Download Cortical Parcellations of the Macaque Monkey

Cerebral Cortex October 2012;22:2227– 2240
doi:10.1093/cercor/bhr290
Advance Access publication November 2, 2011
Cortical Parcellations of the Macaque Monkey Analyzed on Surface-Based Atlases
David C. Van Essen, Matthew F. Glasser, Donna L. Dierker and John Harwell
Department of Anatomy & Neurobiology, Washington University School of Medicine, St. Louis, MO 63110, USA
Address correspondence to David C. Van Essen. Email: [email protected].
Surface-based atlases provide a valuable way to analyze and
visualize the functional organization of cerebral cortex. Surfacebased registration (SBR) is a primary method for aligning
individual hemispheres to a surface-based atlas. We used
landmark-constrained SBR to register many published parcellation
schemes to the macaque F99 surface-based atlas. This enables
objective comparison of both similarities and differences across
parcellations. Cortical areas in the macaque vary in surface area
by more than 2 orders of magnitude. Based on a composite
parcellation derived from 3 major sources, the total number of
macaque neocortical and transitional cortical areas is estimated
to be about 130--140 in each hemisphere.
Keywords: architectonic, areas, maps, registration, retinotopy
Introduction
Primate cerebral cortex contains a complex mosaic of areas
that differ in architecture, connectivity, topographic organization, and/or functional characteristics (Felleman and Van Essen
1991; Kaas 2005). The macaque monkey is the most intensively
studied nonhuman primate, yet despite a century’s effort, an
accurate, consensus cortical parcellation is lacking for most of
macaque cortex. A primary reason is that differences between
neighboring areas are often subtle when assessed by any of the
available methods. Moreover, many individual areas show
internal heterogeneity in architecture, connectivity, and/or
functional characteristics. Another confound is that the size
and shape of each cortical area can vary 2-fold or more across
individuals (Van Essen et al. 1984; Maunsell and Van Essen
1987). Finally, the complexity and variability of cortical
convolutions are an impediment to visualizing results and
making accurate comparisons across individuals.
Many of the challenges posed by cortical convolutions and
their variability can be addressed using surface reconstructions
of individual hemispheres combined with surface-based registration (SBR) to an atlas (Van Essen and Dierker 2007).
Landmark-SBR (Van Essen et al. 1998, 2001, 2005; Van Essen
2005) is a specific type of SBR that uses explicitly delineated
landmark contours to constrain the registration from a source
(individual) to a target (atlas) surface. Here, we introduce the
landmark vector difference (LVD) algorithm for Landmark-SBR,
which achieves accurate alignment even when it entails large local
deformations. The LVD method is useful for registering individuals
to an atlas and for inter-atlas registration within a species or
between species (e.g., for macaque--human comparisons).
We applied this approach to an analysis of cortical parcellations in the macaque. Data from numerous published studies
were registered to the macaque F99 atlas (Van Essen 2002a; Van
Essen and Dierker 2007) using the LVD algorithm or older
Landmark-SBR methods. This facilitates identification of discrepancies as well as commonalities among different parcellation
The Author 2011. Published by Oxford University Press. All rights reserved.
For permissions, please e-mail: [email protected]
schemes. Informative examples illustrated herein include areas
of orbital and medial prefrontal cortex (OMPFC), the middle
temporal (MT) area and nearby areas in the superior temporal
sulcus (STS), and areas in medial and posterior parietal cortex.
Another objective of this study was to estimate the total
number of distinct cortical areas in the macaque. This number
represents a fundamental aspect of brain organization in any
mammalian species, yet has proven difficult to resolve even in
intensively studied species like the macaque. Published parcellations that span most or all the cortical sheet span a wide range
in the total number of areas. Here, we generated a composite
parcellation in which the areas identified in each region are
derived from a parcellation considered to be especially reliable in
that region. We also consider how to assess the ‘‘granularity’’ of
cortical parcellation in regions where the confounds of internal
heterogeneity and subtlety of areal boundaries are particularly
challenging (see Discussion). All atlas data sets are freely available
and are useful for a variety of analyses and comparisons within
and across species.
Materials and Methods
Surface-based visualization and analysis were carried out using Caret
software (v5.616 and earlier versions) (Van Essen et al. 2001; Caret,
http://brainvis.wustl.edu/wiki/index.php/Caret:About).
Macaque F99 Atlas
The macaque single-subject F99 atlas (Fig. 1A,C) includes cortical
midthickness (layer 4) surfaces for the left and right hemispheres,
derived from segmentation of a high-resolution (0.5 mm) T1-weighted
MR scan of an individual adult rhesus macaque (Van Essen 2002a,
2002b). The midthickness representation (top panels) is advantageous
because each unit of surface area is associated with a similar cortical
volume for both gyral and sulcal regions (Van Essen and Maunsell 1980;
Van Essen et al. 2001). Revised demarcations were made of the F99
atlas noncortical ‘‘medial wall’’ boundaries (black in Fig. 1C) with the
aid of published atlases (Paxinos and Franklin 2000; Saleem and
Logothetis 2007). The analysis of cortical surface area included all of
neocortex plus several transitional regions (subicular complex) and
pyriform cortex. The hippocampus proper (CA1-4) was excluded
because it was not accurately segmented.
To bring the left and right hemisphere surfaces into correspondence,
26 landmark contours were drawn in corresponding gyral or sulcal
locations on the F99 native-mesh surfaces. These were projected to the
‘‘spherical standard’’ (distortion-reduced and geographically aligned)
surfaces of each hemisphere. The right and mirror-flipped left
hemisphere spherical landmarks were averaged and used as a target
for registration of each hemisphere using the landmark pin and relax
(LPR) algorithm (Van Essen 2005). This yielded standard-mesh surface
representations (Saad et al. 2004) containing 73 730 nodes that are
identified as being on the ‘‘74k_f99’’ mesh, reflecting both the number
of surface nodes and the F99 target atlas. The left and right
hemispheres are in good geographic correspondence on the 74k_f99
mesh, as assessed by visual inspection. More specifically, highlighting
any given node in one hemisphere identifies a node in the other
hemisphere that is very similar in its position relative to nearby gyral
and sulcal landmarks, usually within 1--3 mm. This assessment was
Figure 1. Landmarks for SBR displayed on the F99 atlas and an individual subject right hemisphere. (A) Lateral views of the F99 atlas midthickness (top), inflated (middle), and
spherical (bottom) surfaces. Fifteen landmark contours are projected to the surface. (B) Corresponding lateral views of an individual macaque (Case CH from Kolster et al. 2009)
generated using the FreeSurfer processing pipeline followed by a semiautomated pipeline in Caret. After conversion to Caret format, the individual-subject surface was inflated,
mapped to a sphere, processed by multiresolution morphing to reduce distortions and rotated to a spherical standard orientation. A standard set of template landmark contours
(originally generated on a different hemisphere and stored as contour points relative to that hemisphere’s spherical standard surface) was projected to the spherical standard
surface and viewed on other configurations (pre-edit landmarks on the inflated surface). Manual editing yielded post-edit contours shown in the other panels. (C) Medial views of
the atlas surface. Revised medial wall borders, compared with earlier version of the F99 atlas (Van Essen and Dierker 2007) include the boundary between neocortex or
transitional cortex and basal forebrain, piriform cortex, and other non-neocortical structures. (D) Medial views of the individual surface.
supported by computing a difference map between cortical folding
(mean curvature) for the left versus the right hemisphere. This
difference map confirmed that creases of high folding are generally in
good agreement (within 2 mm). In a few locations, the mismatch
between folding maps was larger, but these were places where genuine
hemispheric differences in local folding patterns were evident.
Versions of the left and right F99 midthickness surfaces were
generated whose dimensions and surface area matched that of an
average adult macaque. First, the F99 midthickness surfaces were
registered to the ‘‘D99’’ atlas volume of Saleem and Logothetis (2007)
using an F99-to-D99 affine transformation matrix generated by McLaren
et al. (2009). In generating their 112RM-SL population-average adult
rhesus MR volume, McLaren et al. (2009) registered each of 112
individual-subject MR scan to the D99 atlas volume by an affine
transformation (after an initial rigid-body alignment). The scale factors
used for the affine transformations were averaged separately for the 52
male and 32 female brains and for the x, y, and z dimensions. Applying
the inverse of these scale factors to the F99 surfaces in D99 space
yielded F99 atlas surfaces that represent actual rhesus macaque brain
size averaged across sexes.
