A series of five population‐specific Indian brain templates and atlases spanning ages 6–60 years

Abstract Anatomical brain templates are commonly used as references in neurological MRI studies, for bringing data into a common space for group‐level statistics and coordinate reporting. Given the inherent variability in brain morphology across age and geography, it is important to have templates that are as representative as possible for both age and population. A representative‐template increases the accuracy of alignment, decreases distortions as well as potential biases in final coordinate reports. In this study, we developed and validated a new set of T1w Indian brain templates (IBT) from a large number of brain scans (total n = 466) acquired across different locations and multiple 3T MRI scanners in India. A new tool in AFNI, make_template_dask.py, was created to efficiently make five age‐specific IBTs (ages 6–60 years) as well as maximum probability map (MPM) atlases for each template; for each age‐group's template–atlas pair, there is both a “population‐average” and a “typical” version. Validation experiments on an independent Indian structural and functional‐MRI dataset show the appropriateness of IBTs for spatial normalization of Indian brains. The results indicate significant structural differences when comparing the IBTs and MNI template, with these differences being maximal along the Anterior–Posterior and Inferior–Superior axes, but minimal Left–Right. For each age‐group, the MPM brain atlases provide reasonably good representation of the native‐space volumes in the IBT space, except in a few regions with high intersubject variability. These findings provide evidence to support the use of age and population‐specific templates in human brain mapping studies.


| INTRODUCTION
The shape, size, and volume of the human brain is highly variable across individuals, as well as across age, gender and geographical location, or ethnicity. This fact is of prime importance in neuroimaging group studies, where the brains of all subjects are typically aligned to a single template space for data analysis and for the reporting of findings where analogous anatomical structures are mapped on to the same coordinate location across the subjects. A brain template provides a standard 3D coordinate frame to combine and/or compare data from many subjects, across different imaging modalities, structural, or functional and even different laboratories around the world.
The properties of the template (size, shape, tissue contrast, etc.) directly affect the quality of alignment.
An early brain atlas was constructed by Talairach and Tournoux (1988) from a post mortem brain of one 60-yr-old French woman, introducing the concepts of coordinate system and spatial transformation to brain imaging. However, using a single subject brain as a template introduces several idiosyncrasies, as it does not account for groupwide anatomical variability, asymmetry, age-related differences, and so on. In order to address some of these issues, a subsequent initiative from the Montreal Neurological Institute (MNI) resulted in a statistical brain template (MNI-305) using 305 young right-handed subjects (Evans et al., 1993). While this composite template better accounted for anatomical variability, it also had relatively low tissue contrast and structural definition, which can affect the ability of alignment algorithms to provide high quality anatomical matching across a group study. In 2001, the international consortium for human brain mapping (ICBM) introduced the revised MNI-152 template (Mazziotta et al., 2001b) with better contrast and structure definition, where 152 individual brains were linearly registered to MNI305 to make an average template. The ICBM-452 template (Mazziotta et al., 2001a) included all three sites of ICBM and provided even better signal-to-noise ratio due to the nearly threefold increase in the number of subjects. These MNI templates were widely adopted by several image processing pipelines, with the associated set of coordinates known as "MNI space." Furthermore, an unbiased nonlinear average of the adult MNI152 and a pediatric template with 20-40 iterative nonlinear averages has also been made available (Fonov et al., 2011). These templates provide the advantages of retaining group representativeness of the MNI305 or MNI152 while still providing the details that are closer to those apparent in a single subject; however, their "representativeness" is limited to a fairly isolated geographic location and (typically, Western) population, even though neuroimaging studies draw from populations across the globe.
More recently, several research groups around the world have developed and validated brain templates that are representative of their (broadly) local population. Lee et al. (2005) created a set of Korean Brain templates with 78 subjects in an age range between 18 to 77 years (young template <55 years and elderly template >55 years). Additionally, Tang et al. (2010) generated a Chinese brain template of 56 subjects (mean age 24.4 years). In each case, the groups demonstrated significantly reduced warp deformations and increased registration accuracy when applying these templates to studies of local populations. It should be noted that even though the templates draw from subjects within a population, there is still a large amount of inherent variability evident in the brain morphology, due to combinations of factors such as inherent structural variability, multiethnic composition, and differences in genetic influences and environmental exposures.
