The animals used to construct this atlas were 20² young to middle-aged male adults, chosen to have no common parents or grandparents, from the pedigreed Oregon National Primate Research Center breeding colony. The mean ± SD in age was 8.0 ± 1.6 years, distributed over a range from 6.2 to 10.5 years. For approximately 6 months prior to imaging, each monkey was individually housed in a stainless steel cage measuring 1.6 × 0.8 × 0.8 m (Allentown Caging, Allentown, NJ, USA) in a vivarium with a 12 h light/dark cycle (with lights on at 7 a.m.) that was maintained at 21 ± 1°C and 30–50% humidity. Each monkey had visual, auditory, and olfactory access to other monkeys in the vivarium, and limited physical access to a neighboring monkey. The monkeys were fed a diet of fresh fruit and 1 g banana-flavored pellets (consisting of 63% carbohydrate, 4% fat, and 22% protein PJ Noyes, Lancaster, NH, USA) in quantities sufficient to maintain a positive caloric intake.

All animal procedures were conducted in accordance with the “Guidelines of the Committee on the Care and Use of Laboratory Animal Resources” (National Health Council, Department of Health, Education, and Welfare, ISBN 0-309-05377-3, revised 1996). Prior to their implementation, procedures were reviewed by the Institutional Animal Care and Use Committee of the Oregon National Primate Research Center and were in compliance with all local, state, and national regulations pertaining to the humane use of animal subjects.

Magnetic resonance data were acquired using the same general data acquisition protocol, described in an earlier report (Flory et al., 2010). Following a 12-h fast, the monkeys were removed from their home cages under 10 mg/kg of ketamine anesthesia (Vedco, St. Joseph, MO, USA) and transported to the adjacent MRI facility (under 5 min transport time). At the MRI facility, an endotracheal tube was inserted and anesthesia was maintained by the inhalation of 1.5% isoflurane gas (Butler Animal Health Supply, Dublin, OH, USA) and oxygen (Polar Cryogenics Inc, Portland, OR, USA).

The monkeys were then placed inside a Siemens Trio whole-body 3T MRI system (Erlangen, Germany). For 12 animals, their heads were positioned in the center of a circularly-polarized extremity transmit/receive RF coil. For the remaining 8 animals, a circularly-polarized transmit, 8-element receive (InVivo, Orlando, FL, USA) RF coil was used instead.

After positioning the monkey in the scanner, a set of five to six T₁-weighted MP-RAGE images (TE = 4.38 ms, TR = 2500 ms, TI = 1100 ms) were obtained in 7 min 58 s acquisition time per image. For each image, the number of averages (NEX) was equal to 1, voxel sizes were 0.5 mm isotropic, and 128 slices were acquired. In-plane image sampling consisted of 128 and 96 data points in the readout and phase-encode directions, respectively, with no sub-sampling or acceleration factors. The signal-to-noise ratios of images obtained with the two RF coils, averaged over the brain, were within 10% of each other.

Images from all animals were inspected for quality and gross anatomical malformations that would render the images unsuitable for atlas generation. As a result of this survey, one animal was excluded from further processing due to abnormal posterior lateral ventricle anatomy (an image of this animal’s brain with the abnormality visible can, however, be seen in Figure 9 below). A total of 111 images acquired from the remaining 19 animals was thus incorporated into the INIA19 atlas.

A series of image processing operations was first applied to all acquired MR data to correct various types of artifacts, extract brain masks, and compute tissue probability maps. These stages of image preprocessing are illustrated for one animal in Figure 1 and each stage is described in more detail below.

Creation of the brain image template used the same reference-free unbiased groupwise non-rigid image registration approach (Learned-Miller, 2006; Balci et al., 2007) previously used to create the SRI24 atlas of the human brain (Rohlfing et al., 2010). Specifically, the following three stages of progressive alignment were performed: (1)

Initial translational alignment of all input images based on their centers of mass, using CMTK’s groupwise_init tool.

All three stages enforced a zero-sum constraint for each transformation parameter over all 19 individual images. The resulting average coordinate space represents a form of “minimum-deformation template” (Kochunov et al., 2001), i.e., one that requires the least amount of deformation on average to align it with a typical population of images.

