by red arrows in Fig. 3a) were partially overlapping in the projection image (indicated by red arrows in Fig. 3b), resulting in segmentation difficulties and potential counting errors. We compared the accuracy of stereological counting in the full volume and planar cell counting in the z-projection image of five randomly selected data cubes (Fig. 3c). The results showed that the cell numbers obtained by counting in 3D full volumes were higher than those obtained by counting z-projection images. We performed a Wilcoxon signed-rank test and acquired a two-sided P value was 0.043, demonstrating a significant difference between 3D counting and Z-projection counting. This result indicates potential erroneous omission resulting from the overlap of individual neurons along the z-axis, demonstrating that automated stereological cell counting is the only adequate method for unbiased cell quantification in a whole brain. Furthermore, we also estimated the accuracy of automated cell counting in 18 SOM-expressing brain regions from the three datasets (Fig. 3d). We randomly selected a representative volume of 512 × 512 × 512 μm in each region in each brain. The highest and lowest correct identification recall rates in these regions were 100% and 86.5%, respectively. The highest and lowest precision rates for correct identification in these regions were 100% and 85.2%, respectively. All three datasets exhibited identification rates higher than the standard of 85.0%. Average whole-brain recall and precision rates were 94.0 ± 0.4% and 91.8 ± 0.5% (n = 3), respectively. Furthermore, we randomly selected 20 neurons in each data cube and calculated their cell brightness. Pairwise comparisons between different datasets shows that there was no significant difference in the averaged cell brightness (p > 0.5 for all comparisons). It demonstrates slight brightness changes due to individual difference have no effect on cell counting. These results illustrated that stereological cell counting in whole-brain datasets using the NeuroGPS algorithm can accurately detect and segment soma as a potential method for quantifying brain-wide cell distribution. Figure 4 shows the typical distribution of SOM-expressing neurons in a mouse brain. SOM was ubiquitously expressed in the brain, as shown in the volume rendering in Fig. 4a, in brain regions consistent with previous studies 32, 35 . Figure 4b-g show a series of fluorescently labelled representative coronal plane images from the same dataset, including main SOM-expressing brain regions. The insets of the enlarged views of these regions demonstrated that different brain regions exhibit different densities of SOM-expressing neurons. SOM expression was densest in SC in the brain stem 32 . SOM was also relatively highly expressed in cerebral nuclei in the LS, CEA and BST regions, whereas SOM neurons were widely distributed with relatively low densities in the MOp, MOs, SSp, SSs, PIR, VISp, VISl and HPF regions of the cerebral cortex, the BLA and LA in the cortex subplate and the HY in the brain stem. There were relatively few SOM neurons in the CP in the brain stem and cc, although some densely arranged SOM neurons were observed in the CB. Some of this distribution variability within the SOM population in different regions might be functionally significant. These results provide a distribution reference for the functional study of SOM-expressing neurons.

Morphological reconstruction of SOM neurons. Soma segmentation not only provides the cen-
tral location of the cell body but also morphological characteristics of neuronal somas. We observed that SOM-expressing neurons in the same nuclei had similar morphologies, which might be used to classify subtypes of SOM-expressing neurons. We selected and reconstructed a typical 3D morphology of SOM-expressing neurons at the original resolution in 18 brain regions (Fig. 5a). Most cell bodies showed an ovoid or spindle profile with various cell volumes. SOM neuronal cell bodies in the CB were obviously bigger than all of the other somas. Furthermore, we randomly selected a data cube of 100 × 100 × 100 μm in each brain region, measured the longest and shortest radii, surface area and cell volume of all fluorescently labelled somas in the data cube, and calculated the corresponding average radius and longest to shortest radii ratio (Fig. 5b). The results demonstrated that both the longest and shortest radii and the surface area and cell volume of SOM-expressing neurons in the CB were higher than in other regions. However, the longest to shortest radii ratios for neurons in the CB were similar to those in other regions. This result indicated that SOM-expressing neurons had similar soma profiles in different regions, consistent with the reconstructions shown in Fig. 5a. Except for the CB, SOM-expressing neurons in other regions exhibited similar ranges of shortest radii and various longest radii, demonstrating similar characteristics for average radii and the longest to shortest radii ratio. The longest and shortest radii in BST and CEA were overall longer, thus the corresponding surface area and cell volume were bigger. The morphological features of SOM-expressing neurons in the neocortex were relatively consistent. We also performed a statistical analysis on the morphological traits of SOM neuronal somas (Fig. 5c). There was no significant difference in the soma morphological features for the pairwise comparisons between most regions. Some regions presented differently Quantifying the whole-brain distribution of SOM neurons. Quantitative mapping of a given type of neuron in the entire