The benefit of utilizing a population-representative template in the Indian context has also been recognized, with the additional need for age-specific templates due to the increasingly wide range of ages enrolled in studies. Recent attempts at developing brain templates for Indian population have tended to focus on the young adult age group (21-30 years) with relatively small (Rao et al., 2017) to modest sample sizes (Bhalerao et al., 2018;Pai et al., 2020;Sivaswamy et al., 2019), and have utilized data from a single site/scanner. Additionally, to date, whole-brain annotated reference atlases based on segmentation have not accompanied the generated templates. In this study, we present and validate a new set of brain templates that have been created from a large number of subjects from multisite acquisitions across India, with five age ranges provided (between 6 and 60 years), as well as brain atlases for each template. For each age group's template-atlas pair, there is both a "population average" and "typical" version (the latter being the individual brain which most closely matches the population average, which potentially provides higher detail as an alignment target and atlas). We present several validation tests for the accuracy and representativeness of the templates, and we also use data from separately acquired subjects to demonstrate the benefits of these templates over the existing standard MNI templates for studies on Indian cohorts.

| Participants
The datasets used in the present study were selected retrospectively from healthy control subjects of several imaging studies, across multiple centers and different populations across India. They included imaging data from the ongoing Indian multisite developmental cohort study, the Consortium on Vulnerability to Externalising Disorders and Addictions (cVEDA) Zhang et al., 2020) and from stored datasets contributed by researchers at the National Institute of Mental Health and Neurosciences (NIMHANS, Bengaluru, India). All of these studies were approved by the ethics review boards at the corresponding participating sites and informed consent was obtained from each participant (or from their parent, in the case of subjects below 16 years, along with participant's written assent) with a specific request to collect, store and share anonymized data for research.
Inclusion criteria included not having a personal history of prior brain injury, neurological disorder or psychiatric diagnosis. The sample was comprised of 466 subjects from a large number of states across India and acquired at multiple sites. Based on age and demographic distributions, subject datasets were divided into 5 groups: C1, late childhood (6-11 years); C2, adolescence (12-18 years); C3, young adulthood (19-25 years); C4, adulthood (26-40 years); C5, late adulthood (41-60 years). The sample size and demographic information of each cohort are summarized in Table 1. 2.2 | Image acquisition T1-weighted (T1w) three-dimensional high resolution structural brain MRI scans were acquired from five 3T MRI scanners located at three different locations across India: Bengaluru (site A, C, and D), Mysuru (site B), and Chandigarh (site E). The subjects belonged to several neighboring states to these locations, with wide geographical representation throughout India. As with most multisite studies, the acquisition parameters varied slightly across sites and scanners, but were generally similar, with good gray/white matter contrast with a voxel size close to 1 mm isotropic; details are listed in Table 2.

| Data preprocessing and initial quality assurance
This processing primarily used programs in the AFNI (v19.0.20) (Cox, 1996) and FreeSurfer (v6.0) (Fischl, 2012) neuroimaging toolboxes, as well as the "dask" scheduling tool in Python developed by the Dask Development Team (2016). Unless otherwise noted, programs named here are contained within the AFNI distribution. The following processing steps are shown schematically in Figure 1, in the first column.
Datasets were first processed using AFNI's "fat_proc_convert_ dcm_anat." Using this, DICOMs were converted to NIFTI files using dcm2niix_afni (the AFNI-distributed version of dcm2niix; Li, Morgan, Ashburner, Smith, & Rorden, 2016). For uniformity and initialization, with this tool, they were also given the same orientation (RAI), and the physical coordinate origin was placed at the volume's center of mass (to simplify later alignments).
Next, "fat_proc_axialize_anat" was applied to reduce the variance in the spatial orientation of brains for later alignment and for practical considerations of further processing steps, as described here. Each volume was affinely registered to a reference anatomical template (MNI ICBM 152 T1w) that had previously been AC-PC aligned; alignment included an additional weight mask to emphasize subcortical structure alignment (e.g., AC-PC structures), and only the solid-body parameters of the alignment were applied, so that no changes in shape were incurred. Because datasets had been acquired with varied spatial resolution and FOV (see Table 2), the datasets were resampled (using a high-order sinc function, to minimize smoothing) to the grid of the reference base of 1 mm isotropic voxels.
