cortical cartography using the discrete conformal approach ... · introduction the human brain can...
TRANSCRIPT
www.elsevier.com/locate/ynimg
NeuroImage 23 (2004) S119–S128
Cortical cartography using the discrete conformal approach
of circle packings
Monica K. Hurdala,* and Ken Stephensonb
aDepartment of Mathematics, Florida State University, Tallahassee, FL 32306-4510, USAbDepartment of Mathematics, University of Tennessee, Knoxville, TN 37996-1300, USA
Available online 23 September 2004
Cortical flattening algorithms are becoming more widely used to assist
in visualizing the convoluted cortical gray matter sheet of the brain.
Metric-based approaches are the most common but suffer from high
distortions. Conformal, or angle-based algorithms, are supported by a
comprehensive mathematical theory. The conformal approach that
uses circle packings is versatile in the manipulation and display of
results. In addition, it offers some new and interesting metrics that may
be useful in neuroscientific analysis and are not available through
numerical partial differential equation conformal methods.
In this paper, we begin with a brief description of cortical bflatQmapping, from data acquisition to map displays, including a brief
review of past flat mapping approaches. We then describe the
mathematics of conformal geometry and key elements of conformal
mapping. We introduce the mechanics of circle packing and discuss its
connections with conformal geometry. Using a triangulated surface
representing a cortical hemisphere, we illustrate several manipulations
available using circle packing methods and describe the associated
bensemble conformal featuresQ (ECFs). We conclude by discussing
current and potential uses of conformal methods in neuroscience and
computational anatomy.
D 2004 Elsevier Inc. All rights reserved.
Keywords: Circle packing; Computational anatomy; Conformal map;
Cortical flat map; Mapping; Riemann Mapping Theorem; Surface
flattening
Introduction
The human brain can be divided into various regions based on
function and anatomy. The surface is a highly convoluted sheet,
with the folds (gyri) and fissures (sulci) varying in size, position,
and extent from person to person. It is estimated that 60–70% of the
cortical surface is buried in folds and hidden from view. The gray
matter, or cerebrum, is the thin layer of cortical surface where most
of the functional processing of the brain occurs and is about 3–5 mm
1053-8119/$ - see front matter D 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.neuroimage.2004.07.018
* Corresponding author. Fax: +1 850 644 4053.
E-mail address: [email protected] (M.K. Hurdal).
Available online on ScienceDirect (www.sciencedirect.com.)
thick. Due to the two-dimensional sheet topology of the cortex,
some methods have been proposed that bunfoldQ and bflattenQ thecerebrum. These bflat mapsQ of the brain are useful not only as
visualization methods, but may facilitate registration, comparison,
and recognition of structural and functional cortical relationships.
In this paper, we discuss a discrete mapping approach that uses
circle packings to produce bflattenedQ images of cortical surfaces in
the sphere, the euclidean plane, and the hyperbolic plane. This
approach is founded upon a rich body of classical mathematics that
is now becoming available in discrete, computable form via circle
packing. Our maps exploit the conformal geometry of cortical
surfaces and are in fact quasi-conformal approximations of classical
conformal maps (as are bconformalQ maps produced via other
approaches); we refer to them as discrete conformal (flat) maps.
Cortical flattening overview
The main processing sequence for cortical flattening is (1)
acquisition of 3D data; (2) extraction, discretization, and triangu-
lation of the relevant surface(s); (3) flat mapping computations;
and (4) analysis, manipulation, or display of results.
1. Acquisition
For brain flattening applications, the data processing sequence
typically begins with the acquisition of a 3D magnetic resonance
imaging (MRI) volume of a region of interest. There are many data
processing steps involved, including correcting for inhomogene-
ities, stripping extraneous tissues such as the skull and scalp, the
identification and perhaps color-coding of structures of interest,
and possibly alignment of functional data with anatomical data.
Although these steps represent major tasks, they are considered
preprocessing for purposes of this paper; the reader is referred to
the literature for more details on these methods.