brain can improve our understanding of this neuron type in specific brain functions. To explore this diversity, we comprehensively quantified the distributions of SOM-expressing neurons in each region of three brains (Fig. 6). The total number of SOM-expressing neurons in a SOM-IRES-Cre:Ai3-EYFP mouse brain was 1,901,273 ± 67,096 (Fig. 6a). In terms of numbers, the most prominent labelling was observed in the HPF (130,431 ± 2,531 cells) due to its large volume, with denser labelling than reported in a previous study 34 . The numbers in the SSp were also higher than in other regions (111,365 ± 2,965 cells). SC, HY, LS and PIR showed 5~9 × 10 4 SOM-expressing neurons, whereas the CP, MOs, CB, MOp, VISp, CEA, SSs, BST and BLA have 1~5 × 10 4 SOM-expressing neurons. LA, VISl and cc exhibit less than 1 × 10 4 SOM-expressing neurons. There were only 2,996 ± 29 SOM-expressing neurons in the cc. Cell density showed distinct diversity in terms of cell number, reflecting a volumetric difference in various brain regions. The qualitative impression of the brain-wide distribution patterns was supported after calculating the cell density. Across the whole brain, the most abundant labelling was found in the SC, LS and CEA, consistent with Fig. 4. The cell densities in these regions were 48,130 ± 1,278, 41,140 ± 1,708 and 34,370 ± 498 cells/mm 3 , respectively. The regions in the neocortex had similar cell densities, consistent with previous studies 32 , and the cells were much denser than the 2D density reported in a previous study 39 . The average cell density in the neocortex was 10,735 ± 422 cells/mm 3 . The cell density of HY in the brain stem was similar to the neocortex. CP and CB showed low SOM expression densities of 4385 ± 233 and 3882 ± 91 cells/mm 3 , respectively, reflecting their large volumes. SOM was sparsely expressed in the cc (2431 ± 116 cells/mm 3 ). We also compared the cell distributions between different regions (Fig. 6b). Different from the statistical results of the soma morphology (Fig. 5c), both cell number and density presented a significant diversity among different regions. There was a significant difference in the cell number for the pairwise comparisons of most regions. For cell density, there was no significant difference for the pairwise comparisons of sub-regions of cortex, cortex and HY, CP, cc and CB. The other pairs presented differently. This distribution information could potentially be used as a reference for studying SOM neurons, and these results demonstrated that this platform could be used for the quantitative analysis of cell distribution in the whole brain.

Discussion
Understanding the principles of brain functions requires the systematic characterization of participating cell types. Establishing methods to efficiently identify and quantify specific neurons will facilitate our understanding of how neurons are involved in specific functions. Advances in whole-brain optical imaging and stereological cell localization provide a method to answer fundamental questions concerning cell distribution and cell number. BPS enabled the acquisition of fine structural information with the co-located anatomical reference of labelled neurons in the whole brain at subcellular resolution. NeuroGPS is capable of automatically detecting, locating and counting cell bodies in 3D. In this study, we combined these two technologies to constitute a novel platform to quantify a comprehensive distribution of type-specific neurons. To demonstrate the utility of this platform, we obtained high-resolution brain-wide datasets of SOM-IRES-Cre:Ai3-EYFP mouse brains with co-located cytoarchitecture and quantitatively analysed the distribution features of SOM expression and the morphological features of SOM neural somas. The results provide a platform with the capability of precise orientation, fine soma morphological reconstruction and accurate cell counting of a given type of neuron in the whole brain.
The automatic whole-brain optical imaging and stereoscopic cell counting and reconstruction demonstrated an unprecedented ability to accurately and quantitatively detect and analyse the distributional features of specific neurons in the whole brain. Acquiring a 3D brain-wide structural dataset is a prerequisite for accurately locating and counting the cells throughout the whole brain in the stereological approach. In conventional 2D approaches, information overlap between axial projections easily leads to cell omission and increases the difficulty of segmenting and counting. Interval sampling decreases the workload of manual data acquisition but fails to depict the real topological organization of the neurons. In contrast, visualizing neurons with full-volumetric whole-brain imaging overcomes these limitations and records realistic brain-wide cell organization in a reliable and efficient manner. Whether or not to precisely segment axial-aligned or adjacent somas in 3D is the key factor influencing automatic locating and counting accuracy, and this step requires micron-or even submicron-scale imaging resolution in the whole brain, although the sizes of neuronal somas are at the scale of tens of microns. The strategies of combing optical imaging and mechanical sectioning for whole-brain imaging [25][26][27]42 have the potential to meet this requirement. The results of this study demonstrated this technical advantage over conventional 2D methods. Based on the high-resolution whole-brain 3D dataset, stereoscopic cell counting by NeuroGPS avoided the missed counting of axially overlapped cells and possible repeat counting of cells in adjacent slices.