All datasets were visually and systematically checked for quality of both data and registration using the QC image montages that were automatically generated by the previous program. T1w volumes with noticeable ringing or other artifact (e.g., due to subject motion or dicom reconstruction errors) were noted and removed from further analyses. T1w volumes with any incidental findings (for example, large ventricles, cavum septum pellucidum) were also removed.
In several cases, the skullstripped brain volumes output by recon-all (brain_mask.nii) included large amounts of nonbrain material (skull, dura, face, etc.), and so an alternative mask was generated using only the ROIs comprising the parcellation and segmentation maps. For each subject, a whole brain mask was generated by: first making a preliminary mask from all of the ROIs identified by recon-all; then inflating that premask by 3 voxels; and finally shrinking the result by two voxels (thus filling in any holes inside the brain mask and smoothing the outer edges). This produced whole brain masks that were uniformly specific to each subject's intracranial volume.
Finally, AFNI's 3dUnifize was run on each T1w volume in order to reduce the intensity inhomogeneity (e.g., due to the bias field) and to normalize the intensity of tissues within the volume. This ensures that each subject's brain, which had been acquired on different scanners with potentially different scalings, would have equal weight when averaging (e.g., WM is scaled to approximately a value of 1,000 in each brain, and similarly for other tissues), and also reduces the risk of a bright outlier region driving poor alignment.

| Mean template generation
After the above preprocessing steps and QC, the following templatizing algorithm was applied for each cohort (C1-5) separately. F I G U R E 1 Schematic representation of the steps involved in the Dask pipeline (make_template_dask.py) for generating population-average brain templates The general procedure was to alternate between alignment to a reference base (with increasingly higher order of refinement) and averaging the aligned brains to generate a new reference base for the subsequent iteration. In this way one can generate a cohort mean template of successively greater specificity and detail; after several iterations, the alignment essentially converges (i.e., additional refinement becomes negligible) and is halted. Warps were generated and saved at each step. The final nonlinear warps and affine transformations were concatenated for each subject at the end in order to generate the final group average template. These steps are also included in the schematic Figure 1, in the first column (bottom) and second and third columns.
The first level of alignment was made from each anatomical in the cohort to the MNI ICBM-152 T1w template using a 6 degree of freedom (DF) rigid body equivalent registration, meaning a full affine transformation was computed, but only the rigid components were extracted and applied. The average of all subjects' brains, rigidly aligned to the initial template, was used to create a single average volume "mean-rigid"; here and at each alignment stage, a cohort standard deviation map was also created, to highlight locations of relatively high and low variability. That stage's average volume was then used as a base for the next stage of alignment for each subject, using a 12 DF linear affine registration, and with the results averaged to create the next base "mean-affine." For these alignments, AFNI's "lpa" cost function (absolute value of local Pearson correlation) (Saad et al., 2009) was used for high quality alignment of features between volumes of similar contrast. The cost function computes the absolute value of the Pearson correlation between the volume and the current template in patches of the volume at a time.
As a practical consideration, we note that lower level alignments such as these have a general property of producing a smoothed brain, which has the additional effect of increasing the apparent size of the base dataset (i.e., the edge is blurred outward). Therefore, in these initial levels we added a step to control the overall volume of the template. We calculated the mean intracranial volume (ICV) of all the subjects in the cohort V coh , and then calculated the volume of the initial mean-affine brain mask V aff . The volume ratio r vol = V coh /V aff was calculated, and each of the three dimensions of the mean-affine volume were scaled down by the appropriate length scaling factor r 1=3 vol . In this way, the final volume of the templating process retained a representative size for the cohort.
The next alignment stages were comprised of nonlinear registration using AFNI's 3dQwarp (Cox & Glen, 2013). At each successive level the nonlinear alignment was performed to an increasingly higher refinement, resulting in mean volumes of greater detail. Specifically, nonlinear alignment at each stage was implemented to create mean templates as follows (A-E), using 3dQwarp's default "pcl" (Pearson correlation, clipped) cost function to reduce the effects of any outlier values (and unless otherwise specified, applying a 3D Gaussian blur): [A)].