FreeSurfer-to-F99 Registration Pipeline
A semiautomated process was implemented to convert FreeSurfergenerated macaque surfaces into Caret file formats and to register the
individual hemispheres to the F99 atlas.
The FreeSurfer cortical segmentation process generates pial and white
(gray-white boundary) surface representations that can be averaged
to form a midthickness surface, shown for an exemplar individual
hemisphere (Fig. 1B,D). Comparison of the F99 atlas (Fig. 1A,C) and the
individual (Fig. 1B,D) midthickness surfaces (top row) reveals many
shape features in common but numerous differences in detail. A set of
15 template landmark contours is shown on the midthickness (top) and
inflated (middle) and spherical (bottom) configurations of the atlas
surface (Fig. 1A,C). Corresponding template landmarks were projected
to the individual subject’s surface (‘‘preedit’’ in Fig. 1B). These template
landmarks were edited manually, yielding ‘‘postedit’’ landmark contours
that represent corresponding geographic (gyral and sulcal) features
in the source (individual) and target (atlas) surfaces. These landmarks
2228 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
were used to constrain registration of the individual sphere to the atlas
sphere using the LVD algorithm described in the next section. An ‘‘FSto-F99’’ tutorial document and associated data set illustrating this process
is available (http://sumsdb.wustl.edu/sums/directory.do?id=5255961).
LVD Registration
The LVD method of Landmark-SBR is described in detail because of its
broad utility (see Introduction) and because understanding the process
can aid in cases that require customized adjustment of registration
parameters. Although similar in overarching objective to the diffeomorphic landmark-based spherical registration algorithm of Glaunes
et al. (2004) and the elastic energy algorithm of Joshi et al. (2007) and
Pantazis et al. (2010), it is very different in its algorithmic approach (see
Discussion).
The LVD algorithm deforms a source (individual) spherical mesh
until its landmarks match the target (atlas) landmarks, while minimizing
local distortions in regions between neighboring landmarks. Alignment
is achieved in an interactive coarse-to-fine sequence involving multiple
‘‘stages’’ and multiple ‘‘cycles’’ within each stage. In each cycle, the
vector difference between the current location of source and
corresponding target landmark points provides the driving force for
the deformation. The vector difference is distributed across nearby
surface nodes by spatial smoothing of the initial deformation field. Key
steps in the process are illustrated in Figure 2 (same individual and atlas
surfaces as in Fig. 1); see also http://brainvis.wustl.edu/wiki/index.php/
Caret:Operations/SurfaceBasedRegistration.
An initial registration step is to resample landmark contours so that
corresponding source and target landmarks each contain the same
number of points (Fig. 2A,B). Each source landmark point is then
intercalated into the nearest tile of the standard-mesh source sphere.
This yields a landmark-intercalated source tessellation whose topology
reflects the newly incorporated nodes. This is illustrated in Figure 2C
for a small surface patch in the calcarine sulcus (red nodes are the
intercalated landmark points). Similarly, a landmark-intercalated target
tessellation is generated by intercalating each target landmark point
into the nearest tile of the target sphere (Fig. 2D).
At the outset of each cycle, the differences in 3D coordinates of
corresponding landmark nodes in the source and target spheres are
Figure 2. Key stages in LVD registration, using an exemplar individual (source) hemisphere registered to the F99 atlas target. (A, B) Landmark contours on source (case CH of
Kolster et al. 2009) and target (F99 atlas) spheres after resampling to match the number of contour points in corresponding source and target landmarks. (C) Expanded views of
surface mesh in the calcarine sulcus, showing landmark nodes (red) intercalated into the 74 k node spherical mesh. The surface was extensively smoothed and projected back to
a sphere, which caused compression of tiles in the vicinity of the 12 nodes that have only 5 (instead of 6) neighbors. This was important in order to avoid smoothing-induced
distortions at intermediate stages of the deformation process. (D) Expanded views of target nodes intercalated into an equivalent mesh at the locations of the resampled target
landmarks. Red arrows in C and D indicate displacement vectors computed for the highlighted landmark nodes (black). (E) Y-component of the smoothed displacement field.
Yellow--orange signifies nodes to be displaced to the left (positive y), blue signifies nodes to be displaced to the right. (F) Deformed surface mesh at the end of stage 1, which
includes 2 cycles of displacement, smoothing, and morphing, resulting in displacement of the mesh about one-third of the distance to the target. To transfer these results to the
native-mesh surface in preparation for the next stage, each node from the native-mesh source sphere is projected to the initial landmark-intercalated source sphere (panel C) then
unprojected based on its location in the deformed landmark-intercalated source sphere (panel F). (The projection and unprojection steps use barycentric coordinates to specify the
position of nodes in one mesh relative to the nearest tile of the other mesh—see Fig. 3E). This deformed native-mesh sphere becomes the input source sphere for the next stage
of the LVD process. (G) A fresh surface mesh at the beginning of stage 2. Landmark nodes preserve their location from the end of stage 1 (panel F) but are intercalated into an
undistorted spherical mesh (as in panel C but with shifted location of landmark nodes). (H) Alignment of source calcarine sulcus landmarks after 5 stages of LVD registration. The
similarity of landmark node positions in panels D and H signifies successful registration.
computed (red arrows for highlighted nodes in Fig. 2B) and assigned as
vector x, y, and z components for each intercalated landmark node.
This defines a displacement field on the source surface in which only
the intercalated landmark nodes have nonzero values; the displacement
field is stored as a Caret ‘‘surface-shape’’ file with separate fields for the
x, y, and z components. This sharply peaked displacement field is
smoothed by repeatedly averaging each node’s vector component value
with that of its neighbors. The number of smoothing steps is specified
separately for each cycle and stage; it generally entails many iterations
for early coarse registration stages and fewer iterations for the later
fine-tuning stages. In panels C--E, the smoothed displacement field for
the y-axis (horizontal) component (shown in Fig. 2E) diverges, pointing
leftward (orange) on the left and rightward (blue) on the right; there is
also a large upward component (z-axis) (not shown). Surface nodes are
displaced in proportion to the magnitude and direction of this
displacement field. The mesh is then projected back to a spherical
configuration, smoothed, and subjected to spherical morphing (Drury
et al. 1996) to reduce local distortions, while holding the landmark
nodes in their displaced locations. After 2 cycles of displacement and
landmark-constrained smoothing and morphing, source landmark
nodes are shifted substantially toward the target nodes (i.e., arrows in
2F are ~30% shorter than in 2C). To bring the landmark nodes
progressively closer to their targets without creating unacceptably
large local distortions, additional stages are carried out in which the
deformed landmark points are converted back to contours (strings of
contour points) and intercalated from their displaced positions into
a fresh (undistorted) standard-mesh sphere (Fig. 2G). In this exemplar
case, 5 stages of displacement, smoothing, and morphing (Fig. 2C--F),
brought the deformed source landmark nodes (Fig. 2H) acceptably
close to their desired target locations (Fig. 2D).
The deformed native-mesh source sphere from the final stage is used
to create a ‘‘deformation map’’ that encodes the precise relationship
between source and target meshes. This is illustrated in Figure 3 using
low-resolution views (top row) and a zoomed surface patch on the
prelunate gyrus (bottom row). A highlighted node on the ‘‘native’’
midthickness surface mesh (Fig. 3A) is also highlighted in the
‘‘preregistration’’ spherical surface (Fig. 3B) and on the deformed
‘‘postregistration’’ native-mesh sphere (Fig. 3C), with landmarks now in
register with the atlas target landmarks (Fig. 3D). This establishes
a precise spatial relationship between each node on the atlas sphere
and a surface tile in the deformed native mesh sphere as shown in
a highly expanded patch in Figure 3E. The deformation map file
encodes these relationships using barycentric coordinates to describe
the position of each target node within the nearest tile of the deformed
source spherical mesh. In the resampled midthickness surface, each
atlas node has an appropriately interpolated coordinate location based
on the coordinates of the associated tile in the native-mesh midthickness surface. The resampled surface preserves the shape of the nativemesh midthickness surface but has a more regular tessellation pattern
(lower panel) and is slightly smoother as a result of the interpolation
process (Fig. 3F). Thus, the deformation map encodes both the 2D
spatial displacement between source and target across the surface, and
the resampling that allows native mesh surfaces with disparate numbers
of vertices to be brought onto a standard mesh.
The ‘‘source-to-target’’ individual-to-atlas deformation map enables
rapid conversion of data from the individual to the 74k_f99 atlas mesh.
