Neuronal composition of processing modules in human V1: laminar density for neuronal and non-neuronal populations and a comparison with macaque

Abstract

The neuronal composition of homologous brain regions in different primates is important for understanding their processing capacities. Primary visual cortex (V1) has been widely studied in different members of the catarrhines. Neuronal density is considered to be central in defining the structure–function relationship. In human, there are large variations in the reported neuronal density from prior studies. We found the neuronal density in human V1 was 79,000 neurons/mm³, which is 35% of the neuronal density previously determined in macaque V1. Laminar density was proportionally similar between human and macaque. In V1, the ocular dominance column (ODC) contains the circuits for the emergence of orientation preference and spatial processing of a point image in many mammalian species. Analysis of the total neurons in an ODC and of the full number of neurons in macular vision (the central 15°) indicates that humans have 1.2× more neurons than macaques even though the density of neurons in macaque is 3× the density in human V1. We propose that the number of neurons in a functional processing unit rather than the number of neurons under a mm² of cortex is more appropriate for cortical comparisons across species.

Introduction

It has been long established that the connectivity both within and between layers of cortex is an important feature enabling the understanding of the structure and function of cortical circuitry. Recent studies have emphasized the importance of determining the density of different neuron populations within cortical areas (Dombrowski et al. 2001; Charvet et al. 2015; von Bartheld et al. 2016) to understand uniformity principles within cortical areas and across cortical layers (Meyer et al. 2010; Chariker et al. 2016, 2018, 2021, 2022; Wagstyl et al. 2018) as a crucial step in determining the details necessary to make population models of cortical circuits.

Visual cortex is among the most extensively studied cortical areas across many mammalian species, including humans. The early visual pathway of different members of the catarrhines, the parvorder that includes macaque monkeys (family Cercopithecidae) and the Great Apes (Hominidae), shows considerable similarity (de Sousa et al. 2010; Kaas 2020). The photoreceptors, postreceptoral neurons of the retina and those in main relay nucleus to cortex, the lateral geniculate nucleus (LGN), in the different subfamilies are closely matched (Boycott and Dowling 1969; Kolb and DeKorver 1991; Sumner and Mollon 2000; de Sousa et al. 2013). The identification of thalamic terminal distributions using immunocytochemical labeling with the vesicular glutamate transporter (vGluT2) (Fremeau Jr et al. 2001; Fujiyama et al. 2001) has made study of the afferents into human cortex and a comparison with macaque possible (Garcia-Marin et al. 2013). There are many similarities in vGluT2-ir distributions between macaque and human primary visual cortex, such as high densities of vGluT2-ir terminals in layer 4C, patches of VGluT2-ir puncta in the supragranular layers (2/3), lower densities but clear distributions in layers 1 and 6, and very few puncta in layers 5 and 4B (Garcia-Marin et al. 2013). Nonetheless, there are also some major differences. For example, the structure (Preuss and Coleman 2002) and presumptive thalamic input (Garcia-Marin et al. 2013) to layer 4A are quite distinct between macaque and human.

Studies of human V1 have reported substantially different neuronal densities, ranging from 18 × 10³ (Pakkenberg 1966) to 123 × 10³ neurons/mm³ (Selemon et al. 1995; see Supplementary Table 1 for additional studies). In general, studies that used nonstereological techniques to estimate neuronal density yielded overall densities that were lower than estimates of neuronal density using stereological methods (Everall et al. 1993; Pakkenberg and Gundersen 1997; Dorph-Petersen et al. 2007; Inda et al. 2007; van Kann et al. 2017) or 3D counting methods (Selemon et al. 1995). However, even among the studies that used stereological or 3D counting methods, the average densities varied considerably between studies, up to a factor of 3 (c.f. Selemon et al. 1995 and Inda et al. 2007, in Supplementary Table 1). This variability in the results of the overall neuronal density of human V1 is also observed in the results from studies that focused on the laminar densities (see Supplementary Table 1). When laminar densities were determined for a single layer, large differences were observed in the estimates across studies; for example, in layer 4C, estimates range from 47 to 195 × 10³ neurons/mm³ (a factor of 4: Inda et al. 2007; Van Kann et al. 2017; see Supplementary Table 1 for other studies). It is worth mentioning that in addition to the variability between studies, several variables in human neuroanatomical studies, such as cell density and brain volume, vary substantially between individuals, exceeding the within-individuals variation in repeated measures (Leuba and Garey 1989). In prior stereological studies, interindividual coefficients of variation for human neuronal density range from 0.14 to 0.23 (Everall et al. 1993; Pakkenberg and Gundersen 1997; Dorph-Petersen et al. 2007).

Although there have been comparisons between the neuronal density in humans and macaques, advances in neuron identification and large-scale imaging combined with automated image analysis have made it possible to make more rigorous quantitative comparisons. Recent studies using these methodological approaches (Kelly and Hawken 2017; Garcia-Marin et al. 2019; Kelly et al. 2019) have demonstrated that in earlier studies, there was a substantial underestimation of neuronal densities in different layers and sublayers of macaque V1 (O'Kusky and Colonnier 1982; Beaulieu et al. 1992), For example, in the macaque, using the same techniques adopted in the current study, we reported that the overall neuronal density (Kelly and Hawken 2017; Garcia-Marin et al. 2019; Kelly et al. 2019) was about double the density reported in earlier studies (O'Kusky and Colonnier 1982; Beaulieu et al. 1992) but comparable to the density from a more recent study with extensive sampling using NeuN and stereological methods (Giannaris and Rosene 2012).

The aim of the current study was to reevaluate the neuronal quantification in human V1 using large-scale imaging combined with automated image analysis and to make a systematic comparison between the human and macaque focusing on the laminar organization in terms of neuronal composition. While there have been numerous studies that have estimated the variation in neuronal density across layers in macaque, there are only a few studies in human (Leuba and Garey 1989; Inda et al. 2007), and neither study measured continuously through the thickness of the cortex, from layer 1 (L1) to the white matter (WM). We found total cortical densities in human V1 that were 10–15% higher than those reported using stereological methods in Nissl-stained tissue (Everall et al. 1993; Pakkenberg and Gundersen 1997; Dorph-Petersen et al. 2007). The human neuronal densities are about one-third those we found in macaque V1. In addition, we determined the variation in neuronal density continuously through the thickness of cortex and found that there were substantial changes in neuronal density both between layers and within layers, like what we reported in macaque (Kelly and Hawken 2017; Garcia-Marin et al. 2019).

Several studies have assigned a critical role to inhibitory neurons in controlling the flow of activity within local circuits, including shaping receptive fields (Sillito 1975; review in Ferster and Miller 2000), gain control (Atallah et al. 2012; Wilson et al. 2012), and surround suppression (Bair et al. 2003; Angelucci and Bressloff 2006; Adesnik et al. 2012). For a comprehensive understanding of the interaction between excitation and inhibition in the neocortex, it is crucial to obtain data on the distribution of excitatory and inhibitory neurons in a cortical column. Recently, we have reported that GABAergic neurons in macaque V1 accounted for a smaller fraction (11%) of the total neuronal population across layers 1–6 than has previously been reported (Kelly et al. 2019). The parvalbumin (PV) group that includes the chandelier cells and the fast-spiking basket cells represents the largest population of interneurons in the neocortex (40%) (review in Tremblay et al. 2016). Across all layers of macaque V1, we have estimated that the proportion of GABA-ir neurons that were PV-ir was 52%.

Reconciling the differences in neuronal density with known allometric scaling in the catarrhines, primates in general, and across all mammals has major implications for understanding the relationship of structure to function (Finlay et al. 2001; Hill et al. 2010; Herculano-Houzel 2014; Herculano-Houzel, Catania, et al. 2015; Herculano-Houzel, Messeder, et al. 2015). One means of relating the differences in density to other features of cortex is to define a module of cortex. Having an accurate estimation of the neuronal density allows the determination of the size of the populations underlying a processing module in cortex. One module that has been widely used in the comparative process was introduced by Rockel et al. (1980). The proposed unit was the number of neurons under 1 mm² of cortex through the full depth of cortex—from layer 1 to white matter. This module has been used in comparison of different mammalian cortices (Carlo and Stevens 2013; Srinivasan et al. 2015) and within the catarrhines (Colonnier and O'Kusky 1981; O'Kusky and Colonnier 1982) and across a wider range of primate species (Atapour et al. 2019).