Furthermore, this platform takes advantage of accurate neural morphological reconstruction with high resolution. Characterizing a cell type requires cross-validating the molecular, anatomical and physiological proerties 43 . Previous studies demonstrated the predominant cell types of SOM neurons were Martinotti cells (MCs) in the neocortex and oriens-lacunosum moleculare (O-LM) neurons in the hippocampus 34,35,40,44 . Most morphological works focused on soma location, dendritic morphology, axonal tree pattern and neuronal connectivity. Wang, Y. et al. 40 performed a systematic multidimensional study of the detailed anatomical, physiological and molecular properties of MCs in the cortex. They found most MC somas in the cortex were ovoid or spindle shaped, consistent with our results. Figure 5 shows SOM neurons in most regions except CB had a similarity of the soma morphology. The significant difference of SOM neuronal somas between CB and other regions indicates that further molecular and physiological study might be considered to distinguish potential different subtypes. In addition, the digital reconstruction of individual soma is suitable for the quantitative analysis and computational modelling of neuronal cell body shape changes during developmental stages and pathological conditions.
Generating a comprehensive database of the distribution and morphology of type-specific neurons elucidates the role of neurons in neural circuits and informs the design of experiments to study their function. The regional diversity of SOM expression in this study provides an important anatomical reference for SOM-related specific functions. This platform enabled the precise regional analysis of a given type of neuron, reflecting the advantage of simultaneously acquired cytoarchitecture in the same brain with BPS. Regional distribution of soma locations indicates further detailed morphological analysis of dendritic and axonal arbours for future cell type study 40 . Using this platform, researchers can study morphological and distributional alterations of neuronal somas during developmental and evolutionary processes. This platform also facilitates the investigation of the neuronal loss and death in neurodegenerative and psychiatric disease progression 9 . We propose that this platform could potentially become a routine tool in neuroscience research.

Methods
The pipeline for quantifying the cell distribution of type-specific neurons in the whole mouse brain consisted of five steps. (1) Labelling the type-specific neurons in the whole brain using genetic or viral approaches. (2) Imaging the whole brain at subcellular resolution by BPS. Resin embedding. All of the histological procedures have been previously described 45 . Briefly, SOM-IRES-Cre:Ai3-EYFP mouse were anaesthetized with a 1% solution of sodium pentobarbital and subsequently intracardially perfused with 0.01 M PBS (Sigma-Aldrich Inc., St. Louis, MO, USA), followed by 4% paraformaldehyde (Sigma-Aldrich Inc., St. Louis, MO, USA) and 2.5% sucrose in 0.01 M PBS. The brains were excised and post-fixed in 4% paraformaldehyde at 4 °C for 24 h. After fixation, each intact brain was rinsed overnight at 4 °C in a 0.01 M PBS solution and was subsequently dehydrated in a graded ethanol series (50%, 70%, and 95% ethanol at 4 °C for 1 h each). We modified the previous resin-embedding approach to reduce background fluorescence. Briefly, after dehydration, the brains were immersed in a graded series of GMA (Ted Pella Inc., Redding, CA, USA) including 0.2% Sudan Black B (SBB): 70%, 85%, and 100% GMA for 2 h each and 100% GMA overnight at 4 °C. Subsequently, the samples were impregnated in a prepolymerized solution of GMA for 3 days at 4 °C and finally embedded in a vacuum oven at 48 °C for 24 h. The 100% GMA solution contains 67 g A solution, 2.8 g deionized water, 29.4 g B solution, 0.2 g SBB, and 0.6 g AIBN as initiator. The 70% and 85% GMA (wt/wt) were prepared from 95% ethanol and 100% GMA.
Immunostaining. Post-fixed brains were sectioned at 50 µm thickness using a Leica vibratome (Leica VT1200S, Germany). Sections were permeabilized using 0. 3% Triton-X 100 in PBS and blocked with 5% bovine serum albumin (BSA, Sigma) and 0.3% Triton in PBS and subsequently incubated with anti-SOM (goat polyclonal; 1:300; Millipore) at 4 °C overnight. The sections were washed three times for 10 min in 0.1 M PBS at room temperature and incubated with secondary antibodies (Alexa Flour 594, Rb-Anti-Gt-g(ab')2,H, Invitrogen) for 2 h at 37 °C. The sections were washed three times for 10 min in 0.1 M PBS at room temperature. The sections were mounted and coverslipped with Fluorogel. The representative images were acquired with a Zeiss 710 LSM confocal microscope (Zeiss, Jena, Germany).