For example, the map of cortical folding (top, Fig. 3F) appears as
a slightly smoothed version of the native-mesh folding map (Fig. 3A). In
the reverse direction, a target-to-source deformation map specifies the
location of each target node in relation to the 3 nodes of the deformed
source surface tile it overlies. This allows any data of interest on the
Cerebral Cortex October 2012, V 22 N 10 2229
Figure 3. Individual midthickness surface represented using native mesh and atlas mesh tessellations. (A) Native-mesh midthickness surface (top) and expanded to reveal the
native mesh with one node highlighted (bottom panel). (B) The native-mesh spherical surface with landmarks, centered on the highlighted node (top panel) and expanded to show
the preregistration spherical mesh (bottom panel). (C) The deformed native-mesh sphere, with landmarks precisely aligned to those of with the target spherical. (D) Target (atlas)
spherical landmarks. (E) Target spherical nodes overlaid on deformed source spherical mesh. A source-to-target deformation map specifies the barycentric coordinates of each
target node (e.g., red node) relative to the nearest tile in the deformed native-mesh sphere (e.g., blue nodes). (F) A resampled (74k_f99) individual subject midthickness surface is
generated by positioning each target node based on its barycentric coordinates relative to the native-mesh midthickness surface.
atlas to be mapped to the individual’s surface mesh. These maps are
‘‘quasi-invertible,’’ in the sense that transformation of data from one
mesh to another, then back to the first preserves the original
representation except for degradation relating to differences in the
mesh density (e.g., a coarser representation arises if data are mapped
from a fine mesh to a coarse mesh and back) and minimal smoothing.
Resampling of data can be done using either the source node nearest to
the target node or a weighted average of the source nodes defining the
tile within which the target node lies.
Assessing Geographic Correspondences and Distortions
The LVD Landmark-SBR process establishes node-to-node correspondences between the atlas surface and the resampled (74k_f99) version of
the individual surface. The fidelity of registration can be assessed by
visual inspection of highlighted nodes in relation to nearby geographic
features. For example, Figure 4A shows 10 highlighted nodes (black
squares) on the individual (top) and atlas (bottom) midthickness
surfaces. In general, the node locations are in good correspondence
(within ~1--2 mm) relative to nearby gyral and sulcal features. However,
imperfect correspondences occur in places. For example, node 8,
which is near the tip of the inferior occipital sulcus (IOS) in the atlas
surface but is ventral to the IOS tip by about 4 mm in the individual
surface. This is attributable to differences in the trajectories of the IOS
in these 2 hemispheres.
Several types of spatial distortion impact the quality of the registration
process, Figure 4B shows a distortion map based on the ratio of surface
areas for each standard-mesh atlas tile and the corresponding standardmesh individual tile, each measured on the midthickness surfaces. The
distortions are locally quite variable but are moderate in magnitude
overall (only 5% of tiles having distortion more than 2-fold of the mean).
This overall distortion arises from 2 main sources. One occurs in the
mapping of 3D midthickness (or other anatomical) surfaces to a sphere.
Local distortions are concentrated in regions associated with intrinsic
(Gaussian) curvature, that is, concavities and convexities that are not
simple folds or creases (Van Essen and Drury 1997; Van Essen 2005).
The sphere versus midthickness distortion maps (Fig. 4C) show broad
regional similarities for the individual and atlas surfaces, including
compression of the frontal and occipital poles (dark blue regions) and
2230 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
expansion of the Sylvian fissure (red regions) that are characteristic of
both hemispheres. However, many local bumps and dents are
idiosyncratic to each hemisphere, so the maps differ in many local
details.
Another source of distortion arises from the local deformations that
occur during spherical registration. Figure 4D shows an areal distortion
map for pre- versus post-registration spheres, with the landmark
contours overlaid. The distortions are generally modest in magnitude
(99.7% within 2-fold of the average) and are concentrated near
landmark terminations. One reason is that the correspondences near
the terminations of sulci and gyri are the most difficult to estimate (see
Fig. 1). Also, the LVD registration algorithm does not completely
eliminate distortions in the region between adjacent landmarks.
Landmarks translated over a large distance tend to have residual
compression in front of the direction of translation and residual
expansion in the region immediately behind.
LVD Parameter Specification
The LVD algorithm can achieve accurate registration of landmarks
(within 1 mm) even when large local deformations are required, as long
as an adequate number of stages with appropriate parameters at each
stage are used to compensate for the initial misalignment. Numerous
free parameters can be adjusted to achieve the desired degree of
alignment and to reduce local distortions. A standard set of parameters
suitable for registering individual macaque hemispheres is available and
can be read in automatically (see SumsDB data set associated with Fig. 2
and the FS-to-F99 tutorial). Any topological defects (crossovers)
occurring at any stage of a deformation are reported and can be used
as a guide in adjusting parameters, the surfaces generated at
intermediate stages are also available. The parameters used successfully
for any one registration can be read in and used for other registrations
that have similar requirements.
Registration quality is impacted by the number and distribution of
landmarks, based on the number of local features considered reliable,
the shape differences between the source and target surfaces and the
desired degree of alignment in different regions. Computation time
depends heavily on the number of landmark contours, the number of
Figure 5. Preparatory steps for Landmark-SBR of a partial hemisphere reconstruction. (A) Anterior view of 3D anatomical surface of macaque case om43,
with landmarks drawn along gyral and sulcal landmarks and along perimeter of
prefrontal cortex. (B) Inflated surface generated using a combination of smoothing
iterations and expansion from the center of gravity. A cut along the principal sulcus
(gap at top) reduced distortions in the cortical flat map (not shown). (C) A partial
sphere generated by 1) projecting all nodes in the inflated surface to a fixed radius, 2)
reducing distortions relative to the 3D anatomical surface using spherical morphing, 3)
applying a spherical compression step that shifts all nodes in the front portion to
subtend a solid angle approximately matching that of the target atlas landmarks, and
4) rotating the compressed sphere into approximate alignment with the atlas. (D)
Corresponding landmarks on the F99 atlas anatomical surface. (E) Target landmarks
on the F99 atlas sphere.
Figure 4. Geographic correspondences and areal distortion maps for different stages
of spherical registration using LVD Landmark-SBR. (A) Highlighted nodes (black
squares) indicated corresponding geographic locations in the individual (top) and atlas
(bottom) midthickness surfaces, each shaded by is own map of cortical folding. (B)
Ratio between the surface area of each atlas midthickness tile and the corresponding
tile in the registered and resampled individual surface. (C) Distortion map between
individual sphere and its 3D midthickness surface (top) and between F99 atlas sphere
and its 3D midthickness surface (bottom). The individual sphere versus midthickness
distortion map is based on the Caret-generated (multiresolution morphing) sphere,
whose distortions differ from the FreeSurfer process. (D) Distortion map and between
atlas and individual spheres, showing the local distortions needed to align source and
target landmarks.
processing stages and cycles per stage, and the density of the surface
mesh used for spherical registration. For the cases used in this study, it
typically took about an hour on a MacBook Pro (2.53 GHz Intel CPU, 8
GB RAM).
Surfaces Reconstructed from Contours
Some architectonic parcellations had been mapped to surfaces
reconstructed from contours of histological sections. The PHT (Paxinos
and Franklin 2000) and MMF11 (Markov et al. 2011) atlases were
available as complete hemisphere reconstructions based on contours
drawn along the estimated trajectory of layer 4 in images of each
section loaded into Caret. Landmark-SBR registration from these atlas
surfaces to the F99 atlas was similar to that described above for
Freesurfer-to-F99, except that a larger number of landmarks were used
in order to better compensate for the local irregularities in the
reconstructed surfaces. The current version of the PHT00 atlas
supplants an older version that had been registered using the LPR
algorithm and fewer landmarks.
Customized preprocessing was required for cortical surfaces that
included only a partial hemisphere (e.g., only prefrontal cortex). Figure 5
illustrates preprocessing steps used to prepare a partial hemisphere of
macaque prefrontal cortex (Ferry et al. 2000) for Landmark-SBR.
Landmark contours were delineated along geographically corresponding
locations in the individual (Fig. 5A) and atlas (Fig. 5D) surfaces. This
example included 15 gyral and sulcal landmark contours, of which 9 ran
along the perimeter of the partial-hemisphere surface. The individual
surface (Fig. 5A) was inflated (Fig. 5B) and mapped to a partial sphere
(Fig. 5C) whose landmarks were rotated and scaled to approximately
align with the corresponding landmarks on the target sphere (Fig. 5E).