However, as clearly pointed out by Rakic (2008), there are numerous types of columns or modules. A module that has been of interest since its discovery and elaboration is the ocular dominance column (ODC; Hubel and Wiesel 1962, 1968; Hubel et al. 1978; Horton and Hedley-Whyte 1984; Adams et al. 2007) that is thought to contain the circuits for the emergence of orientation preference and spatial processing of a point image (Carlo and Stevens 2013) in many mammalian species. In the current study, we determined the population of neurons within each layer underlying an eye dominance module to compare with the size of the populations underlying the same module in macaque (Garcia-Marin et al. 2019) thereby evaluating the uniformity (Rockel et al. 1980; Carlo and Stevens 2013) or nonuniformity (Herculano-Houzel et al. 2008; Lent et al. 2012) of cortex hypothesis at the level of the eye dominance module.

Although neuronal density and the size of ODCs and minicolumns are different in human and macaque, our current results show that there are a similar number of neurons in one minicolumn in both human and macaques. However, there are 1.7× more neurons per ODC in human V1 compared with macaque V1 and 1.2 more neurons in the visual field representation of the central retina. Our results suggest that for the same eccentricity, more neurons are devoted to the first visual processing steps in humans than in macaques, which, in turn, indicates the potential for a refined level of perceptual processing.

Material and methods

Human brain tissue

Brain sections from seven adult humans were obtained from NIH NeuroBioBank. Specific information on each individual is presented in Supplementary Table 2. Briefly, all sections were from males without any reported neurological disease (mean age, 51.7 ± 5.2 y; range, 23–66 y). After death, the occipital left hemisphere was blocked (1–1.2 cm thick), fixed in 4% PFA, and cut in 40-μm-thick coronal cryostat sections. The sections selected for this study were located around 8–10 mm from the tip of the occipital pole. We estimated that our samples represent regions of primary visual cortex (V1) that had a visual field representation between 2.5° and 5° eccentric, and the images were obtained at the floor of the calcarine sulcus—near to the representation of the horizontal meridian.

Immunofluorescence

To estimate the total neuronal density and the density of the PV inhibitory neurons, sections were immunofluorescence stained for the pan-neuronal marker NeuN and for the calcium-binding protein PV (Fig. 1A and B). Sections were incubated for 1 h in a blocking solution of 0.01 M PBS with 0.25% Triton-X and 3% normal goat serum (NGS) and then incubated overnight at 4°C with rabbit anti-NeuN antibody (1:2,000, MAB5504, Millipore, Temecula, CA, United States) in 0.01 M PBS with 0.25% Triton-X and 3% NGS. Sections were rinsed in 0.01 M PBS and incubated overnight at 4°C with Alexa biotinylated goat anti-rabbit 488 (1:2,000, Vector Laboratories, Burlingame, CA). After rinses, sections were incubated overnight at 4°C in guinea pig anti-PV polyclonal serum (1:1,000; 195,004, Synaptic Systems, Göttingen, Germany), and later in Alexa biotinylated goat anti-guinea pig 594 (1:500; BA-7000, Vector, Burlingame, CA, United States). After rinsing in PBS, the sections were counterstained with 10-μg/mL DAPI (D9542-5MG, Sigma-Aldrich, St. Louis, MO, United States; Fig. 1C) and were treated with Autofluorescence Eliminator Reagent (Millipore) to minimize lipofuscin-like autofluorescence. Finally, the sections were washed and mounted with ProLong Gold Antifade Reagent (Invitrogen Corporation, Carlsbad, CA, United States).

Fig. 1.

Laminar distribution of neurons through the thickness of human V1. Section was triple labeled for NeuN, PV, and DAPI. A) Neuronal population immunoreacted with the pan neuronal marker NeuN. B) Neurons in V1 immunoreactive for PV. C) The total cellular population including neurons, glial, and epithelial cells labeled with DAPI. The laminar boundaries are drawn according to the scheme of Brodmann (Brodmann 1909; Braak 1982) with Hassler’s (1967) labels indicated in parentheses. Scale bar in C 250 μm.

Calculation of the neuronal density

For neuronal quantification, fluorescence sections were imaged using a Leica TCS SP8 confocal system (Leica Microsystems, Wetzlar, Germany). A 488-nm Argon laser, a 564-nm DPSS laser, and a 405-nm diode laser were used to excite the Alexa Fluor 488, and 594 and DAPI, respectively. Gain and offset levels were set for each channel such that there were few saturated pixels and minimal background noise. Image stacks were acquired by specifying an upper and lower z position, which correspond to the top and the bottom of the section, as we were able to get complete penetration of the antibodies through the 40-μm section.

For quantification of neuronal density, we followed the methodology previously described in our recent studies (Kelly and Hawken 2017; Garcia-Marin et al. 2019; Kelly et al. 2019). Briefly, a column of tissue was defined spanning from the pial surface to the white matter and oriented orthogonal to the pial surface (Fig. 2A). Two columns were acquired for each section from each subject. The Leica Application Suite X software automatically adjusted the number of stacks needed to cover the selected region; in our case, this number ranged from 9 to 14 stacks, with a constant horizontal overlap of 110 pixels between stacks. Each stack was acquired using a 40× oil-immersion objective lens (NA 1.4, refraction index 1.45), zoom 1, with a pinhole size of 1 airy unit, a z-step of 0.5 μm, and a range of 54–103 (mean 74) optical planes. The z stack acquisition routinely included planes above and below the section surfaces. The image resolution in x–y was 1,024 × 1,024 pixels (290.6 × 290.6 μm), and the scanning speed was 400 Hz.

Fig. 2.

Methods for quantifying cellular (DAPI), neuronal (NeuN), and PV densities through the different cortical layers. A) A column of tissue spanning from the pial surface to the white matter and oriented orthogonal to the pial surface. A vertical series of z-stacks, spanning the thickness of cortex (each stack: 1024 × 1024 pixels, 290.6 × 290.6 μm, 40× oil-immersion objective lens, zoom 1, z-step of 0.5 μm), was acquired. The number of z-stacks required to span the thickness varied between 9 and 14, depending on the individual section, for example in (A), 10 z-stacks were acquired; z-stack#1 was acquired at the pial surface and z-stack#10 at the layer 6/WM boundary. Within each z-stack, a range of 54–103 optical planes were acquired to span the 40-μm-thick section. For the neuronal quantification, within each z-stack, we defined inclusion (right, down, and top), and exclusion borders (left, up, bottom), of the counting frame, (e.g. z-stack #5).z-stack 5. B) Maximum projection of 70 images from the DAPI channel. Three groups of cells are shown at higher magnification in the squares (D1, E1, F1). C) Image of a maximum projection of 14 images from the DAPI channel. (D1–D2), (E1–E2), and (F1–F2) inserts from B and C, respectively, showing the cluster of neurons in each square (D1, E1, and F1) and how they are individually resolved in the segmentation process (D2, E2, and F2). Each purple dot shows the centroid of the cluster and each green dot shows the centroid of the individual cells in each cluster extracted by the automated segmentation algorithm. G) Image of a maximum projection of 14 images from the NeuN channel, in which DAPI centroids are labeled with a green dot and the DAPI centroids that colocalize with a neuron are identified with a red dot. Note that the green (DAPI) centroids that do not colocalize with a NeuN-ir neuron belong to a nonneuronal cell. The NeuN-ir neurons that do not colocalize in this subset of z-planes have their centroid in other image planes. H) Image of a maximum projection of 14 images from the PV channel, in which DAPI centroids are labeled with a green dot and the DAPI centroids that colocalize with NeuN and PV are identified with a blue dot. The red dots indicate the NeuN labeled neurons shown in G. Note, some PV-ir neurons that have neither DAPI or NeuN centroids have their centroid in other image planes. Scale bar: 290 μm (A), 85 μm (B, C, G, H), 20 μm (D1–F1 and D2–F2).