Although the brain is clearly the focus of interest for the vast majority of imaging-based neuroscience studies, it is advantageous to have a whole-head template image also available. This facilitates the alignment of individual data sets from which the non-brain tissue has not been removed (i.e., images that have not been skull-stripped). In short, an atlas that provides both brain and whole-head template images can be used equally with and without skull-stripped data.

Creation of a template from unstripped images would be ill-advised, however, because the presence of non-brain tissue can interfere with image registration (Battaglini et al., 2008). Thus, by creating a template from unstripped images only would compromise the quality of the template in the brain areas. At the same time, creating whole-head and brain templates separately and compositing them would produce discontinuities between brain and non-brain regions in the combined whole-head image.

Thus, we applied a sequential approach to create both brain and whole-head templates, in this order, by adding a second non-rigid normalization stage after the non-rigid averaging of skull-stripped images. In this second stage, we simultaneously registered all whole-head images of the 19 input animals, but (a) initialized the registration transformation for each animal with the result of the brain-only registration, and (b) fixed the transformation in the brain region. As a result, the previously aligned brains remained in alignment, while the only coarsely aligned (via spatial proximity and smoothness of the registration) non-brain areas were selectively refined (see Figure 3).

To provide anatomically interpretable image coordinates, the final brain template was re-aligned from its original, essentially random position and orientation. To this end, the midsagittal plane was first computed using CMTK’s symplx tool as the plane of maximal symmetry (Citardi et al., 2004). A rigid coordinate transformation was created that aligned this plane with the mid-plane of the template image grid.

One of the authors (Christopher D. Kroenke) then manually identified the locations of the anterior commissure (AC) and posterior commissure (PC) landmarks, each defined as the center of the commissure at the midline, in the brain template image. A rotation transformation was determined that brought both landmarks into the same axial plane and also rotated the PC to a location exactly posterior from the AC. All atlas image files were then re-generated, with the symmetry and AC/PC alignment transformations concatenated onto the front of any other normalization transformations, to obtain final images with only a single interpolation from the original image data.

Finally, the image-to-physical space transformation matrices stored in all NIFTI image files of the different atlas images were adjusted such that the AC landmark became the origin (i.e., zero coordinate) of the image coordinate system. This adjustment should facilitate the use of the INIA19 template images in common neuroimaging software such as SPM (Friston et al., 2007)³.

The NeuroMaps rhesus macaque brain atlas⁴ consists of an MRI template image and a detailed cortical and sub-cortical label map for that image. Since this template itself has not been fully described in the literature, we first briefly summarize its creation below.

The MRI template is an image of the ex vivo brain of a single, 3-year-old male rhesus macaque obtained from the Russian primate center at Sochi, Russia. The brain was first perfused with papaverine and heparin to prevent clotting during perfusion with paraformaldehyde for fixation. The specimen was then immersed in 1 mM gadopentetate dimeglumine, Gd-DTPA (Magnevist), for several days to enhance gray-white contrast. For scanning, the specimen was suspended in cotton batting in a cylindrical canister, which was filled with Fomblin to minimize image noise in the surrounding medium and pressurized to prevent artifacts due to air bubbles. A high-resolution MR image with 0.15 mm³ isotropic spatial resolution was acquired on a 4.7T Bruker Biospec Avance scanner with a three-dimensional T₂-weighted gradient echo sequence.

The imaged brain was then made symmetric by flattening the brain’s mid-plane and reflecting one hemisphere with respect to the flattened mid-plane as described by Bowden et al. (2011) for an analogous mouse brain atlas creation procedure.

The cortical and sub-cortical label map was created manually by one of the authors (Mark F. Dubach), outlining structures on coronal image sections and initializing each subsequent slice by copying the regions outlined in the previous one. Discontinuities were then edited in orthogonal cross-sections. A total of 502, pixel-wise mutually exclusive regions were thus labeled in the map used herein.

The NeuroMaps label map, while arguably the most complete and accurate digital segmentation available, is still being actively improved⁵. Currently, confidence in the match of structural boundaries defined on the basis of the MRI to boundaries found in conventional atlases based on Nissl-stained sections varies for different parts of the brain. We have confidence in the current representation of the classical cortical structures (e.g., gyri and lobules), cortical white matter, major myelinated tracts, basal ganglia, ventricles, and cerebellum. The segmentations of the basal forebrain, hypothalamus, and amygdala are incomplete. The segmentations of the midbrain and hindbrain have not been fully reviewed. The incomplete internal segmentation of the thalamus is not considered reliable at this time, but this labeling will be replaced in the near future with boundaries mapped from the atlas by Ilinsky et al. (2002).