• mean-NL0: after registering to mean-affine with a minimum patch size of 101 mm and blurring of 0 mm (base) and 9 mm (source); • mean-NL1: after registering to mean-NL0 with a minimum patch size of 49 mm and blurring of 1 mm (base) and 6 mm (source); • mean-NL2: after registering to mean-NL1 with a minimum patch size of 23 mm and blurring of 0 mm (base) and 4 mm (source); • mean-NL3: after registering to mean-NL2 with a minimum patch size of 13 mm and blurring of 0 mm (base) and 2 mm median filter (source); • mean-NL4: after registering to mean-NL3 with a minimum patch size of 9 mm and blurring of 0 mm (base) and 2 mm median filter (source).
Each mean-NL* volume was resized in the same manner as the initial stages, although the correction factors were much smaller here.
Additionally, each mean-NL* volume was anisotropically smoothed (preserving edges within the volume, for detail) using 3danisosmooth, in order to sharpen its contrast for subsequent alignments.
The mean-NL4 volume became the final group mean template for each cohort, as in all cases results appeared to have essentially converged after this number of step. The coordinate system of this mean volume defines the template space for that age group, and is labeled "IBT_C1," "IBT_C2," and so on.

| "Typical" subject template generation
We used the following approach to find the maximally representative individual brain for the mean template from the underlying cohort, in order to generate an additional "typical" template for that space, in complement to the mean template.
To find the most typical subject for the mean template quantitatively, the lpa cost function value from aligning each subject's anatomical to the final mean-NL4 was compared across the group; that is, the degree of similarity of each subject's aligned volume to the mean template base was compared across the cohort. The individual brain in that mean template space with the lowest cost function value was selected to be the "typical template" brain. Alignment results were also visually verified for each typical template. We note that the typical template volume uses the same coordinate system as the mean template, and thus no additional "coordinate space" is created in this process.

| Validation and tests
The fractional volumes of each ROI in the MPM atlases were checked for being representative of each cohort. For this we calculated the logarithm of the relative volume ratio of each ROI: where the numerator is the fractional volume of a given ith ROI in the MPM (i.e., volume of the ROI divided by that template's ICV), and the denominator is the fractional volume of that ith ROI averaged across all N subjects (i.e., for each jth subject, volume of the ROI divided by the subject's ICV, in native space). Thus, r i values close to 0 reflect high similarity of the MPM ROI to the cohort mean, and negative or positive values reflect a relative compression or expansion, respectively, of the MPM ROI relative to that for a particular cohort.
In order to quantify the intersubject brain morphological variability for participants in each age-band, we calculated a region-wise mean deformation value (mDV) from the deformation warp fields generated during nonlinear registration to the age-specific IBT. For this, the absolute warp value was summed across all three axes (L1-norm) and averaged across all the voxels within each ROI in the age-specific MPM atlas. A larger mDV indicates greater intersubject brain morphological variability.

| RESULTS
The first part of the output consists of both "population average" and "typical" Indian brain templates for five specific age-ranges: latechildhood (C1), adolescence (C2), young adulthood (C3), adulthood (C4) and late adulthood (C5) [see Table 1   increasing. Additionally, the variance decreases in the gray and white tissues with each stage. The contrast-to-noise ratio (CNR) between GM and WM improved through the successive stages in all the template age-groups (see Figure S1).  Warps to MNI were highly significantly greater (p < .05, corrected for N = 3 × 5 multiple comparisons) along the PA and IS axes in all cases.
Along the LR axes, differences were smaller but still significant at the same level for 4/5 cohorts (again, warps to MNI being larger); the C4 cohort showed no significant difference along the LR axis, but overall F I G U R E 4 Evaluation of the region-wise similarity of the MPM volumes as measured (left panel) by the relative volume ratio for each ROI via Equation (1), and (right panel) by mean deformation value (mDV) of each ROI; rows A-E show results for each age-specific group C1-C5, respectively. In the left-panel ROIs with notably different volume fractions are highlighted in purple (increases) and green (decreases), and in the right-panel ROIs with greater intersubject variability are shown as increasingly yellow F I G U R E 5 Validation cohort T1w results: (A-E) IBT-based results are in orange, and MNI-based results in blue. Wilcoxon's signed-ranks test was used to compare the distributions; p-values are shown at the top of each panel. For each validation group (V1-5), boxplots of the median warp magnitude along each major axis (LR, PA, IS) to a given template are shown in panel A-E. The warp distributions to MNI space are significantly larger along the AP and IS axes in all cases. While the differences tend to be smallest along the LR axis (particularly for C4), warps to MNI are nevertheless significantly larger for 4/5 of the cohorts along this axis, as well differences for this group were still large, due to the warps along the other axes.