Next, DAPI, NeuN, and PV images were automatically analyzed, and each cell was categorized according to its pattern of immunoreactivity, using the method previously described (Kelly and Hawken 2017; Kelly et al. 2019) (Fig. 2B–H). Briefly, cell centroids are identified in 2D for each optical plane in the DAPI channel (Fig. 2B and C). In a single image plane, the DAPI-labeled cells could be in close proximity to each other, so the 2D segmentation provided one single centroid for a cluster of cells (Fig. 2D1–F2; magenta dots). The next step identified the 3D centroids of the cells and split the clump into individual cells (Fig. 2D1–F2; green dots). Once the 3D centroids were identified for each cell, the other channels were evaluated for marker expression at the locations of the centroids (Fig. 2G–H). Cells were quantified using stereological exclusion boundaries (Sterio 1984; Gundersen et al. 1988) where each stack was treated as a 3D counting brick probe (Williams and Rakic 1988; Howard and Reed 2005). Objects touching the top, left, or back planes were excluded from quantification, whereas objects touching the bottom, right, or front planes were included.

Finally, the neuronal counts were converted to densities by dividing the number of counted neurons by the sampled volume in each stack, corrected to account for tissue shrinkage. The sample volume in each stack was determined as the region in which the penetration of the antibodies was uniform. Although the antibodies penetrated through the whole depth of the sections, optical planes acquired near the cut tissue surfaces yielded lower counts than optical planes acquired in the middle of the section. For each z stack in the column, we determined the z-range where the counts plateaued, then found the minimum plateau-range across all the z stacks and applied this range to all the stacks in the same column. On average, we sampled 20 ± 3 optical planes (Table 1). To determine the laminar density, laminar boundaries in V1 were identified in each column by visually identifying transitions in cell density and neuronal composition (Fig. 1) consistent with previous descriptions of these features in V1 (Brodmann 1909; Braak 1982). Although Brodmann’s (1909) laminar scheme for V1 is widely used, the distinction of layer 4 into three sublayers (4A, 4B, 4C) does not reflect that only layer 4C is equivalent to layer 4 in other areas (Hassler 1967). Layer 4C merges with layer 4 of V2, while Brodmann’s layers 4A and 4B are considered sublayers of layer 3. Although many studies in V1 in primates now use Hassler’s terms (see Balaram et al. 2014; Balaram and Kaas 2014), there are still many influential studies, particularly those that work on the circuitry and function (see Callaway 1998; Sincich et al. 2003; Adams et al. 2007; Federer et al. 2013) that use Brodmann’s nomenclature. To compare our work with other V1 studies in human and in nonhuman primates that use different nomenclature, both sets of terms are used to label the layers in the figures and tables. Brodmann’s nomenclature has been used through the text.

Table 1.

Total number of cells (#DAPI) or neurons (#NeuN, #PV) for each of the 7 humans (H1–H7) in each sampling column (b1–b2). Volume of tissue analyzed (mm³), cortex length (mm), and number of optical planes taken for each human and column. Densities for DAPI, NeuN, and PV labeled cells were calculated using the total number of cells (#DAPI, #Neun, #PV) and the volume (mm³) in each column. (A column of tissue was defined spanning from the pial surface to the white matter and oriented orthogonal to the pial surface).

Human	# DAPI	Dens_Dapi cell/mm3	# NeuN	Dens_NeuN Neurons/mm3	#PV	Dens_PV PV/mm3	%PV	# Planes	Vol mm3	cx length mm
H1b1	1,049	145,855	480	66,740	25	3,476	5.2	12	0.007	2.77
H1b2	1,915	142,812	814	60,704	38	2,834	4.7	22	0.013	2.82
H2b1	2,388	156,827	1,131	74,276	62	4,072	5.5	25	0.015	2.82
H2b2	3,476	186,882	1,825	96,053	93	5,000	5.1	29	0.019	3.00
H3b1	2,903	181,125	1,266	78,989	84	5,241	6.6	23	0.016	3.23
H3b2	2,526	146,837	1,216	70,686	68	3,953	5.6	27	0.017	2.95
H4b1	1,749	166,949	859	81,995	29	2,768	3.4	18	0.010	2.69
H4b2	1,035	168,858	498	81,248	34	5,547	6.8	12	0.006	2.36
H5b1	853	152,727	419	75,021	27	4,834	6.4	10	0.006	2.58
H5b2	1,741	167,302	796	76,492	58	5,574	7.3	19	0.010	2.53
H6b2	2,269	156,581	1,204	83,087	103	7,108	8.6	30	0.014	2.23
H6b3	1,071	197,919	534	98,682	29	5,359	5.4	9	0.005	2.78
H7b2	2,450	173,348	1,214	85,896	58	4,104	4.8	22	0.014	2.97
Mean	1,956	164,925	943	79,221	54	4,605	5.8	20	0.012	2.75
SEM	224	5,833	114	3,025	8	396	0.411	2.97	0.002	0.08

Using the same z-stacks, we also calculated the volume occupied by the soma and proximal dendrites of the neurons in the images labeled with NeuN. We applied a Gaussian filter of 2.0 to each stack and used the 3D object counter plugin in FIJI (Schindelin et al. 2012) that automatically applied a threshold to the images to obtain the volume occupied for the neurons.

Calculation of shrinkage

The degree of tissue shrinkage was measured by comparing sections in x, y, and z dimensions before and after tissue preparation with the aid of Olympus VS120 and confocal Leica SP8. The average linear shrinkage (S) was calculated by using the following formula: S = (A − 0)/A, where A is the absolute value before processing and 0 is the observed valued after processing. This yielded an average linear shrinkage of 18%, 18%, and 20%, in x, y, and z, respectively.

Image processing for figures

Images presented in figures were captured using a Leica TCS SP8 confocal system or an Olympus VS120-FL virtual slide scanning system. ImageJ was used to generate maximum intensity z projections. Adobe Photoshop CS software (Adobe Systems, San Jose, CA) was used to adjust the images for brightness and contrast and to generate the figure plates. Images were not otherwise altered in any way, e.g. by removing or adding image details.

Statistical analysis

Statistical comparisons were performed by using either a parametric (one-way ANOVA) if the datasets were normally distributed and passed the test for homogeneity of variances (Bartlett test) or nonparametric test (Kruskal–Wallis) if the data did not pass the test for the homogeneity of variances, followed by suitable post hoc tests. The data are presented as the mean ± standard error mean (SEM) or standard deviation where appropriate. All statistical comparisons were performed using MATLAB (MathWorks, Natick, MA).

Results

Initially, we show the neuronal laminar density distribution in human V1 and the laminar density distribution of the largest population of interneurons, the PV neurons. Next, we estimate the number of neurons in each layer through the depth of an ocular dominance processing module in human V1 and make a comparison to the same processing unit in macaque monkey visual cortex. The results indicate that although the density of neurons (per mm³) is considerably lower in humans than macaque, the number, when normalized to the same cortical processing module, is higher in humans than macaque.

Neuronal and glial density using 3D-confocal imaging

Many previous studies have estimated the neuronal density in human V1 using nonstereological techniques in 2D and reported a wide range of average density values. We recently demonstrated (Kelly and Hawken 2017; Garcia-Marin et al. 2019) that these techniques produced a substantial underestimation of the total neuronal density. Average neuronal density values estimated using stereology are higher than those reported from nonstereological studies but also show substantial variation between studies (40.5–102 × 10³ neurons/mm³).

We estimated the total neuronal density using NeuN and the total density of the largest subpopulation of interneurons, the PV neurons. Overall, we found a neuronal density of 79.2 ± 3.0 × 10³ neurons/mm³ (mean ± 1 SEM; Table 1, n = 13 from 7 individuals). We found a DAPI density of 164.9 ± 5.8 × 10³ cells/mm³ and a total PV density of 4.6 ± 0.4 × 10³ PV-ir neurons/mm³, which represents 5.8% of the total number of neurons. The mean vertical cortical thickness (pia to white matter) was 2.8 ± 0.1 mm.

A total of 12,256 neurons were counted in a total volume of 0.146 mm³, with an average volume of 0.012 mm³ per individual. For six individuals (H2–H6), the mean neuronal density across all layers showed a relatively narrow range (74,840–90,880 neurons/mm³). However, for one individual (H1), the mean neuronal density was low compared with the other cases (63,720 neurons/mm³) although this difference is not significant. The mean pairwise difference between measurements within an individual was about 10% of the mean, while the average between-subject difference was 16% of the mean. These two distributions were not significantly different (Student’s t test, P = 0.25). The direction of this result is in agreement with previous observations that variation in neuronal density between individuals of similar ages exceeds the variation in repeated measures within an individual (Leuba and Garey 1989). Furthermore, there was no correlation between the neuronal density and age (r² = 0.04126, ns) (Supplementary Fig. 2).