To transfer the label map to the newly created INIA19 template space, the NeuroMaps data were prepared as follows: (1)

To be able to distinguish left from right hemispheric regions in the label map image, the original NeuroMaps assignment of region names to label values was kept for the brain left hemisphere, whereas 1000 was added to each label value for the right hemisphere (but note that, for simplicity, Figure 4A uses the same colors to visualize the two hemispheres).

The region-averaged NeuroMaps MR image was then used for structural image alignment with the INIA19 MR template image and transfer of the label map to the INIA19 template space. It is now well-established, that the quality of label map propagation is greatly increased when using multiple source atlases (Rohlfing et al., 2004a,b; Rohlfing and Maurer, 2005), a process commonly referred to as multi-atlas segmentation and label fusion. Here, however, we have only one source atlas available, so we chose instead a method suggested by Heckemann et al. (2006) as “indirect fusion” and used by us also in creating the SRI24 atlas of the human brain (Rohlfing et al., 2010).

We used all 19 animals that were used to create the INIA19 template as intermediates, because their images were already aligned to the template space by construction. The image data set of each animal was also co-registered with the region-averaged NeuroMaps MR image via non-rigid registration using CMTK’s warpx tool. The NeuroMaps label map was then propagated 19 times, once per intermediate animal, into INIA19 space and all 19 propagated maps were finally combined using shape-based averaging (Rohlfing and Maurer, 2007) as implemented in CMTK’s sba tool.

Given the lower resolution of our in vivo MR images, and thus the INIA19 template image, relative to the NeuroMaps MR image and label map, 362 of the 502 manually outlined unilateral labels survived bilateral projection, mapping to INIA19 space, label map averaging, and down-sampling. As a result, there are 724 unique bilateral labels in the INIA19-mapped NeuroMaps parcellation map, one left/right pair for each of the surviving 362 regions.

All atlas files are distributed in NIFTI-1 data format⁶, stored in “RAS” index order⁷, with a grid size of 168 × 206 × 128 pixels and 0.5 mm isotropic pixel size. The template coordinate space is aligned with the image grid such that the midsagittal plane of the template anatomy coincides with the mid-plane of the image grid. The physical image coordinates, represented as direction vectors and offset in the NIFTI headers, were adjusted such that the zero physical space coordinate coincides with the AC landmark, and the PC landmark is exactly posterior from PC (see Figure 5 for illustration).

The distributed files are:

inia19-t1.nii: Whole-head T₁-weighted template image.

The structural channels of the INIA19 atlas (T₁-weighted whole-head and brain-only MRI, tissue probabilities, and ML tissue labels) are shown in Figure 6 for axial slices with 5 mm spacing (recall that the actual template images have 0.5 mm slice spacing).

The NeuroMaps labeling of the INIA19 brain template image is visualized in Figures 7 and 8. (A label description file containing label index, anatomical structure name, and RGB color information matching these renderings is provided with the atlas.)

After computing a coordinate transformation, via image registration, from an individual animal image to the atlas template image, data such as labels can be mapped back from the atlas into the individual image. For the INIA19 atlas, this includes specifically tissue labels and probabilities as well as NeuroMaps labels. In Figure 9, we demonstrate specifically the usefulness of the atlas tissue probabilities, mapped to each of our 20 originally imaged animals (including the one excluded from atlas construction), as priors for tissue segmentation.

After computing a coordinate transformation, again via image registration, from the atlas template image to an individual animal image, data from the individual can be mapped into atlas space. This process is commonly referred to as “spatial normalization” and serves the purpose of enabling the pixel- or region-wise comparison of properties across a population. This operation is central to the widely used group analysis technique Voxel-based Morphometry (VBM; Ashburner and Friston, 2000). Figure 10 shows, for illustrative purposes, that the INIA19 atlas is well-suited for spatial normalization using both affine and non-rigid registration methods.