Registration of Cortical Flat Maps
Many parcellations were registered initially to the 79-0 atlas (Drury
et al. 1996) and then to the F99 atlas using an older landmarkconstrained diffeomorphic (fluid-based) algorithm applicable to 2D flat
maps—the landmark flat-fluid (LFF) algorithm (Joshi 1997; Van Essen
et al. 2001). The input parcellations were represented on computerized
surface reconstructions or on manually drawn flat maps (see Table 2).
Data Availability
Data sets associated with this study are freely available in the SumsDB
database (http://sumsdb.wustl.edu/sums/directory.do?id=8286148&dir_
name=MACAQUE_ATLAS_CC11). Associated with each figure in this
study is a ‘‘scene’’ file that allows regeneration of all the constituent
figure panels. To facilitate cross-platform data migration, most surfacerelated files are in GIFTI format (http://www.nitrc.org/projects/gifti).
Data sets can be viewed online using WebCaret without downloading
data or software and can be downloaded for offline analysis using Caret
or other GIFTI-compliant software.
Results
Macaque Cortical Surface Area and Variability
Table 1 shows data on cortical surface area and its variability in
the macaque. The macaque F99 atlas has an average surface
Cerebral Cortex October 2012, V 22 N 10 2231
area of 122 cm2 per hemisphere in its native dimensions. After
normalization to the average dimensions for adult rhesus
macaques (male and female combined; see Materials and
Methods), surface area averages 105 cm2 per hemisphere.
Sex differences and individual variability in macaque brain
dimensions were estimated using scale factors that relate each
of 112 individual brains to the 112RM-SL population-average
atlas volume (see Materials and Methods). Average brain size in
males exceeds that in females by 17% in volume (11% in surface
area). Estimated brain volume was not correlated for with age
in either males (r2 = 0.009, range = 3--36 years) or females
(r2 = 0.05, range = 4--29 years).
similarities and differences among parcellations at any given
location are difficult to resolve in this figure. When viewing the
same data sets in Caret, the assignments at any location can be
immediately ascertained by clicking on a surface node of
interest and viewing a list of areas in the ‘‘Identify node’’
window (Fig. 6L,M). For example, the white highlighted node in
each panel (white arrow) has relatively consistent assignments,
listed in Figure 6L: it is identified as MT in 6 parcellations, and
the other assignments (areas 19, OAa, and TA) reflect older
parcellations in which area MT had not been identified. In
contrast, the nearby black highlighted node (black arrow) has
a distinct areal identity in all 6 parcellations where an
assignment was made (Fig. 6M: area PA, TP, 22, STP, PBc, or TA).
In places where parcellations disagree, a few involve
differences are only in name, that is, different names used for
what is fundamentally the same well-defined area (e.g., areas
V1, 17, and OC are synonymous). At the other extreme, areal
boundaries in one parcellation can show minimal correlation
with those in another parcellation. Intermediate situations are
by far the most common, in which areal boundaries are
correlated but imperfectly on the atlas map. Lack of correlation
can arise from several factors, including 1) differences in the
size (2-fold or more), shape, or location of areas in the
contributing hemispheres (biological variability); 2) inaccuracies in delineating areal boundaries owing to methodological
bias or error; or 3) inaccuracies or bias in the registration
process. Another major factor relates to the distinction
between ‘‘lumping’’ and ‘‘splitting.’’ For example, one parcellation may identify a single area (e.g., area 3) in a region where
another parcellation identifies 2 (e.g., areas 3a and 3b),
resulting in areal boundaries that are strongly correlated along
part but not all of their trajectories.
The following sections discuss several cortical regions where
cross-parcellation comparisons benefit from having data registered to the F99 atlas and from having access to individual
hemisphere surfaces that contributed to the atlas. These regions
include orbital and OMPFC; the STS, especially area MT; and
parietooccipital cortex. We also illustrate a composite map based
on multiple parcellations spanning most of cerebral neocortex.
Macaque Parcellations
Table 2 provides information on 15 published cortical
parcellations mapped to the F99 atlas, including 5 newly
incorporated or revised maps, plus 9 previously registered
parcellations (Van Essen 2004). The parcellations are based on
architectonic studies, retinotopy using neurophysiology or
functional magnetic resonance imaging (fMRI), connectivity, or
a combination of the above.
Figure 6 displays all 15 parcellations on the very inflated atlas
surface. Six parcellations (Fig. 6A--F) each encompass most or
all of cerebral neocortex and transitional cortex. The others
cover only a minority of cortex; panels H and I each combine 3
nonoverlapping parcellations. For the 6 pan-hemispheric
parcellations, the number of identified areas ranges from 26
(BB47) to 161 (PHT00) (Table 2, column 4). Specific
Table 1
Macaque neocortical surface area
Macaque F99 atlas
Surface area (cm2)
Left
Right
Average
Individuala
Normalized
Sex
Scaling factorsb
125
119
122
107
103
105
Male (n 5 52)
Female (n 5 30)
Combined
1.04 ± 0.12 (0.77--1.37)
1.22 ± 0.13 (1.03--1.65)
1.11
a
Surface areas of the original mesh (~30 000 nodes each for the left and right hemisphere) were
about 3% larger than the standard-mesh values (because resampling slightly smooths the
surface) but were not a discernibly better fit to cortical shape when overlaid on the MR volume.
b
Average 3D scaling factors for individuals relative to the 112RM-SL average MR, based on the
product of the x, y, and z scaling factors.
Orbital and Medial Prefrontal Cortex
The F99 atlas includes 8 parcellations having cortical areas in
the OMPFC region. Figure 7 compares 2 of the most widely
Table 2
Macaque cortical parcellations registered to the F99 atlas (in the order illustrated in Fig. 6)
Study
Abbreviation
Lobe(s)
Number of areas
Number of cases
Modality
Input (source) surface
Registration algorithm
Lewis and Van Essen (2000)
Felleman and Van Essen (1991)
Paxinos and Franklin (2000)
Markov et al. (2011)
Brodmann (1905)
von Bonin and Bailey (1947)
Ungerleider and Desimone (1986)
Seltzer and Pandya (1978, 1980, 1986)
Galletti et al. (1999)
Preuss and Goldman-Rakic (1991)
Preuss and Goldman-Rakic (1991)
Lyon and Kaas (2002)
Bayliss et al. (1987)
Kolster et al. (2009)
Ferry et al. (2000)
LV00
FV91
PHT00
MMF11
B05
BB47
UD86
SP78-80-86
GFG99
PGR91a
PGR91b
LK02
BRL87
KMA09
FOA00
All
All
All
All
All
All
O, T, P
P, T
P
P
F
O
T
O, P, F
F
88
80
161
78
28
26
17
13
1
5
38
11
9
12
20
5
1
1
1
1
1
1
1
1
1
1
1
1
4
5
Architectonic
Multiple
Architectonic
Architectonic
Architectonic
Architectonic
Retinotopic Physiology
Architectonic
Retinotopic physiology
Architectonic
Architectonic
Connectivity
Architectonic
Retinotopic fMRI
Architectonic
Caret flat map
Manual flat map
3D reconstruction
3D surface
Manual flat map
Manual flat map
Manual flat map
Manual flat map
3D partial hemisphere
Manual flat map
Manual flat map
Manual flat map
Manual flat map
FreeSurfer surface
3D partial hemisphere
LFF to F99
LFF
LVD (30)
LVD (24)
LFF via 79-0
LFF via 79-0
LFF via 79-0
LFF
LVD (9)
LFF via 79-0
LFF via 79-0
LFF via 79-0
LFF via 79-0
LVD (16)
LVD-ph (15)
Note: LVD; the number of landmarks is indicated in parentheses LVD-ph: partial-hemisphere, registered via LVD LFF (Joshi 1997; Van Essen et al. 2001) via 79-0 signifies initial registration to the Case 79-0 atlas
(Drury et al. 1996), followed by registration to F99.
2232 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
Figure 6. Cortical areas from 15 parcellations mapped to the F99 atlas surface. (A--K) Areas for one or more nonoverlapping parcellations, presented in the same sequence as in Table 1.