Glia to neuron ratio

DAPI labels not only neurons but also all other cell types, which in the cortex primarily, consist of glial cells and endothelial cells. The density of nonneuronal cells is the total density of DAPI-labeled cells (165 × 10³ cells/mm³; Table 1) minus the neuronal density (79 × 10³ neurons/mm³; Table 1), resulting in a density of nonneuronal cells of 86 × 10³ cells/mm³. Using a prior estimate that endothelial cells make up about 30% of the total nonneuronal cells (von Bartheld et al. 2016), we estimated that about 70% of the nonneuronal population were glial cells (60 × 10³ cells/mm³). Therefore, we determined that the average glia to neuron ratio (60/79) was 0.76 and the nonneuronal cell to neuron ratio (86/79) was 1.1.

Laminar distribution of neuronal density

The density was determined continuously, from sequentially acquired images, from the pial surface to the white matter (a column through the thickness of cortex), and this allowed us to analyze the neuronal distribution within layers (Fig. 3A) and across the normalized cortical thickness (Fig. 3B). Neuronal density varies at a finer scale than the divisions between the major cortical layers (Table 2). Because of the different relative position of the laminar boundaries in each individual section and even between columns within a section and the fine scale density variation, we adopted an alignment strategy where we found the peaks and the valleys in each individual column. The peaks correspond to the position of the highest density in layer 2, layer 3B/4A, layer 4Cβ, and layer 6; the valleys correspond to the positions of layer 4B; and the border 4Cβ/5. Using this strategy, we aligned the different columns and show the result in Fig. 3B and Table 3. Below, we present the results based on conventional layer boundaries and then compare these results to those determined by aligning peaks and valleys between columns.

Fig. 3.

A) Laminar density of neurons in human V1. Average density for 13 columns from 7 individuals for each layer shown by the horizontal bars. Individual layer density values for each column are represented by “o.” The values from each individual (H1–H7) are shown by the colors in the legend. There are two columns from each individual except H7 where there is one. B) Average continuous density through the relative cortical depth. The cortical depth was normalized for each section. The densities were individually smoothed for each section. Next, the density distributions were aligned to a peak—Layer 2 (pink), border 3/4A (orange), layer 4Cβ (purple) and layer 5 (blue)—or valley—Layer 4B (green) and 4Cβ/5 (brown)—and averaged. The values from each section sampled at the aligned peak or valley are shown by the open circles. C) Laminar density of PV-ir neurons per mm³ in human V1. Individual values for each column are represented by “o.” D) Average density of PV neurons per mm³ is shown by the height of the individual bars, the error bar indicates the SEM. Average density determined for 13 columns from 7 individuals.

Table 2.

Human neuronal density by layers (neurons/mm³) for each of the 7 human (H1–H7) and each sampling column (b1–b2).

Layers human	1	2/3	L4A (3Bβ)	L4B (3C)	4Cα (4A)	4Cβ (4B)	5A	5B	6A	6B
H1b1	9,452	82,280	85,037	56,363	32,710	88,841	80,877	52,956	97,829	56,751
H1b2	5,770	73,387	83,880	46,950	72,110	62,584	80,574	53,330	79,442	29,064
H2b1	5,929	99,198	125,100	52,176	45,051	100,085	90,912	43,480	55,115	37,483
H2b2	9,970	116,903	117,293	78,631	88,500	160,004	124,186	60,389	111,051	59,944
H3b1	11,862	92,077	97,066	56,237	84,266	129,453	106,336	44,439	116,464	50,432
H3b2	10,909	77,154	123,653	41,409	73,163	123,365	104,200	27,719	75,860	19,991
H4b1	13,420	111,650	97,927	43,135	80,903	135,321	87,177	54,771	90,583	16,775
H4b2	13,396	100,959	142,620	36,454	89,931	157,873	76,429	50,726	57,034	22,792
H5b1	29,439	125,158	98,125	11,008	59,834	173,919	56,268	13,877	12,788	3,450
H5b2	2,571	120,349	141,436	25,427	110,406	155,070	77,109	57,210	73,771	12,437
H6b2	16,541	88,651	165,519	42,720	116,544	175,857	76,729	51,762	6794	24,155
H6b3	17,137	87,863	184,344	31,970	143,025	266,047	173,617	78,150	64,702	6,836
H7b2	8,150	118,999	128,712	50,507	73,349	137,485	128,097	42,529	84,496	6,334
Mean	11,888 (16%)	99,587 (133%)	122,362 (164%)	44,076 (59%)	82,292 (115%)	143,531 (194%)	97,116 (125%)	48,564 (66%)	71,225 (95%)	26,650 (34%)
SEM	2,746	5,036	9,858	3,954	8,720	15,163	9,355	5,954	11,206	6,884

Table 3.

Human neuronal density (neurons/mm³) when peaks or valleys are aligned for each of the 7 human (H1–H7) and each sampling column (b1–b2).

Layers human	2	3/4A (3/3Bβ)	4B (3C)	4Cβ (4B)	4Cβ /5 (4B/5)	5
H1b1	144,620	105,680	38,935	105,680	38,935	116,810
H1b2	133,805	118,940	47,575	101,100	35,681	83,260
H2b1	160,258	133,990	42,035	139,240	36,780	73,560
H2b2	146,830	136,190	65,970	193,640	61,710	114,900
H3b1	124,016	124,020	44,646	146,340	59,528	124,020
H3b2	93,017	134,870	44,183	127,900	27,905	79,060
H4b1	183,290	133,650	30,548	152,740	45,822	133,650
H4b2	189,269	182,740	39,159	169,690	39,159	97,900
H5b1	193,376	85,940	7162	143,240	21,486	21,486
H5b2	149,912	176,820	34,595	188,350	57,659	80,720
H6b2	226,371	193,240	27,606	218,090	5,521	57,970
H6b3	184,820	229,170	22,178	295,700	36,963	96,100
H7b2	175,487	152,840	45,287	147,180	28,304	90,570
Mean	161,929	146,776	37,683	163,761	38,112	90,001
SEM	10,438	11,222	4,265	14,706	4,772	9,541

Supragranular layers

In the supragranular layers of cortex, layer 1 had a significantly lower density than all the other layers, 12 × 10³ neurons/mm³ (One-way ANOVA, Supplementary Table 3). The low density in layer 1 is a characteristic of this layer in all mammals (Marin-Padilla and Marin-Padilla 1982; Gabbott and Somogyi 1986; Balaram and Kaas 2014). It is common to assign layers 2 and 3 together in density estimates. When this was done, the mean density across individuals was 100 × 10³ neurons/mm³, which was significantly lower than layer 4Cβ and higher than layers 4B, 5B, and 6B (Fig. 3A; Table 2; One-way ANOVA, Supplementary Table 3). However, within layers 2/3, there are quite consistent variations in density. In all the cases studied, there was a local density peak in layer 2 (Fig. 3B, pink trace) followed by a trough and then another peak in lower layer 3 near the layer 3/4A border (Fig. 3B, orange trace). The local density in layer 2, when all the continuous estimates were aligned by the layer 2 peaks (Fig. 3B), reached a mean value of 162 × 10³ neurons/mm³ (Table 3). This suggests that there is sublayer organization that is not captured by the laminar density estimates. A second peak in lower layer 3—probably including layer 4A—had a density of 147 × 10³ neurons/mm³ when all the columns were individually aligned (Fig. 3B, orange trace; Supplementary Table 3).

Granular layers

The layer with the highest mean laminar density was layer 4Cβ with 144 × 10³ neurons/mm³ (Fig. 3A), which was almost twice the average density across the whole cortex. This density was significantly higher than all layers except for layer 4A (Table 2, Supplementary Table 3). There was also considerable variability between individuals (Fig. 3A); H6 had an average density of 221 × 10³ neurons/mm³ in layer 4Cβ, whereas H1 had an average density of 76 × 10³ neurons/mm³ (Fig. 3A; Table 2). The within-individual values were considerably less variable.

The other main TC recipient layer, 4Cα, where the main excitatory cell type is also the spiny stellate cell, had an average density of 82 × 10³ neurons/mm³ that was significantly different from layers 1, 4A, 4B, and 6B (Fig. 3A, Table 2, Supplementary Table 3). The density ratio between 4Cα and 4Cβ was 0.60. This ratio was remarkably consistent across individuals ranging from 0.52 to 0.69. When the densities were aligned to the peak in 4Cβ, the mean density at the peak depth bin was 164 × 10³ neurons/mm³ (Fig. 3B, purple trace; Table 3).