Finally, we investigated the practical difference when using IBT versus MNI as a template space for fMRI processing, using the valida-

| DISCUSSION
We have introduced five new India brain template (IBTs) spaces, spanning an age range from 6-60 years. Additionally, corresponding atlases (IBTAs) from widely used segmentations were also created for each space. These should form useful reference templates and region maps for brain imaging studies involving predominantly Indian populations. Both the creation of age-specific templates and the inclusion of associated atlases make the present study distinct from previous Indian population brain template projects (Bhalerao et al., 2018;Pai et al., 2020;Rao et al., 2017;Sivaswamy et al., 2019); additionally, we have generated both "population mean" and highcontrast "typical" templates for each age band. The IBT volumes and corresponding atlases are publicly available for download, in standard NIFTI format, and freely usable by the wider neuroimaging community.
The need for age-specific templates in particular has been recognized across different populations (Fonov et al., 2011;Wilke, Schmithorst, & Holland, 2002;Yoon et al., 2009); however, Indian versions of age-specific brain templates have not been available to date. While adult brain templates may still provide reasonably accurate anatomical priors for normalizing lower resolution smoothed functional data, they may not be appropriate for high resolution structural and functional data (Wilke et al., 2002). For example, Yoon et al. (2009) examined the "template effect" in a pediatric population and noted significantly greater amount of deformation required for nonlinear normalization to the MNI152 adult template than compared to an age-appropriate template (2.2 vs. 1.7 mm). Further, the authors also noted significant differences in both volume-based and surfacebased morphological features between data warped to pediatric and adult brain templates. Such discrepancies are also reported in aging studies, where use of young-adult template (such as the MNI) for older adults can result in biases such as regional distortion and systematic over-expansion of older brains (Buckner et al., 2004).
Age-appropriate template for older adults have also been shown to provide more accurate tissue segmentation for structural imaging (Fillmore, Phillips-Meek, & Richards, 2015) and more focused activation patterns with improvement in sensitivity for fMRI group analyses (Huang et al., 2010).
In addition to age, consideration should also be given to the ethnic or population-specific differences (Lee et al., 2005;Rao et al., 2017;Tang et al., 2010), when choosing the appropriate brain template. As expected, there are noticeable structural differences when comparing the new IBTs with existing, popular standard templates (such as the MNI), which have been made from very different subject populations. Overall, registration to the IBTs from the Indian population validation groups required much less deformation of the input datasets and resulted in more accurate stereotactic standardization and anatomical localization. The relative differences in warping along the major axes of the brain were shown here using validation groups from the local population. The differences in warping magnitudes varied both by axis and by the age of subjects. Thus, the structural differences in templates are not trivial, that is, just scaling, but Such aspects were highlighted in the differences of outcomes in fMRI processing when using IBT versus MNI templates: the IBT-based output tended to have higher ReHo values among ROI pairs. The latter fact in particular suggests that the IBTs provided better functionto-anatomical alignment across groups, so that voxel with functionally similar time series tended to be grouped together more preferentially.
One might expect this to be a relatively small effect, because alignment to the MNI templates still appears generally reasonable; one would expect the overlap pattern differences to be occurring fractionally within ROIs and predominantly at boundaries. Indeed, the ReHo differences were relatively small, but with a noticeable trend toward higher values in the IBT-based datasets.
It is important to emphasize that these structural differences are only with regards to morphology; they do not relate to functional or behavioral outcomes, nor to intelligence, and so on. The purpose and goal of population-specific templates is for the practical consideration of maximizing the matching of structures across a group during an alignment step of processing, as well as to better match functional regions to structures. These are geometric and signal-to-noise considerations, which are important in brain studies (as demonstrated here), but which are unrelated to the brain behavior itself.