(L, M) Areas identified for the white (panel L) and black (panel M) nodes highlighted in panels A--J. Each area is identified using a standard format: \area-name[_\parcellation[.
used, the FOA00 parcellation of Ferry et al. (2000) and the
PHT00 parcellation of Paxinos and Franklin (2000). For
the FOA00 parcellation, OMPFC areas were delineated histologically in 5 hemispheres using criteria based on cytoarchitecture, myeloarchitecture, and 3 types of chemoarchitecture
(Carmichael and Price 1994). Areas were mapped to surface
reconstructions of prefrontal cortex in each hemisphere
(Fig. 7A) and then registered to the F99 atlas (Fig. 7B; see
Materials and Methods). A probabilistic map (Fig. 7C, n = 5)
shows that intersubject alignment was reasonably good for all 3
Cerebral Cortex October 2012, V 22 N 10 2233
Figure 7. Comparing OMPFC areas in the FOA00 and PHT00 parcellations. (A) Ventral view of orbital cortex an individual subject (case om43 of Ferry et al. 2000). (B) Case om43 areas
registered to the F99 atlas. (C) Probabilistic areas 13a, Iai, and 11l (n 5 5 hemispheres). (D) Composite map of FOA00 areas on ventral (top) and medial (bottom) views. (E) FOA00
composite borders overlaid on PHT00 areas registered to F99 atlas. (F) PHT00 areas on ventral and medial views of the reconstructed PHT00 atlas surface (Paxinos and Franklin 2000).
illustrated areas (13a, Iai, and 11l). Composite boundaries
drawn using the complete probabilistic map show the most
likely location and extent of each area on ventral and medial
atlas views (Fig. 7D).
The PHT00 parcellation was based on cytoarchitecture and
chemoarchitecture (cholinesterase) of a single atlas hemisphere
reconstructed from section contours (Fig. 7F; see Materials and
Methods). Side-by-side comparison of the FOA00 and PHT00
parcellations (Fig. 7D,F) shows many similarities but also many
differences. Figure 7E allows a more direct comparison by
showing the PHT00 areas registered to the F99 atlas, with the
FOA00 borders overlaid. A direct comparison reveals that some
areas are reasonably well aligned (e.g., PHT00 13M, 13L, and 11L
vs. FOA00 13m, 13l, and 11l), whereas other areas (e.g., 25) are
in substantial disagreement; most areas lie in between (some
overlap but extensive nonoverlap). Inspection of both parcellations on the original hemispheres (Fig. 7A,F) indicate that the
mismatches generally cannot be attributed to inaccurate
registration to the atlas for either parcellation.
Area MT and the Posterior Bank of the STS
Area MT is one of the most intensively studied extrastriate area
in the macaque, yet its location and extent vary markedly in the
7 parcellations mapped to the F99 atlas. Macaque MT was
originally defined using 4 independent criteria that revealed
2234 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
coextensive boundaries: it is heavily myelinated, contains
a complete visuotopic map, has a high incidence of directionselective cells, and receives a strong projection from area V1
originating in layer 4B (Van Essen et al. 1981; Maunsell and Van
Essen 1983). Using these criteria (especially myeloarchitecture), MT was mapped to the medial half of the posterior bank
of the STS, in a region typically ~3--4 mm wide, ~1-cm long, and
with an average surface area of 33 mm2 (Van Essen et al. 1981;
Maunsell and Van Essen 1987). In Figure 8A, the white squares
outline the posterior bank of the STS, and red shows the extent
of myeloarchitectonic MT in the FV91 parcellation, in good
agreement with the aforementioned location and extent. (It is
more elongated in the figure than in the actual atlas midthickness surface because the inflation process compresses the
surface anisotropically in this region.)
Figure 8B shows MT and the other FV91 areas displayed on
the very inflated atlas surface, with white squares again
outlining the posterior bank of the STS. Figure 8C--H shows
the diversity in position, shape, and size of area MT (red) in 6
other studies mapped to the atlas, presented in order of
decreasing similarity to MT of FV91 (black contour in each
panel). The LV00 (Lewis and Van Essen) MT map (Fig. 8C),
based mainly on myeloarchitecture, is centered on the same
location but is narrower and longer than FV91 MT (black
contour). The UD86 (Ungerleider and Desimone 1986) MT
map (Fig. 8D) is similar in size and shape to FV91 MT but is
Figure 8. Comparison of macaque area MT size and location across multiple studies. (A) Area MT (red) from FV91 (Felleman and Van Essen 1991), based on myeloarchitectonic
MT on inflated F99 atlas, with posterior bank of the STS highlighted (white squares). MT is several-fold longer than wide when visualized in 3D space (see Fig. 5, Van Essen et al.
1981) or on the midthickness surface, but it appears more highly elongated in Figure 8 owing to anisotropies that occur during surface inflation. (B) FV91 visual areas with black
contour outlining MT on the very inflated atlas surface. (C--H) Cortical areas from 6 other parcellation schemes, with MT as read and FV91 MT outlined (black contour). (C) LV00
5 Lewis and Van Essen (2000). (D) UD86 5 Ungerleider and Desimone (1986). (E) LK02 5 Lyon and Kaas (2002). (F) PHT00—Paxinos and Franklin (2000). (G) MMF11 5
Markov et al. (2011). (H) Probabilistic area MT from KMA09 5 Kolster et al. (2009, n 5 4). (I) STS parcellation of KMA09. (J, K) KMA09 parcellations from the left and right
hemispheres of case CH (panel K) and case TO (panel J), with the composite borders of MT and MTp outlined. MTp was identified as a distinct visuotopic map labeled as ‘‘[2]’’ by
Kolster et al. (2009, their Figs 2, 4, 6, and 7) and designated as MTp by Kolster et al. (2010, their Fig. 1) in a reinterpretation of the same data set.
oriented differently so that the 2 maps overlap only about 50%.
The LK02 (Lyon and Kaas 2002) MT map (Fig. 8E) is smaller
and lies mostly lateral to FV91 MT. The PHT00 (Paxinos and
Franklin 2000) (Fig. 8F) and MMF11 (Markov et al. 2011) (Fig.
8G) maps are both based mainly on cytoarchitecture and are
more than twice as large; they extend to the dorsolateral lip of
the STS and also into the anterior bank of the STS.
Kolster et al. (2009) mapped the central visual field
representation of MT and other visual areas using retinotopic
fMRI in 4 hemispheres that were mapped to the F99 atlas. A
probabilistic map of their MT (Fig. 8H) overlaps partially with
FV91 MT but is centered in the lateral half of the STS. The
missing peripheral field representation should be more dorsal
(up on the map) based on the observed retinotopy. (Their visual
stimuli extended to 12 degrees, but their areal boundaries
covered only the central ~6 degrees—cf. their Fig. 2) In the
composite areal parcellation (Fig. 8I), FV91 MT overlaps more
extensively with KMA09 areas MTp (pink), MSTv (blue), and
FST (brown) than with their MT. In maps of the individual
inflated hemispheres (Fig. 8J,K), FV91 MT overlaps substantially
with KMA09 MT in only one hemisphere (TO right) and generally
shows greatest overlap with their MSTv. Kolster et al. (2009, their
Fig. 6, case CH right) noted a good correspondence between
their MT and that of Ungerleider and Desimone (1986) at one
slice level, consistent with our maps (Fig. 8D,J). However, the
discrepancy for the population-average MT as a whole is puzzling.
One possibility is that the KMA09 retinotopic data might be
consistent with a revised parcellation in which MT would include
cortex they assigned to MTp and perhaps part of MSTv.
Comparisons across parcellations for the other extrastriate
visual areas displayed in Figure 8 reveals numerous additional
differences and discrepancies. For some parcellations, even the
early extrastriate areas V2 (yellow), V3d and V3v (green), and
V4 (blue) differ substantially in terms of size and location.
These differences are likely to reflect a combination of
methodological and biological factors (see Discussion).
Parietooccipital Areas
Figure 9 shows parcellations of parietooccipital cortex from 6
studies, with the LV00 borders shown as an overlay on all
Cerebral Cortex October 2012, V 22 N 10 2235
Figure 9. Comparison of macaque parietal regions from 6 studies displayed on the very inflated F99 atlas from medial (top) and dorsal (bottom) views, all with LV00 borders
overlaid. (A) Cortical areas in the macaque parietooccipital cortex from the LV00 architectonic parcellations. White arrow points to the gap between area V1 and the corpus
callosum. (B) FV91 areas. Arrow points to FV91 prostriate area (blue). (C) PHT00 parcellation. Arrow points to PHT00 prostriate area (PS, blue). (D) Area V6 (Galletti et al., 1999)
with LV00 borders overlaid. (E) Parietal areas from Preuss and Goldman-Rakic (1991) and Lyon and Kaas (2002), with LV00 borders overlaid.
panels to facilitate comparisons. Three parcellations—LV00
(Fig. 8A), FV91 (Fig. 9B), and PHT00 (Fig. 9C)—each span most
of this region. Several observations warrant comment, especially in relation to interspecies comparisons (Van Essen et al.