The precise boundaries of layer 4A are difficult to define in human visual cortex because there are no distinct upper or lower layer boundaries such as those provided by the honeycomb arrangement of CO (Horton 1984; Garcia-Marin et al. 2013) and vGlut2 (Garcia-Marin et al. 2015) that provides clear boundaries in macaque V1. We defined layer 4A at low power as a thin strip of high neuronal density, below layer 3. Using these criteria to determine the layer 4A boundaries, our best estimate of the layer 4A density was 122 × 10³ neurons/mm³ (Fig. 3A and B, Table 2). The neuronal density in this layer was significantly higher than in layers 4B, 4Cα, 5B, and 6A-6B. Finally, layer 4B had an average density 44 × 10³ neurons/mm³ (Fig. 3A), and when the 4B troughs were aligned, the density was reduced to 38 × 10³ neurons/mm³ (Fig. 3B, green trace; Table 3).

Infragranular layers

The highest densities are in the upper halves of layers 5 and 6 (97 × 10³ neurons/mm³ and 71 × 10³ neurons/mm³, respectively, Table 2). The neuronal density in layer 5A was significantly higher than in layers 5B and 6B. The average density in layer 6B (27 × 10³ neurons/mm³, Table 2) was lower than in layer 5B, and significantly lower in layer 6B than in 6A (Supplementary Table 3). There is a consistent valley in layer 5 and when these valleys were aligned, the density at the trough was 38 × 10³ neurons/mm³ (Fig. 3B, brown trace; Table 3). Finally, there is a local peak at the layer 5/6 border (Fig. 3B, blue trace; Table 3) where the density rises to 90 × 10³ neurons/mm³.

Comparison with layer distribution in macaque

Recently, we studied the neuronal laminar density in macaque V1 (Kelly et al. 2019), using the same methodology as in the current study. To compare the laminar densities between macaque and human V1, we used the mean density across all layers to normalize the layer data and then compared the densities within layers as deviations around the mean (Fig. 4A). Although the absolute density in macaque V1 is about 3× greater than the density in human V1, the relative distribution across layers within cortex is remarkably similar between macaque and human (Fig. 4A).

Fig. 4.

Comparison of the laminar distribution of neurons between human and macaque. A) Relative percentage of Neun-ir neurons in each layer. B) Relative percentage of PV neurons (PV-ir). The values shown for each layer were determined as a percentage relative to the total average cortical density (shown for visualization as the gray horizontal line at 100%). Macaque data from Garcia-Marin et al. 2019, Kelly et al. 2019. Error bars indicate the SEM.

Using the same z-stacks, we calculated the volume of neuropil occupied by neuronal somata and proximal dendrites in the NeuN sections using ImageJ. In human, neurons occupied 7.5% of the neuropil, on average, across all the layers. Using the data from macaque (Kelly and Hawken 2017), we also calculated the volume of neuropil occupied by the neuronal cell bodies and proximal dendrites in the NeuN sections across all layers. In macaque, the soma of the neurons and their proximal dendrites occupied 15.5%, suggesting that there is more neuropil volume in human than macaque that could be occupied by other elements rather than soma and proximal dendrites of neurons.

Distribution of PV neurons

PV-ir interneurons comprise the largest population of inhibitory interneurons in cortex (Rudy et al. 2011). Previously, in macaque V1, we found that the population of GABAergic interneurons was about 11% of the total population and PV-ir neurons were ~52% of the GABAergic population, therefore 5–6% of the total population (Kelly et al. 2019). In the current study, we found a total density of 4.6 × 10³ PV neurons/mm³, which represents 5.8% of the total neuronal population (Table 1), similar to the percentage that we previously reported for macaque (Kelly et al. 2019).

We observed the highest density of PV neurons in 4Cβ with 10.3 × 10³ neurons/mm³ followed by layer 4A with 9.9 × 10³ neurons/mm³ and layer 4Cα with 7.0 × 10³ neurons/mm³ (Fig. 3C-D; Table 4). Layers 2/3, 6A, 5A, and 5B have lower densities ranging from 2 × 10³ to 4 × 10³ neurons/mm³. The lowest densities were found in layers 1 and 6B, with 1.0 and 0.9 × 10³ PV neurons/mm³, respectively.

Table 4.

Mean PV density (neurons/mm³) by layers across all human (H1–H7) and columns (b1–b2), and the percentage of PV in each layer with respect to the mean PV density across all layers. (Mean PV density across all layers 4605 neurons/mm³ (Table 1)).

Layers	1	2/3	L4A (3Bβ)	L4B (3C)	4Cα (4A)	4Cβ (4B)	5A	5B	6A	6B
Mean	978 (23%)	4,310 (99%)	9,919 (229%)	1,856 (43%)	7,014 (162%)	10,263 (237%)	2,624 (61%)	2,110 (49%)	3,366 (78%)	915 (21%)
SEM	444	503	1,418	508	1,224	1,781	785	764	707	376

When we compared the normalized densities across layers in human to those in macaque V1 (Kelly et al. 2019), we found that the PV neurons within each layer—as a proportion of the mean density across all layers—were remarkably similar between the human and the macaque (Fig. 4B). A similar close correspondence was found for the total neuron density (Fig. 4A). However, there is nearly a 2-fold difference in the proportion of neurons that are PV-ir in the putative output layers 2/3, 4B, 5A, 5B, 6A, and 6B, where about 4% of neurons are PV-ir, compared with input layers 4A, 4Cα, and 4Cβ, where more than 8% of neurons are PV-ir.

Connectivity

We next sought to compare not just differences in density between humans and macaque monkeys but the composition of comparable processing units. The cortex is often thought of as a collection of repeating columnar structures. At the finest scale, the processing unit is the cortical microcolumn (Mountcastle et al. 1957; Mountcastle 1997). In primary visual cortex, at a meso-scale, the cortical ODC, structurally demarcated by cytochrome oxidase patches (Horton and Hedley-Whyte 1984; Adams et al. 2007), is considered an important processing unit (Hubel and Wiesel 1977; Lund et al. 2003) that contains representations of the full range of orientations at a range of spatial scales for one eye.

Adams et al. (2007) found that ODC widths in human V1 were, on average, 863 μm. If it is assumed that each eye dominance patch defines a column (Hubel et al. 1978) then, in human cortex where the vertical extent (pia to white matter) is on average 2,700 μm, an ODC volume is 863 μm × 863 μm × 2700 μm. In the cynomolgus macaque monkey, the ODC volume is 530 μm × 530 μm × 1500 μm (Horton and Hocking 1996). The ratio of the ODC area tangential to the cortical surface is 2.7 (human:macaque) and the ratio of ODC volumes is 4.8 (human:macaque).

There is some debate as whether each ODC has the same dimensions as an orientation domain (see Mazade and Alonso 2017 for summary across species). Nonetheless, the human/macaque ratio of surface area is likely to be similar for orientation domains and ODCs, and the difference in thickness of cortex between the human and macaque is well established (Wagstyl et al. 2015; Alvarez et al. 2019). In what follows, we will use the ODC to compare between human and macaque.

We estimated the total number of neurons under one ODC (Table 5). Using the total neuronal density calculated in the current study (Table 1), we determined that there were 159 × 10³ neurons in one ODC (Table 5) in human V1. The same calculation applied to our previous macaque data (Garcia-Marin et al. 2019) showed that there were 96 × 10³ neurons in one ODC. Consequently, there are 1.7 more neurons in an ODC in human V1 than in macaque. Although there was a considerable range in the ratio of the number of neurons in an ODC when comparing within layers between human and macaque, two factors were of considerable interest. The principal thalamic recipient layers (4Cα and 4Cβ) had about the same number of neurons in an ODC in human as in macaque. In contrast, the output layers had a greater number of neurons in the human ODC when compared with macaque. Because we could not easily distinguish a layer 4A/4B boundary in human V1, we summed the neurons in these two layers (4A and 4B) and compared the number to the sum from the macaque; there were 2.4× the number of neurons in the human ODC. Layers 2/3 and 5 had 1.7× and 1.9× more neurons in a human ODC. Layer 6 had 1.2× the number. Layer 1, which has considerably greater thickness in human than in macaque, has 3.8× more neurons in the human ODC.

Table 5.