2. Surface extraction
A topologically correct, triangulated surface mesh is required
before flattening can begin. The surface of interest may cover the
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128S120
entire cortex, a hemisphere, or a region within a hemisphere, such
as a lobe. A surface representation is created from the preprocessed
3D volumetric data, typically using an isosurface or threshold-
based algorithm that approximates the white matter–gray matter
interface, the gray matter–cerebrospinal fluid (CSF) interface, or a
surface midway between. We assume here that the surface of
interest is either a full cerebral hemisphere (or cerebellum) and
hence a topological sphere, a surface region that is a topological
disc (i.e., without holes), or a multiply connected surface region (a
topological disc with one or more holes) such as a topological
annulus. The most common algorithms for surface extraction are
marching cubes (Lorensen and Cline, 1987) and its derivatives
(e.g., marching tetrahedra) due to their speed and ease of
implementation. Inadequate spatial resolution of MRI scans, as
well as certain ambiguities that can arise during isosurface creation,
means that the initial surface extraction will generally require
correction. In addition to routine artifacts, there are several
common topology errors, including holes in the surface (more
than one boundary component), multiple connected components,
nonmanifold edges (edges contained in three or more triangles),
and handles. Many defects can be detected via the Euler
characteristic formula (Massey, 1967) and automatically corrected
(Hurdal, 2004b); these issues are actively being pursued by Hurdal.
There is also progress on extraction methods which directly
generate topologically correct surfaces (Han et al., 2002; Mac-
Donald et al., 2000; Mangin et al., 1995; Stern et al., 2001).
3. Flattening
We repeat that all flattening approaches require the final
surface that is embedded in R3 to be a topologically correct
triangulated mesh. (Meshes can be generated with nontriangular
faces, but our method requires that those faces be further broken
into triangles.) Flattening refers to the creation of a one-to-one
simplicial map (i.e., triangles go to triangles) between the
triangulated 3D surface and a triangulation in bflatQ space.
Flat space is interpreted to mean a space of constant Gaussian
curvature: the sphere, the euclidean plane, or the hyperbolic plane.
Surfaces with boundary edges can be mapped to any one of these;
surfaces without boundary must be mapped to the sphere unless a
boundary is introduced by removing one or more topological
discs. (Note: bflat mapQ refers equally to the mapping itself as a
function and to its image.)
Flattening methods fall into three broad categories: graph-
based, metric-based, and conformal-based methods. Combinato-
rially speaking, triangulated meshes are planar graphs, and every
planar graph has a straight-line embedding (i.e., flattening) (Fary,
1948; Wagner, 1936). There is a substantial body of established
embedding techniques with various target criteria, such as
convexity or resolution, often having extremely efficient imple-
mentations. These graph-based methods disregard geometric
information in the surface and hence have not been used in
cortical flattening.
Metric-based algorithms were the first implemented for the
cortical flattening problem, in part, because metric features—
euclidean lengths and areas—are the most familiar of surface
features. As a result, metric-based approaches are the most popular
and well-known. These algorithms typically involve an iterative
strategy that attempts to minimize some combination of areal,
linear, and angular distortion between the original 3D mesh and its
flattened counterpart. Van Essen et al. (Drury et al., 1996; Van Essen
and Drury, 1997; Van Essen et al., 1998) have developed methods
based on longitudinal and torsional forces; Fischl et al. (1999a,b)
use gradient methods incorporating geodesic distance and area.
Other methods using similar premises have also been developed for
cortical flattening (Goebel, 2000; Wandell et al., 2000) and in
visualization and computational geometry. We make two points vis-
a-vis our later discussion: First, flat maps that preserve metric
information do not exist; algorithms attempt to minimize distortion;
some even introduce ad hoc cuts into the surfaces for this purpose.
Second, performance of these algorithms in terms of distortion (be it
linear, areal, or angular) is judged based on noncanonical and
blocalQ features.Conformal-based methods include our circle packing approach
and alternative numerical partial differential equation (PDE)
methods. bConformalQ notions are not widely known outside of
mathematics, and a principal goal of this paper is to introduce and
illustrate the central concepts. Among conformal methods, PDE
approaches rely on standard numerical techniques; they tend to be
very fast, but they apply in only limited settings. The circle packing
approach is known for its versatility, as will be demonstrated in our
illustrations. It is more soundly grounded in theory and more
comprehensive in the features of conformal geometry that it can
bring to bear in applications.
4. Results
After flat maps have been computed, they can be used for
manipulation, analysis, display, and even storage. We show many
images in this paper, but wish to emphasize that flat mapping is
much more than (perhaps not even primarily) a visualization
method. The desire to encode binformationQ about the target
surfaces, which will be useful in applications such as neuroscience,
is driving interest in conformal mapping methods. In the next
section, we discuss the nature of bconformalQ information; later we
discuss its potential uses.
Conformal surfaces
The mathematical study of surfaces recognizes properties of
two types: those intrinsic to the surface itself and those having to
do with how the surface sits in 3-space. Surface-based techniques
naturally target the intrinsic properties. The most familiar are
metric properties: lengths of surface curves, areas of surface
patches, Gaussian curvature. (Note: mean curvature is not intrinsic,
though useful for coloring of flat maps.)