The wide variety of brain structural patterns in any group, even in an apparently homogeneous one, is also worth commenting on. This inherent variability affects both the creation and utilization of brain templates (Yang et al., 2020). In any population brain structures can vary to the degree of having different numbers of sulci in the same region (e.g., [Thompson, Schwartz, Lin, Khan, & Toga, 1996] and op cit); this is true even in a group of controls who are highly localized, genetically related, similar age and background, and so on. Thus, there is a minimum and nontrivial degree of variability in alignment that one can reasonably expect both when combining multiple subjects to generate a template, as well as in the overlap of anatomical structures when applying the template. Indeed, the Indian population (currently over 1.3 billion people) is spread across a wide range of geographies with diversity in linguistic-ethnic compositions as well as extensive genetic admixtures (Basu, Sarkar-Roy, & Majumder, 2016). In this study, the final mean template for each cohort contained variability. However, this was relatively low compared to the mean dataset values, and the final mean template contained a large amount of clearly defined structure. Moreover, the fractional overlap of ROIs when generating the maximum probability map atlases showed a high degree of agreement across the group through most of the brain.
The variability present in the template generation is also observable in the atlases. The intersubject variability (as measured by the mean deformation values for various regions during nonlinear registration to age and population-specific template) also correlated positively with the expansion of MPM volumes, in all age groups (see Figure S7).
While the final MPM atlases indicate the most frequent positions of each brain region in a given cohort, we also provide the probability density maps for each ROI in the atlas (see Figure S8 for example), which can be of additional use in ROI-based analyses.
While spatial normalization to IBT offers distinct advantages in terms of spatial accuracy and detection power, it may still be desirable to have the results from any particular analysis also reported in another space. For example, for comparisons with previously published studies, one might want to compare the locations of a finding with those reported in MNI, Talairach, or Korean template coordinate spaces.
Therefore, a nonlinear coordinate transformation mapping between IBT and the common MNI space has also been calculated, and a similar coordinate warp between any coordinate frames can be calculated easily.
There are several methodological strengths and limitations related to the current study that should be noted. We used combined stateof-the-art linear and nonlinear averaging techniques using AFNI's completely automated pipeline "make_template_dask.py," which uses the Dask python parallelization to efficiently make a template from a large group of subjects. We addressed several specific challenges involved in the template creation, such as intensity normalization from different scanners, scaling, resizing of the overall brain size to be representative of the cohort at each iteration, and anisotropic smoothing with preservation of edges. While the overall sample size of the study was relatively large, the late childhood and the late adulthood templates had relative modest sample sizes. Therefore, it will be of benefit for the constructed templates to continue to be updated with larger sample sizes as we collect more MRI datasets. Future work should also expand the templates for ages <6 yr and >60 yr. We will also expand this work to include development of a cortical surface atlas, which may allow for a registration procedure involving alignment of highly variable cortical folding patterns.

| CONCLUSIONS
In conclusion, the present work demonstrates the appropriateness of using age and population-specific templates as reference targets for spatial normalization of structural and functional neuroimaging data.
This database of age-specific IBTs and IBTAs is made freely available to the wider neuroimaging community of researchers and clinicians worldwide. We hope that these tools will facilitate research into neurological understand in general and into the functional and morphometric changes that occur over life-course in Indian population in particular.
02/2014-15). GS was supported by the Horizon 2020-funded ERC Advanced Grant "STRATIFY" (brain network-based stratification of reinforcement-related disorders; 695313), ERANID (understanding the interplay between cultural, biological and subjective factors in drug use pathways; PR-ST-0416-10004), BRIDGET (JPND brain imaging, cognition, dementia and next generation GEnomics; MR/N027558/1), the Human Brain Project (SGA 2, 785907, and SGA 3, 945539), the National Institute of Health (NIH) (R01DA049238, A decentralized macro and micro gene-by-environment interaction analysis of substance use behavior and its brain biomarkers Vivek Benegal contributed to the interpretation of the findings and edited the manuscript for important intellectual content. All authors discussed the results and contributed to the final manuscript.
ENDNOTE 1 FreeSurfer distinguishes between cortical parcellations and subcortical segmentations; here, we use "parcellation" generically to refer to final map of all ROIs.