Calculations of the total number of neurons, excitatory neurons, and inhibitory neurons per ODC in human by layers.

Layer	Layer width (mm)	Density (neu/mm3)	Neurons/ODC (0.86 × 0.86 × width)	GABA %	Excitatory neurons/ODC	Inhibitory neurons/ODC
1	0.281	11,888	2,471	67	815	1,655
2/3	0.722	99,587	53,178	13	46,265	6,913
4A (3Bβ)	0.208	122,362	18,914	12	16,645	2,270
4B (3C)	0.301	44,076	9,812	12	8,635	1,177
4Cα (4A)	0.205	82,292	12,477	13	10,855	1,622
4Cβ (4A)	0.192	143,531	20,382	11	18,140	2,242
5A	0.163	97,116	11,708	8	16,670	1,450
5B	0.162	48,564	6,142
6A	0.238	71,225	12,063	6	16,000	1,021
6B	0.238	26,650	4,691
	2.7	79,221	159,304	11	141,780	17,523

Next, we asked whether the increased number of neurons in the human ODC is because there are more minicolumns per ODC or more neurons in each minicolumn. Minicolumns are usually defined by a central core area that contains the majority of the neuronal cell bodies, apical dendrites, and myelinated afferent fibers, flanked by cell-poor areas (peripheral neuropil space) that are rich in unmyelinated axon fibers, dendritic arborizations, and synapses (Jones and Burton 1974; Szentagothai 1978; Seldon 1981, 1982; Ong and Garey 1990; Peters and Payne 1993; Peters and Sethares 1996; Mountcastle 1997). Previously, we used the Vesicular Glutamate Transporter-2 (vGluT2), a specific marker for thalamic afferents, to measure the minicolumnar width as the average center-to-center spacing of two adjacent peripheral neuropil columns that are vGluT2-ir dense (Garcia-Marin et al. 2013). We found that the average minicolumn width in human and macaque V1 was 24.7 μm and 18 μm, respectively (Garcia-Marin et al. 2013).

Currently, the exact perimeter or shape of the cortical columns is not well defined. The important factor is the center-to-center spacing between columns that determines the mosaic (Adams et al. 2007). Here, we have adopted the square shape as proposed in the conceptual columnar arrangement of Hubel and Wiesel (1977) for our numerical estimates. Assuming a square array distribution of the minicolumns, we estimated that the number of minicolumns per ODC was 1221 for human and 867 for macaque (Fig. 5). Using the total number of neurons per ODC, we estimated that there were 130 neurons per minicolumn in each human ODC and 111 neurons per minicolumn in each macaque ODC. These estimates indicate that at the finest columnar scale—the minicolumn—there is a close match in the number of neurons between human and macaque. Yet at the mesoscale, the human cortex has a greater number of neurons because there are more minicolumns per ODC than in the macaque. Furthermore, the increase in number is disproportionately distributed in the supra and infragranular layers.

Fig. 5.

Schematic summary of the neurons within a hypercolumn (HC) in human and macaque. A hypercolumn is composed of 1,221 minicolumns in human and 867 minicolumns in cynomologus macaque. For each layer, the number of excitatory (Ne) and inhibitory (Ni) neurons is calculated and shown as the number of neurons in each minicolumn in each layer. Layers 2/3, 4Cα, and 4Cβ are expanded twice their length and width to give a clearer impression of the distribution of neurons within a single minicolumn in each of these layers, excitatory neurons (filled circles), and inhibitory neurons (open circles).

Discussion

Areal and laminar neuronal density in V1

In the current study, we found that the neuronal density was about 10–15% greater than in previous studies that have used stereological methods for neuronal counting that have more than three subjects (Everall et al. 1993—n = 13; Pakkenberg and Gundersen 1997—n = 63; Dorph-Petersen et al. 2007—n = 9). Several studies have determined the distribution of neuronal density between males and females (Pakkenberg and Gundersen 1997; Rabinowicz et al. 1999, 2002). In the largest study of gender on density in V1, Pakkenberg and Gundersen (1997) reported about a 6% difference in mean density between females and males (70.9 vs 66.9 × 10³ neurons/mm³, Supplementary Table 1). To allow for a large enough sample and given the availability of postmortem tissue, we studied only male subjects. When we compare the density measurements from males from the three prior studies with the results of the current study (also all males), the ratio of densities (current study/previous study) ranges from 0.84 to 0.89 (Supplementary Table 1). This confirms that the current method provides about 10–15% higher density estimates than earlier studies when taking gender differences into account. In the three earlier studies, the coefficients of variation were in the range of 0.14–0.23, again similar to the variation found in the current study (0.13). Another study that used a 3D counting regime (Selemon et al. 1995) reported substantially higher mean density (1.55–1.8× higher) and lower CV (0.08) than the three studies using stereology and our current study.

One reason for this difference in density may be the higher fidelity of specifically labeling the nuclei of neurons using the immunocytochemical labeling protocol rather than having to make a judgment about whether a Nissl stained cell is a neuron or nonneuronal cell.

There have been no previous studies that have determined laminar density where the cortex has been divided into the layer structure that includes the accepted subdivisions for layer 4 (Preuss and Coleman 2002) that match those in macaque monkey (Lund 1973). Nor did any of the studies using stereology undertake a laminar analysis. Hence, the current study is the first comprehensive description of the variation of density across layers in the parafoveal regions (2.5°–5° eccentricity) of human V1. Further studies need to be done to study the laminar neuronal density at different eccentricities in V1, as it has been reported that the neuronal density and cortical thickness vary across the V1 retinotopic map, with higher neuronal density and cortical thickness in the central region versus the periphery (Collins et al. 2010, 2016; Alvarez et al. 2019).

There was a clear variation in density between layers (Table 2; Fig. 3A) where the highest densities were found in upper layer 2, layer 4A, and layer 4Cβ (Table 3; Fig. 3B). The peak densities in these three sublayers were around 150,000 neurons/mm³ (Table 3), almost double the average density across all layers. In contrast, the low-density regions, layer 4B and at the layer 4Cβ/5 border, had densities of just less than 40,000 neurons/mm³, a density that is about half the average across cortex. These laminar variations were consistent between individuals (Table 3; Fig. 3B). The mean density between individuals showed a coefficient of variation of 0.14 which is a close match to those studies using stereology (see Supplementary Table 1).

Glia:neuron density ratio

We found that the average nonneuronal cell to neuron ratio for human V1 was 1.1, lower than the ratio (~1.4) reported with the isotropic fractionator in the cortical gray matter (Azevedo et al. 2009). Since glia, on average, represents about 70% of nonneuronal cells, we estimated that the glia to neuron ratio was about 0.76 in human V1. The V1 glia:neuron ratio in our current study is smaller than the values reported using stereological counting methods in the neocortex as a whole (1.3–1.7, Pelvig et al. 2008, Pakkenberg et al. 2003). It is worth noting that this glia to neuron ratio is not constant across all cortical areas; as noted by Azevedo et al. (2009), the ratio changes depending on the density of neurons; the extreme example is the cerebellum where neuronal density is high and the glia to neuron ratio drops to 0.23. V1 has a neuronal density that is greater than other cortical areas therefore, following the arguments of Azevedo et al. (2009), it might be expected that V1 has a lower glia to neuron ratio than the neocortex as a whole, as our results demonstrate. The glia to neuron ratio also varied through the thickness of cortex. We found an inverse relationship between neuron density and glia:neuron ratio (Supplementary Fig. 1B). This largely reflects that glial and nonneuronal densities are relatively constant through the thickness of cortex (Supplementary Fig. 1A), while neuron density shows local peaks and valleys (Fig. 3B, Supplementary Fig. 1A).

The nonneuronal:neuronal ratio has also been determined in different studies of macaque of V1 with a ratio of 0.44 (Giannaris and Rosene 2012; Kelly and Hawken 2017). This is considerably lower than the ratio we found in human in the current study in human V1 (1.1). If we assume that about 70% of the nonneuronal cells are glial cells, as we have in human, then the ratio of glia:neurons is 0.3 in macaque V1, using the same methods as in the current study (Kelly and Hawken 2017). Our value is slightly smaller than the previous glia:neuron ratios reported in macaque V1 (0.49, O'Kusky and Colonnier 1982; 0.58, Christensen et al. 2007; 0.44, Giannaris and Rosene 2012) probably due to the different sampling methods. Our values of glia:neuron ratio in macaque V1 is smaller than the glia:neuron ratio reported in the whole cortex (0.56, Christensen et al. 2007), as was expected from an area with a high neuronal density. The glia:neuron ratio observed in the current study is also different from values reported in other cortical areas (Christensen et al. 2007: temporal, 0.46; frontal, 0.82; Sherwood et al. 2007: a4, 1.29; a9, 0.84: a32, 0.68; a44, 0.95) confirming the idea that glia:neuron ratio in macaque also varies across different cortical areas, as was observed in human (Christensen et al. 2007; Sherwood et al. 2007).