2011).
Macaque V1 extends nearly to the anterior tip of the
calcarine sulcus, where it adjoins a narrow strip of prostriate
cortex (white arrows in Fig. 8A--C) that separates it from the
subiculum and the termination of cortex at the corpus
callosum. In this anterior region, dorsal and ventral halves of
area V2 (yellow) are separated by the prostriate area (blue) and
also by area 29 (in the FV91 parcellation). (See also Van Essen
et al. 1982 for supporting evidence based on interhemispheric
connections.) In contrast, human V1 (area 17) terminates
much farther posterior (near the junction of the POS with the
CaS), and there is a wider expanse of cortex between it and the
corpus callosum (Van Essen et al. 2011).
Retinotopically mapped area V6 of GFG99 (Galletti et al.
1999; pink in Fig. 8D) includes a complete map that emphasizes
the visual periphery. It adjoins the peripheral visual field
representation of area V2. (Indeed, it overlaps partially with
V2 in all 3 illustrated parcellations.) Its retinotopic organization
is similar to that of human hV6. However, hV6 is widely
separated from V2 (Pitzalis et al. 2006; Van Essen et al. 2011), so
it is unclear whether they are genuinely homologous to one
another.
In the intraparietal sulcus (top-bottom row, Fig. 9A--C), the
5 parcellations mapped to the atlas include areas that are
generally elongated parallel to the sulcus. The LV00 parcellation
of the lateral IPS includes LIPv (dark orange), similar to PHT00
area POaI and PGR91 area LIP, and LIPd (light orange), similar
to PHT00 area POaE and PGR91 area 7am. In the fundus and
medial IPS, the LV00 parcellation includes VIPl (light green) and
VIPm (dark green), which are nearly coextensive with PGR91
2236 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
area VIP and overlap partially with FV91 area VIP and with PHT00
areas DIP and PEa.
Size Range and Total Number of Macaque Cortical Areas
The parcellations discussed in this study span a wide range in
terms of their average size and their total number within the
region covered (Table 2). The PHT00 parcellation is the most
fine-grained and includes a total of 161 parcels. However, this is
likely to be a substantial overestimate because some of the
entries represent transitional regions between 2 well-defined
areas (e.g., area 2/1 in between area 2 and area 1). In addition,
some of the smaller areas have not been validated in other
architectonic studies or using other methods.
We generated a composite pan-hemispheric architectonic
map (Fig. 10) that entailed merging of 3 independent parcellations (LV00, FOA00, and PHT00). The rationale is that each of
these parcellations is likely to be more accurate than the others
in some regions but is incomplete or arguably less accurate in
other regions. In any one region, only one parcellation was used.
The composite map includes 129 distinct areas of neocortex and
transitional cortex. These are primarily from the LV00 parcellation in dorsal, lateral, and posterior cortex, primarily from the
FOA00 parcellation in orbital cortex and from all 3 parcellations
in medial cortex. This entails subjective judgment to select
the parcellation for any given region, and the composite
should not be considered ‘‘ground truth’’ or a ‘‘gold standard’’
(see Discussion).
These 129 parcels have an average area of 74 mm2 and
encompass total surface area of 96 cm2, which is 93% of total
right hemisphere cortical surface area (103 cm2). As indicated
in the area-size histogram (Fig. 10E), these areas vary greatly in
size. The total range exceeds 100-fold range between the
2 largest areas (V1, V2) and 9 areas that are all smaller than
10 mm2. A large majority of areas (96) are within an 8-fold
range in the middle (20--160 mm2). If the remaining 7% of
Figure 10. (A) Map of 129 areas from the composite LV-FOA-PHT parcellation shown in lateral, medial, dorsal, and ventral views. Unassigned regions (gray with asterisks, 7% of
total surface area) either lacked a well-defined areal assignment in any of the 3 contributing parcellations or were assigned to a cortical area in one parcellation (e.g., V2v of
PHT00) that was considered less accurate than the corresponding area of a different parcellation (e.g., V2v of LV00). (B) Histogram of areal sizes in the composite parcellation,
calculated using the midthickness surface for the right hemisphere normalized to average rhesus macaque brain size (cf. Table 1).
unassigned cortex (gray regions with asterisks in Fig. 10A)
contains areas similar in size to the average, then macaque
cortex (neocortex plus transitional) contains an estimated 140
total cortical areas based on this composite parcellation.
Discussion
In this study, we used a new method of SBR to expand and
improve upon the set of cortical parcellations registered to the
macaque F99 atlas. Our analysis suggests that macaque cortex
contains about 130--140 areas that span more than 2 orders of
magnitude in average surface area. Comparisons across
different parcellations reveal considerable diversity in the
estimated location and extent of many areas, even ones such
as MT that have been intensively studied for decades. Four
general issues warrant additional discussion: 1) surface-based
atlases and registration; 2) evaluating and reconciling cortical
parcellations; and 3) estimating the total number of cortical
areas in the macaque; and 4) interspecies comparisons.
Registration Algorithms and Atlases
A primary objective of registering data to a cortical atlas is to
maximize intersubject alignment of cortical parcellations and
functional maps. The Landmark-SBR algorithms used in the
present study constrain registration using structural (geographic) features as a surrogate for underlying functional
organization. We evaluated the fidelity of registration based on
the alignment of geographic features themselves and also based
on the consistency of areal alignment in probabilistic maps (Fig.
7C and also Fig. 4 in Van Essen et al. 2001).
The consistency of folding patterns in the macaque enables
accurate delineation of corresponding geographic landmarks in
individual and atlas surfaces. For the Landmark-SBR approach
described here, geographic correspondences in the regions
between landmarks are typically accurate within about 2 mm
based on visual inspection (cf. Fig. 4A). Because areal boundaries
have relatively consistent locations relative to folds in the
macaque, Landmark-SBR also achieves reasonably good intersubject alignment of cortical areas, as shown for probabilistic
architectonic areas (Ferry et al. 2000; Lewis and Van Essen
2000) and retinotopic visual areas (Kolster et al. 2009). The
residual variability present in these probabilistic maps arises from
multiple factors, including: 1) inaccuracies or different criteria for
identifying areal boundaries in individual subjects, 2) interindividual variability in areal size, 3) variability in the location of
areal boundaries relative to the folds used as landmarks, and 4)
local distortions associated with the Landmark-SBR method. While
it is difficult to quantify the relative impact of each factor for the
macaque data sets analyzed here, distortions and biases
attributable to the registration process are likely to be modest
relative to the other factors. Inaccuracies in mapping of areal
borders and/or criterion differences in identifying areal boundaries may be the dominant factors in accounting for misalignment
of many areas (see below).
Some parcellations were mapped to the F99 atlas using older
data sets that had initially been mapped using older algorithms
(the LFF method for flat maps) and in some cases an older atlas
(Case 79-0) as an initial target. These algorithmic differences
might impact comparisons between parcellations to a limited
degree, but they are unlikely to be a major confound. Only
modest deformations are needed to map from one macaque
surface to another, and the density of landmarks used to
constrain the registration was adequate to prevent large local
deformations. In addition, direct visual comparison with results
as displayed on the primary source generally reveal only
modest bias or distortion relative to the other uncertainties in
the parcellation or its representation on the cortical surface.
Such comparisons are clearest when the input parcellation is
available on explicit surface reconstructions (e.g., Figs 4, 5, 7,
and 8). However, good correspondence is generally evident
when comparisons are made in relation to section drawings or
schematic brain drawings in the original published study in
order to assess the fidelity of mapping to the atlas.
It would be of interest to compare the LVD algorithm
introduced here to other Landmark-SBR algorithms derived
using different principles and approaches (e.g., Glaunes et al.
2004; Joshi et al. 2007; Pantazis et al. 2010) and also with EnergySBR methods (Fischl et al. 1999; Yeo et al. 2010) if adapted to
Cerebral Cortex October 2012, V 22 N 10 2237
register macaque surfaces. We were unable to make direct
comparisons using these other algorithms as our data sets.
However, as noted above, the impact of the choice of algorithm
depends greatly on the magnitude of deformations required and
the density of landmarks or other features that constrain the
registration. These issues are much more important when
considering registration of human cortical surfaces, owing to the
greater complexity and variability of human cortical convolutions (Van Essen et al. 2011) and for interspecies registration
(Orban et al. 2004; Van Essen and Dierker 2007).