Whole-brain imaging. We performed the data acquisition with BPS 27 . Briefly, we fixed the sample on a high-precision 3D translation stage. During data acquisition, the sample was immersed in a 2 μg/ml propidium iodide (PI) solution to counterstain the cytoarchitecture of the sample in real time during data acquisition. We used two-channel structured illumination microscopy (SIM) 46 to simultaneously image EYFP-labelled SOM neurons and PI-counterstained cytoarchitecture at a voxel resolution of 0.32 × 0.32 × 2 μm. Optical sectioning of SIM enabled the acquisition of the image slightly below the top surface to avoid the effect of sectioning marks. Mosaic scanning was performed to acquire the entire coronal image of the sample surface. There was an overlap of 12 pixels between adjacent mosaics for lateral registration. Subsequently, the sample on the stage was moved to a diamond knife to remove imaged tissue. This imaging-sectioning-staining cycle was repeated until the entire sample was imaged. The total data acquisition time for each brain was approximately 130 h. The whole-brain datasets had a size of approximately 4 terabytes, including approximately 5400 coronal images, for each channel. Segmenting brain region. We identified anatomical brain regions based on the Allen mouse Reference Atlas 47 using PI-counterstained signals. We downsampled the original data set at the voxel resolution of 3.2 × 3.2 × 3.2 μm and rotated and translated the data for 3D spatial correct to easily figure out the corresponding coronal planes in the reference atlas. Subsequently, we resampled the transformed data of PI-stained images into 3.2 × 3.2 × 20 μm and depicted the contours of each region using Amira software (v 5.2.2, FEI, Merignac Cedex, France). The segmentation editor module of Amira was utilized for the manual outline segmentation of the whole mouse brain. Referring to the cytoarchitecture of each region in the Allen mouse Reference Atlas 47 , we manually identified the contour of each region in the PI-stained cytoarchitecture. The outlined contour was saved as a TIFF file.
Localization and counting. We used the NeuroGPS algorithm 28 to automatically segment, locate and count somas in 3D space. With the NeuroGPS algorithm, we employed binarization and erosion operations to extract the foreground and subsequently used an L1 minimization model 48 to locate the somas in the extracted regions. Limited by the memory capability of the graphic workstation (T7600 with two Intel E5-2687w CPUs, 256 GB memory and an Nvidia K6000 graphics card, Dell Inc., Round Rock, Texas, USA), each dataset of EYFP-labelled SOM neurons was divided into 50,000 blocks of 512 × 512 × 512 μm with a resolution of 1 × 1 × 2 μm. The NeuroGPS analysis was performed in the graphic workstation cube by cube to identify all of the EYFP-labelled neurons. The cell locations and numbers of SOM neurons in the whole brain were collected and saved as an SWC file for each brain.
To evaluate the counting omission and commission errors made by NeuroGPS, we compared the results of automatic and manual recognition. Ground truth was back-to-back validated by three skilled researchers. B1 and B2 indicate the total neuronal numbers by manual and automatic recognition, respectively, whereas B indicates the cell number using both approaches. The recall R is defined as R = B/B1, whereas the precision P is defined as P = B/B2. We randomly selected 5 data cubes of 512 × 512 × 512 μm from a whole-brain dataset and generated 10 images of 50 μm-thick z-projection for each volume. Subsequently, we performed stereological and planar cell counting in full volumes and z-projection images, respectively, and calculated the R and P for each data cube.

Reconstruction and visualization.
We reconstructed and visualized the cell body morphology of SOM neurons in the midsagittal plane of the brain using Amira software. The process included extracting the data in a range of interest, sampling or interpolation, reslicing the images, identifying the maximum intensity projection, and volume and surface rendering using the main Amira modules. Statistical analysis. We compared the morphological features of SOM neuron somas. We measured the longest and shortest radii, surface area and cell volume of the SOM neuron somas with a resolution of 0.32 × 0.32 × 2 μm by NeuroGPS. We also analysed the SOM distribution of each region by comparing the cell number and density of each region. The corresponding SWC files were recalled according to the 3D contour of each region in the PI-stained signal and merged together as the counting result of each region. The volume for each region was measured according to the 3D contour of each region for the PI-stained signal in the Amira software program. Subsequently, we divided the cell number of the brain region by the corresponding volume to get the cell distribution density for the region. This process was performed in all three whole-brain datasets. All of the results are presented as the MEANs ± SEM. Moreover, we assessed the statistical significances of soma morphology and cell distribution between different regions using analysis of variance (ANOVA). We also statistically compared stereological and planar cell counting using Wilcoxon signed-rank test. We performed these statistical calculations using SPSS Statistics (v 22, IBM, New York, USA). P values <0.05 were considered significant.