The difference between the glia:neuron ratio in human and macaque V1 (0.76 vs 0.33, respectively), when the same methods are used, indicates that there is an increase in the glial density in human V1. Glial cells, particularly astrocytes, play a crucial role in the flux of energy substrates to neurons by regulating the rate of glucose uptake and phosphorylation in response to glutamate concentrations in the synaptic cleft (Tsacopoulos and Magistretti 1996). Thus, the glia-to-neuron is considered an indirect measure of the metabolic support supplied to the neurons in one brain region (Sherwood et al. 2006). A larger glia–neuron ratio has also been reported in the frontal lobe (1.65 vs 0.84, human and macaque, respectively) (Sherwood et al. 2006). The higher glia:neuron ratio in human may suggest that human brain requires a higher metabolic demand to support neurons with larger dendritic fields (Elston and Rosa 1998; Elston et al. 2001, 2005). As humans possess lower neuronal density, a larger neuropil volume could accommodate expanded dendritic fields. Moreover, a higher glial density would be necessary to compensate for the increased metabolic demand in the human cortex.

Higher glial density in humans than in macaques could be also an evolutionary advantage for humans in the context of the visual system. Glial cells play an important role in the development, function, and plasticity of the visual system (see Benfey et al. 2022). Particularly in V1, computational models have demonstrated that by altering the radius of astrocytes, one can alter the radius of lateral excitatory connections, ultimately modifying the size and presence of orientation preference maps in the visual cortex (Philips et al. 2017). It is likely that astrocytes contribute to the formation of functional maps throughout the visual system in multiple ways given that they express important axon guidance molecules such as ephrins and their receptors (Murai and Pasquale 2011) and form spatially compartmentalized networks through gap junctions (Houades et al. 2008; Roux et al. 2011, review in Benfey et al. 2022).

Density of PV neurons across layers

Several studies have assigned a critical role to inhibitory neurons in controlling the flow of activity within local circuits, including maintaining excitatory/inhibitory balance (van Vreeswijk and Sompolinsky 1996), shaping receptive fields (Sillito 1975; review in Ferster and Miller 2000), gain control (Atallah et al. 2012; Wilson et al. 2012), and surround suppression (Bair et al. 2003; Angelucci and Bressloff 2006; Adesnik et al. 2012). For a comprehensive understanding of the interaction between excitation and inhibition in the neocortex, it is crucial to obtain data on the distribution of excitatory and inhibitory neurons in a cortical column and within a cortical layer. The fixation protocol that was used for the human postmortem tissue was incompatible with GABA immuno-histochemistry, and consequently, we were not able to measure the total density of GABAergic interneurons directly. PV-ir interneurons account for the largest subpopulation of the inhibitory interneuron population in rodent somatosensory cortex (Rudy et al. 2011), as well as in human and macaque V1 (Leuba et al. 1998; Kelly et al. 2019). We found that the average density of the PV population was 4.6 × 10³ neurons/mm³, a value that accounts for ~ 6% of the total neuronal population. Similar percentages of PV had been reported previously in the catarrhines, including human and macaque, in different areas (Leuba et al. 1998; Gabbot and Somogyi 1986; Glezer et al. 1998; Smiley et al. 2016; Kelly et al. 2019; Sherwood et al. 2007). If the PV population makes up ~50% of the total inhibitory population, as is the case for macaque and human V1 (Leuba et al. 1998; Kelly et al. 2019), then we estimate that the total proportion of inhibitory neurons would be about 12%. Recently, we estimated the density of GABAergic neurons in macaque V1 and found that they represent 11% of the total neuronal population. In human V1, Leuba et al. (1998) quantified the density of the other calcium-binding proteins, Calbindin (CB) and Calretinin (CR). Comparing their GABAergic density (from their measurements of PV, CB, and CR) with our neuronal density, interneurons would represent 9.3% of the total neuronal population. Within human V1, the data indicate that the total inhibitory neuronal population is between 9% and 12% of the total neuronal density. Hence, our numerical estimates provide quite narrow bounds on the density of inhibitory neurons in V1. Nonetheless, caution should be taken when extrapolating these numbers to other cortical areas, as it has been reported that the PV-ir neurons do not represent the largest inhibitory subpopulation in prefrontal or primary auditory areas in primates (Conde et al. 1994; Smiley et al. 2016; Fish et al. 2018).

Interareal comparisons

In catarrhines in particular and all mammals in general, the relative density in different cortical areas is thought to be of central importance from a number of views on how cortex functions. From a developmental viewpoint, the neuronal density in each cortical area provides a strong prediction of the likely connectivity profile of that area with other cortical areas (Finlay et al. 2001). When predictions are made within this framework they are, by necessity, made within a single species of primate as the neuronal density expressed in neurons/mm³ of cortex shows a wide range between species. Within the catarrhines, we have shown that the average V1 neuronal density in macaque monkey is 230 × 10³ neurons/mm³ (Kelly and Hawken 2017; Garcia-Marin et al. 2019; Table 6) and, in human, the density is 79 × 10³ neurons/mm³ (Table 5). Nonetheless, the regional areal variation within a species is the important factor when considering the developmental connectomic viewpoint (Finlay et al. 2001; Barbas 2015; Atapour et al. 2019). From this perspective, there are currently few studies of human cortex that provide a comprehensive account of the regional areal densities of neurons. Using the methods in the current paper would provide a basis for such a series of studies.

Table 6.

Calculations of the total number of neurons, excitatory neurons, and inhibitory neurons per ODC in macaque by layers.

Layer	Layer width (mm)	Density (neu/mm3)	Neurons/ODC (0.53 × 0.53 × width)	GABA %	Excitatory neurons/ODC	Inhibitory neurons/ODC
1	0.112	20,700	651	67	215	436
2/3	0.434	257,800	31,429	13	27,343	4,086
4A (3Bβ)	0.056	268,800	4,228	12	3,721	507
4B (3C)	0.154	173,800	7,518	12	6,616	902
4Cα (4A)	0.172	235,000	11,368	13	9,891	1,478
4Cβ (4B)	0.172	411,000	19,838	11	17,656	2,182
5	0.154	213,000	9,214	8	8,477	737
6	0.238	214,500	14,340	6	13,480	860
	1.5	229,100	96,451	11	85,841	10,610

Functional considerations

A complementary view of the importance of cortical density is in terms of cortical function (Carlo and Stevens 2013; Srinivasan et al. 2015). In the visual system of the catarrhines, there is a very clear homology across the different members of the parvorder that includes humans, nonhuman apes, and macaque monkeys. In the cortex, the commonly used measure of neuronal density (neurons/mm³) works well for the within-species areal comparison from the developmental viewpoint but is not well suited for a comparison of cortical function, especially when the comparison is between species (Herculano-Houzel et al. 2008). Functionally, the columnar organization of cortex has been a fruitful viewpoint. Here, we argue that one of the most frequently adopted columnar measures (Rockel et al. 1980; Dombrowski et al. 2001; Carlo and Stevens 2013; Srinivasan et al. 2015; Atapour et al. 2019), the neuronal population under a mm² column of cortex from the pial surface to the white matter, is not the most appropriate measure when making comparisons across species from a visual-processing functional viewpoint.

Neural populations in an ODC

As clearly discussed by Rakic (2008), there are numerous types of columns or modules. In mammalian vision, a module that has been of interest since its discovery and subsequent elaboration is the ODC (Hubel and Wiesel 1962, 1968; Hubel et al. 1978; Horton and Hedley-Whyte 1984; Adams et al. 2007) that is thought to contain the circuits for the emergence of orientation preference and spatial processing of a point image (Carlo and Stevens 2013), in many mammalian species. In the current study, we determined the population of neurons within each layer underlying an eye dominance module to compare with the size of the populations underlying the same module in macaque (Garcia-Marin et al. 2019) thereby providing a comparison based on functional similarity rather than evaluating the uniformity (Rockel et al. 1980; Carlo and Stevens 2013) or nonuniformity (Herculano-Houzel et al. 2008; Lent et al. 2012) of cortex hypotheses.