Several other freely available macaque atlases besides the
F99 atlas have been derived from MR scans and can be viewed
as volumetric data sets in stereotaxic coordinates. Population
average atlases include the 112RM-SL atlas for the rhesus
macaque (McLaren et al. 2009), the Montreal Neurological
Institute (MNI) rhesus plus cynomolgus macaque atlas (Frey
et al. 2011), and the F6 cynomolgus atlas (Vincent et al. 2007),
and the M. nemestrina atlas (Black et al. 2001). Single-subject
atlases include the macaque D99 atlas (Saleem and Logothetis
2007). As with human stereotaxic atlases, methods for accurate
interatlas registration are desirable, so that data generated or
represented on one atlas can be easily transferred to any other.
The F99 atlas considered here includes midthickness surfaces
that have already been registered to D99, 112RM-SL, and F6
space. Registration of F99 to the MNI and D99 (112RM-SL)
atlases will be straightforward to execute as well. Surface maps
of cortical parcellations illustrated here can be readily
converted to volumetric format in any of these alternate
stereotaxic spaces using existing surface-to-volume capabilities
in Caret.
Reconciling Parcellations
A fundamental challenge is to determine whether spatially
overlapping parcels as defined in different studies are coextensive and therefore represent fundamentally the same
underlying cortical domain, whether or not they were delineated by the same or different methods. Another challenge is
to determine which parcellation most accurately reflects the
underlying neurobiological organization (i.e., the ground truth).
Atlas-based comparisons of parcellations as illustrated here do
not resolve these issues, but they do provide useful insights.
Area MT is an informative example, as it has been charted in
many studies using multiple methods. The discrepancies among
architectonically based delineations of MT likely reflect criterion differences between, for example, the cytoarchitectonic
boundaries of PHT00 MT (Paxinos and Franklin 2000) and the
more restricted myeloarchitectonic boundaries of LV00 MT
(Lewis and Van Essen 2000) and FV91 MT (Felleman and Van
Essen 1991). It is difficult to reconcile the discrepancies
between maps of MT using retinotopic-fMRI (Kolster et al.
2009) versus any of the architectonic maps. Retinotopic fMRI
involves complex analysis procedures applied to noisy experimental data; inaccurate estimation of areal boundaries using this
modality might arise from a variety of factors, including
incomplete mapping of the visual field and also inadequate
registration of blood oxygen level--dependent EPI data sets to
the anatomical MR volume. Achieving consistency across
methods remains a highly desirable but an often elusive
objective in the realm of cortical parcellation.
We generated a composite pan-hemispheric parcellation
in order to deal with limitations in available single-study
2238 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
parcellations—either in the amount of cortex covered or in the
reliability of the parcellation in certain regions. For example, in
OMPFC, the FOA00 parcellation has multiple advantages over
other parcellations of this region. It is based on 5 independent
architectonic criteria, it is a population average (n = 5) and thus
less susceptible to the impact of individual variability, and
connectivity data provide independent support for many of these
areal assignments (Carmichael and Price 1996; Ferry et al. 2000).
For other cortical regions, it is more difficult to assess which
parcellation best reflects ground truth. Using a combination of
the LV00 and PHT00, parcellations for non-OMPFC regions has
advantages over either one alone but many alternative hybrids
could plausibly be considered. One issue that arises is that the
boundaries between separate parcellations rarely abut precisely.
This leads to occasional gaps (unassigned cortex) and also
modest overlaps that were adjudicated by trimming one or both
areas.
Patches, Gradients, Modules, and Parcellations
Our estimate of ~140 cortical areas per hemisphere in the
macaque provides one useful value for a fundamental feature of
cortical organization. However, it is subject to the well-known
tension between lumping and splitting of cortical areas. Some
regions considered a single area in this composite parcellation
arguably might be split into finer-grained subdivisions that
warrant being considered separate areas. Alternatively, some
fine-grained parcels arguably might be more appropriately
lumped together in order to constitute neurobiologically welldefined areas.
On what grounds and by what criteria, should any given
transition in properties be considered an areal boundary versus
heterogeneity within an area? There is no simple answer to
such questions. An instructive example involves inferotemporal
cortex, which is implicated in high-level processing of
complex objects, including faces. IT neurons sharing similar
functional properties tend to be clustered. Most notably,
multiple face-selective patches have been identified by
fMRI and neurophysiology (Tsao et al. 2006), and they are
preferentially interconnected (Moeller et al. 2008). Should
each face patch be considered a separate cortical area because
it is distinct from neighboring cortex based on function and
connectivity? A plausible alternative is to consider each face
patch as a module within a larger area, whose other
components might consist of 1) a additional modules that are
comparably distinct but have different specializations (e.g., for
other object categories) or 2) a heterogeneous region with
fluctuations and gradients in functional properties that lack
crisp modularity. For these and other reasons, the debate
between cortical ‘‘lumpers’’ and ‘‘splitters’’ is unlikely to be
quickly resolved. The concept of cortical areas remains
a powerful organizing principle, but properties of modularity
and internal gradients are valuable and complementary
concepts as well. Progress on these issues will benefit from
neuroimaging methods that enable systematic analysis across
large cortical expanses and objective analysis of cortical
properties and their spatial gradients (see Glasser and Van
Essen 2011; Van Essen et al. 2011)
Interspecies Comparisons
A better understanding of macaque cortical parcellations
will provide valuable comparative data that can aid in
understanding human cortical parcellation. A companion study
(Van Essen et al. 2011) also uses a surface-based analysis
strategy to map published human cortical parcellations to an
improved surface-based atlas. Progressively more accurate
parcellations that are mapped to atlas surfaces in both species
will facilitate evaluations of presumed homologies between
species. This will also aid in quantitative assessments of
differential cortical expansion during evolution (Orban et al.
2004; Van Essen and Dierker 2007). Interspecies comparisons
will also be aided by analyses of features that tend to be
evolutionarily conserved, such as myelin content (Glasser and
Van Essen 2011) and cortical connectivity patterns.
Funding
National Institutes of Mental Health grant RO1 MH60974.
Notes
We thank S. Danker for assistance in manuscript preparation, E. Reid for
assistance with the data analysis, and Drs J. L. Price, H. Kolster,
G. Orban, H. Kennedy, N. Markov, and C. Galletti for providing data sets
used to map cortical parcellations to the atlas. Conflict of Interest :
None declared.
References
Bayliss GC, Rolls ET, Leonard CM. 1987. Functional subdivisions of the
temporal lobe neocortex. J Neuroscience. 7:330--342.
Black KJ, Koller JM, Snyder AZ, Perlmutter JS. 2001. Template images
for nonhuman primate neuroimaging: 2. Macaque. Neuroimage.
14:744--748.
Brodmann K. 1905. Beitraege zur histologischen Lokalisation der
Grosshirnrinde: dritte Mitteilung: Die Rindenfelder der niederen
Affen Journal fuer Psychologie und Neurologie. Neuroimage. 4:5/
6:177--226.
Carmichael ST, Price JL. 1994. Architectonic subdivision of the orbital
and medial prefrontal cortex in the macaque monkey. J Comp
Neurol. 346:366--402.
Carmichael ST, Price JL. 1996. Connectional networks within the
orbital and medial prefrontal cortex of macaque monkeys. J Comp
Neurol. 371:179--207.
Drury HA, Van Essen DC, Anderson CH, Lee CW, Coogan TA, Lewis JW.
1996. Computerized mappings of the cerebral cortex: a multiresolution flattening method and a surface-based coordinate system.
J Cogn Neurosci. 8:1--28.
Felleman DJ, Van Essen DC. 1991. Distributed hierarchical processing in
the primate cerebral cortex. Cereb Cortex. 1:1--47.
Ferry AT, Ongur D, An X, Price JL. 2000. Prefrontal cortical projections
to the striatum in macaque monkeys: evidence for an organization
related to prefrontal networks. J Comp Neurol. 425:447--470.
Fischl B, Sereno MI, Tootell RB, Dale AM. 1999. High-resolution
intersubject averaging and a coordinate system for the cortical
surface. Hum Brain Mapp. 8:272--284.
Frey S, Pandya DN, Chakravarty MM, Bailey L, Petrides M, Collins DL.
2011. An MRI based average macaque monkey stereotaxic atlas and
space. MNI monkey space. Neuroimage. 55:1435--1442.
Galletti C, Fattori P, Gamberini M, Kutz DF. 1999. The cortical visual
area V6: brain location and visual topography. Eur J Neurosci.
11:3922--3936.
Glasser M, Van Essen D. 2011. Mapping human cortical areas in vivo
based on myelin content as revealed by T1 and T2-weighted MRI.