The lattice of the ODCs can be determined from CO-stained sections of flatmounted V1, and the average size of the OD domain can also be determined. The area occupied by the macaque OD domain is smaller than that of the human, and the ratio of areas is about 0.39. In cynomolgus macaque, the ODC width is, on average, 0.53 mm (Horton and Hocking 1996), while in humans, it is considerably larger, about 0.86 mm (Adams et al. 2007). In human, the number of neurons through the thickness of cortex in a single ODC is 159 × 10³ (Table 5) compared with 96 × 10³ for the similar column in macaque (Table 6, see Table 7 for calculations). Using the ODC as a core processing unit, human cortex has 1.7× more neurons than macaque for undertaking the same local computations (Fig. 5).

Table 7.

Relationship between neurons and ODC in human and macaque.

		Human	Macaque	H/M ratio
[A]	Neuronal density (neurons/mm 3 ) (Current study,  Garcia-Marin et al. 2019)	79,221	229,100	0.3
[B]	ODC volume (mm 3 ) (Horton and Hocking 1996;  Adams et al. 2007)	0.863*0.863*2.70 = 2.011	0.530*0.530*1.50 = 0.421	4.8
[C] = [A]*[B]	# neurons/ODC	159,303	96,531	1.7
[D]	Minicolumn volume (mm 3 ) (Garcia-Marin et al. 2013)	0.0247*0.0247*2.70 = 0.00165	0.018*0.018*1.5 = 0.00049	3.4
[E] = [B]/[D]	# Minicolumns/ODC	1,221	867	1.4
[F] = [A]*[D]	Neurons/minicolumn	130	111	1.2
[G]	Total area of V1 (mm 2) (Adams et al. 2007)	2,637	1,189	2.2
[H]	% total area devoted to macula representation (central 15 deg)   (Adams et al. 2007) Total area devoted to macula representation (mm2)	53 1,397	61 725	1.9
[I]	ODC area (mm 2 ) (Horton and Hocking 1996;  Adams et al. 2007)	0.863*0.863 = 0.745	0.530*0.530 = 0.281	2.7
[J] = [H])/[I]	# ODCs in macula	1,877	2,582	0.7
[K] = [C]*[J]	# neurons in macula representation	299 × 106	249 × 106	1.2

Neural populations in central vision

The next question we addressed was the number of ODCs in a large spatial extent of visual cortex and their underlying neural populations. Often, the visual cortex is divided into the central and peripheral regions (Adams et al. 2007). In this parcellation, the central region is termed the macula region—the central 15° of eccentricity from the fovea to the optic disc. The remaining cortex contains a representation of the peripheral visual field, including the monocular crescent. The total area of V1 in the human is 2.2× larger than in macaque (2637 vs 1189 mm², respectively) (Adams et al. 2007). When comparing the macula representation of the central 15°, 53% of the striate cortex is devoted to this central representation in humans versus 61% in macaques (Adams et al. 2007). As a result, in the macula region, macaques have more ODC than humans (2,580 vs 1,877 ODCs, macaque and human, respectively). Each single human ODC has 1.7× more neurons than each macaque ODC (Tables 5–7); however, since humans have fewer ODCs in the same central region, humans only have 1.2× more neurons in the central 15° than macaque (299 × 10⁶ neurons vs 249 × 10⁶ neurons, human and macaque, respectively). Therefore, the number of neurons available for the first stage of cortical processing in central vision is ~20% greater in humans than macaques, even though the difference in volume is 4× greater in humans.

Implications of allometric scaling differences between humans and macaques

Across mammalian species, differences in brain size are systematically related to various other aspects of neural organization and structure. These relationships may derive from differences in the duration of development across species: members of species with longer developmental trajectories will develop larger brains, larger volumes devoted to analogous brain structures, and proportionally larger isocortex, among other co-regulated differences (Finlay and Darlington 1995; Cahalane et al. 2014). By this model, the much larger size of the human brain compared with the macaque is thought to derive from a slower developmental time course (Workman et al. 2013) and, therefore, is predicted to have more V1 neurons overall and more neurons per column. Although the most relevant aspects of development—the rate and duration of neurogenesis (Cahalane et al. 2014)—are difficult to establish in humans, the timing of other developmental landmarks is generally consistent with a slower developmental trajectory in humans (Clancy et al. 2001). Neuron density in human V1 (79 × 10³ neurons/mm³) is only 34% of that in macaque (229 × 10³ neurons/mm³) and, correspondingly, the number of neurons under 1 mm² column of cortex in human V1 is 62% of the number in macaque (214 × 10³ neurons in human vs. 343 × 10³ neurons in macaque; Supplementary Table 4). This seeming violation of allometric scaling may reflect that not only neuron number but also the volume devoted to the same region scale with body size, and both contribute to the calculation of density. Our results are qualitatively consistent with the allometric scaling prediction when considering the number of neurons allocated to the representation of the macula (20% higher in human V1; Table 7) and the number of neurons per ODC (66% higher in human V1; Table 7). Future work measuring the numbers of neurons per V1 ODC in other species would be useful to evaluate whether, in general, this aspect of structure follows an allometric scaling pattern, which would imply the possibility that it is regulated by conserved developmental programs rather than modified by specific evolutionary selection pressure (Cahalane et al. 2014).

Minicolumns, Soma size, and neuropil

We found that the number of minicolumns in each ODC was about double in human compared with macaque, while the number of neurons in each minicolumn was nearly constant: 111 neurons in macaque and 130 in human. Because individual minicolumns in humans and macaques have about equal numbers of neurons and the volume of each minicolumn is 4× larger in human than macaque (Fig. 5), it is likely that the perikaryon is expanded in human cortex and the additional neuropil volume could be occupied with other elements: dendrites, synapses, and glial cells. Analysis of Lucifer Yellow-labeled pyramidal cells of layer 3 reveled that the cross-sectional area of the soma of macaque V1 neurons is smaller than soma area of human V1 neurons (118 vs 154 μm², macaque and human, respectively) (Elston and Rosa 1998; Elston et al. 2001, 2005; Oga et al. 2016; Benavides-Piccione et al. 2023). The same Lucifer Yellow studies have also compared the basal dendritic field areas between human and macaque. These studies showed that the basal dendritic field area of the labeled layer 3 pyramidal neurons was also larger in human than macaque (Elston and Rosa 1998; Elston et al. 2001, 2005; Oga et al. 2016; Benavides-Piccione et al. 2023). We can conclude that both neural soma volume and dendritic arbor size are expanded in human to occupy the larger proportion of the neuropil. More studies are needed to determine if there are also more synapses in the human neuropil than in the macaque neuropil.

Conclusion

In human V1, the total cortical visual field representation and the region representing the macula are expanded compared with macaque, yet the human has a lower neuronal density than macaque. However, using a comparison based on ocular dominance modules, our results give human V1 a 1.2 × advantage in terms of neuronal numbers in the cortical region representing the central visual field—the macula region of the retina. It is important to note that while humans and macaque share many visual perceptual similarities due to our common ancestry (e.g. De Valois et al. 1974; Dahl et al. 2009; Horwitz 2015; Kiorpes 2016; Furtak et al. 2022), there are distinct differences in the nuances of our visual systems that may reflect different ecological and cognitive requirements (Yovel and Freiwald 2013; Horwitz 2015; Ridder et al. 2019; Kim et al. 2023). Having an accurate estimate of the number of neurons in human V1 is also a critical step to determine coding capacities of this area. When a similar visual processing region is compared between humans and an animal model, it will be essential to make rigorous comparisons of the cellular distributions (as in the current study) and their functional characteristics to gain insights into their processing capacity and how they may be altered in dysfunction (Selemon et al. 1995; Dorph-Petersen et al. 2007).

Author contributions

Virginia Garcia-Marin (Conceptualization, methodology, investigation, writing—original draft, review & editing), Jenna G. Kelly (software, software validation, writing—review) and Michael J. Hawken (conceptualization, writing—review & editing)

Funding

This work was supported by the National Institutes of Health (EY17945; P30EY013079; T32EY007136; SC3NS127766).

Conflict of interest statement: None declared.

Data availability

The data that support the findings of this study are available from the corresponding author upon request.