J Neurosci. 31:11597--11616.
Glaunes J, Vaillant M, Miller MI. 2004. Landmark matching via large
deformation diffeomorphisms on the sphere. J Math Imaging Vis.
20:179--200.
Joshi AA, Shattuck DW, Thompson PM, Leahy RM. 2007. Surfaceconstrained volumetric brain registration using harmonic mappings.
IEEE Trans Med Imaging. 26:1657--1669.
Joshi SC. 1997. Large deformation landmark based diffeomorphic
for image matching [PhD thesis] [St. Louis (MO)]: Sever Institute,
Washington University.
Kaas JH. 2005. The future of mapping sensory cortex in primates: three
of many remaining issues. Philos Trans R Soc Lond B Biol Sci.
360:653--664.
Kolster H, Mandeville JB, Arsenault JT, Ekstrom LB, Wald LL,
Vanduffel W. 2009. Visual field map clusters in macaque extrastriate
visual cortex. J Neurosci. 29:7031--7039.
Kolster H, Peeters R, Orban GA. 2010. The retinotopic organization of
the human middle temporal area MT/V5 and its cortical neighbors.
J Neurosci. 30:9801--9820.
Lewis JW, Van Essen DC. 2000. Mapping of architectonic subdivisions in
the macaque monkey, with emphasis on parieto-occipital cortex.
J Comp Neurol. 428:79--111.
Lyon DC, Kaas JH. 2002. Connectional evidence for dorsal and ventral
V3, and other extrastriate areas in the prosimian primate, Galago
garnetti. Brain Behav Evol. 59:114--129.
Markov NT, Misery P, Falchier A, Lamy C, Vezolia J, Quilodran R,
Giroud P, Gariel MA, Ercsey-Ravasz MM, Pilaz LJ, et al. 2011.
Computational methods for reconstructing and unfolding the
cerebral cortex. Cereb Cortex. 21:1254--1272.
Maunsell JH, Van Essen DC. 1983. Functional properties of neurons in
middle temporal visual area of the macaque monkey. I. Selectivity
for stimulus direction, speed, and orientation. J Neurophysiol.
49:1127--1147.
Maunsell JH, Van Essen DC. 1987. Topographic organization of the
middle temporal visual area in the macaque monkey: representational biases and the relationship to callosal connections and
myeloarchitectonic boundaries. J Comp Neurol. 266:535--555.
McLaren DG, Kosmatka KJ, Oakes TR, Kroenke CD, Kohama SG,
Matochik JA, Ingram DK, Johnson SC. 2009. A population-average
MRI-based atlas collection of the rhesus macaque. Neuroimage.
45:52--59.
Moeller S, Freiwald WA, Tsao DY. 2008. Patches with links: a unified
system for processing faces in the macaque temporal lobe. Science.
320:1355--1359.
Orban GA, Van Essen D, Vanduffel W. 2004. Comparative mapping of
higher visual areas in monkeys and humans. Trends Cogn Sci.
8:315--324.
Pantazis D, Joshi A, Jiang J, Shattuck DW, Bernstein LE, Damasio H,
Leahy RM. 2010. Comparison of landmark-based and automatic
methods for cortical surface registration. Neuroimage.
49:2479--2493.
Paxinos G, Franklin KBJ. 2000. The mouse brain in stereotaxic
coordinates. San Diego, CA: Academic Press.
Pitzalis S, Galletti C, Huang RS, Patria F, Committeri G, Galati G,
Fattori P, Sereno MI. 2006. Wide-field retinotopy defines human
cortical visual area v6. J Neurosci. 26:7962--7973.
Preuss TM, Goldman-Rakic PS. 1991. Architectonics of the parietal and
temporal association cortex in the strepsirhine primate Galago
compared to the anthropoid primate Macaca. J Comp Neurol.
310:475--506.
Saad ZS, Reynolds RC, Argall RC, Japee S, Cox RW. 2004. SUMA: an
interface for surface-based intra- and inter-subject analysis with AFNI.
Proceedings of the 2004 International Symposium on Biomedical
Imaging, Arlington, VA, IEEE. p. 1510--1513.
Saleem KS, Logothetis NK. 2007. A combined MRI and histology atlas of
the rhesus monkey brain. San Diego, CA: Academic Press.
Seltzer B, Pandya DN. 1978. Afferent cortical connections and
architectonics of the superior temporal sulcus and surrounding
cortex in the rhesus monkey. Brain Res. 149:1--24.
Seltzer B, Pandya DN. 1980. Converging visual and somatic sensory
cortical input to the intraparietal sulcus of the rhesus monkey. Brain
Res. 192:339--351.
Seltzer B, Pandya DN. 1986. Posterior parietal projections to the
intraparietal sulcus of the rhesus monkey. Exp Brain Res.
62:459--469.
Tsao DY, Freiwald WA, Tootell RB, Livingstone MS. 2006. A cortical
region consisting entirely of face-selective cells. Science.
311:670--674.
Cerebral Cortex October 2012, V 22 N 10 2239
Ungerleider LG, Desimone R. 1986. Cortical connections of visual area
MT in the macaque. J Comp Neurol. 248:190--222.
Van Essen D, Newsome WT, Bixby JL. 1982. The middle temporal visual
area in the macaque: myeloarchitecture, connections, functional
properties and topographic organization. J Neurosci. 2:265--283.
Van Essen D, Newsome WT, Maunsell JH. 1984. The visual field
representation in striate cortex of the macaque monkey: asymmetries, anisotropies and individual variability. Vision Res. 24:420--448.
Van Essen DC. 2002a. Surface-based atlases of cerebellar cortex in the
human, macaque, and mouse. Ann N Y Acad Sci. 978:468--479.
Van Essen DC. 2002b. Windows on the brain: the emerging role of atlases
and databases in neuroscience. Curr Opini Neurobiol. 12:574--579.
Van Essen DC. 2004. Surface-based approaches to spatial localization
and registration in primate cerebral cortex. Neuroimage.
23(Suppl 1):S97--S107.
Van Essen DC. 2005. A population-average, landmark- and surfacebased. PALS. Atlas of human cerebral cortex. Neuroimage.
28:635--662.
Van Essen DC, Dierker DL. 2007. Surface-based and probabilistic atlases
of primate cerebral cortex. Neuron. 56:209--225.
Van Essen DC, Drury HA. 1997. Structural and functional analyses of
human cerebral cortex using a surface-based atlas. J Neurosci.
17:7079--7102.
Van Essen DC, Drury HA, Joshi S, Miller MI. 1998. Functional and
structural mapping of human cerebral cortex: solutions are in the
surfaces. Proc Natl Acad Sci U S A. 95:788--795.
2240 Macaque Surface-Based Cortical Parcellations
d
Van Essen et al.
Van Essen DC, Glasser MF, Dierker D, Harwell J, Coalson T. 2011.
Parcellations and hemispheric asymmetries of human cerebral
cortex analyzed on surface-based atlases. Cerebral Cortex. (in press).
Van Essen DC, Harwell J, Hanlon D, Dickson J. 2005. Surface-based
atlases and a database of cortical structure and function. In:
Koslow SH, Subramaniam S, editors. Databasing the brain: from
data to knowledge. Neuroinformatics. Hoboken, NJ: John Wiley &
Sons. p. 369--388.
Van Essen DC, Lewis JW, Drury HA, Hadjikhani N, Tootell RB,
Bakircioglu M, Miller MI. 2001. Mapping visual cortex in monkeys
and humans using surface-based atlases. Vision Res. 41:1359--1378.
Van Essen DC, Maunsell JH. 1980. Two-dimensional maps of the
cerebral cortex. J Comp Neurol. 191:255--281.
Van Essen DC, Maunsell JH, Bixby JL. 1981. The middle temporal visual
area in the macaque: myeloarchitecture, connections, functional
properties and topographic organization. J Comp Neurol.
199:293--326.
Vincent JL, Patel GH, Fox MD, Snyder AZ, Baker JT, Van Essen DC,
Zempel JM, Snyder LH, Corbetta M, Raichle ME. 2007. Intrinsic
functional architecture in the anaesthetized monkey brain. Nature.
447:83--86.
von Bonin G, Bailey P. 1947. The neocortex of Macaca Mulatta. Urbana
(IL): University of Illinois.
Yeo BT, Sabuncu MR, Vercauteren T, Ayache N, Fischl B, Golland P.
2010. Spherical Demons: fast diffeomorphic landmark-free surface
registration. IEEE Trans Med Imaging. 29:650--668.