Less well-known intrinsic properties are those termed
bconformalQ, which are grounded in the ability to measure angles
between curves on the surface at points where they intersect. If
there is a consistent way to measure such angles, the surface is
termed a Riemann surface and is said to have a conformal
structure. All surfaces encountered in cortical studies may be
assumed to be Riemann surfaces.
Flattening can be used purely for visualization, but the main
scientific value will lie in the ability of a flattening method to
preserve binformationQ about the surface, so that properties
observed in the flat image reflect properties inherent in the surface
rather than artifacts of the flattening process. It has been known
since Gauss that a surface with variable Gaussian curvature (such
as a cortical surface) cannot be flattened in a way that preserves
metric structure, lengths, and areas (Polya, 1968). On the other
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128 S121
hand, it has been known since Riemann (Gauss’s student) that a
surface can be flattened in a way that preserves conformal
structure. Mappings between surfaces that preserve angles between
intersecting curves (in magnitude and orientation) are known as
conformal maps.
The three classical geometric spaces (those of constant
Gaussian curvature) are the sphere P, the plane C, and the
hyperbolic plane D. In computations, P is the unit sphere at the
origin in R3, C is R2 identified as the complex plane, and
D = f x; yð Þ: r bffiffiffiffiffiffiffiffiffiffiffiffiffiffix2 þ y2
pb 1g is the unit disc which is treated as
the Poincare model for the hyperbolic plane. It is well-known
how to measure angles in these spaces; moreover, these are
archetypes for the geometries of all Riemann surfaces, as
indicated in the famous Riemann Mapping Theorem (RMT) of
1851 (as extended to surfaces by Koebe) (Riemann, 1876).
Theorem 3.1 (RMT). If S is a simply connected Riemann surface,
then there exists a one-to-one conformal mapping of S onto one of
the standard spaces P; C or D.
There are several observations regarding conformal geometry
and the mathematics of conformal mapping(s) that is pertinent to
scientific uses:
(a) Conformal structure has been a core concept of mathematics
for nearly two centuries, from number theory and the famous
Riemann Hypothesis to analytic functions to the latest
conformal string theory. Teichmqller theory for Riemann
surfaces is an example of the exquisitely nuanced information
resident in conformal structures.
(b) For maps between plane regions, being conformal is
equivalent to being analytic. Such maps are characterized
as those having derivatives (in complex arithmetic) or
equivalently those satisfying the Cauchy–Riemann PDEs.
The maps are also intimately related to the Laplace operator
and harmonic functions.
(c) Conformal maps between plane regions have been central to
physical and engineering studies for decades—fluid flow,
electrostatics, airfoil design, residue theory. There are literally
thousands of papers on numerical approximation of such
planar conformal maps (many predating computers).
(d) Only in the last decade have methods become available for
approximating conformal mappings for general (i.e., non-
planar) surfaces, motivated by mathematical applications,
visualization p imaging, and computational geometry. Circle
packing was among the first and remains the most compre-
hensive and versatile approach. Others involve numerical PDE
methods for solving the Cauchy–Riemann equations, har-
monic energy minimization for solving the Laplace–Beltrami
equation, and differential geometric methods based on
approximation of holomorphic differentials (Angenent et al.,
1999; Gu and Yau, 2002; Gu et al., 2003; Ju et al., 2004; Levy
et al., 2002). These methods do not require equilateral surface
triangulations and are still able to obtain conformal maps.
(e) In applications, all flat mapping methods involve (1)
discretization of the surface (as with triangulations) followed
by (2) numerical flattening of the result. However, computed
maps are never conformal in the strict sense—angle is a local
infinitesimal feature that can survive neither (1) nor (2). In
technical terms, all computed maps are quasi-conformal
rather than conformal. One measure of the extent to which a
mapping distorts conformal information is reflected in a
parameter j z 1 known as dilatation. A j-quasi-conformal
map is one whose worst local dilatation is bounded above by
j. One can get a sense of this by picturing a map of a circle to
an ellipse having ratio (major axis–minor axis) = j; the closerj is to one, the less distortion. The objective goal is the
construction of flat maps with j as close to 1 as possible. A 1-
quasi-conformal map is in fact conformal. However, that goal
is more nuanced in practice. Regarding a given conformal
method: Can one estimate j? Can large j be avoided? Can jbe brought close to 1? Close to 1 outside some small set? j is
in a sense a worst-case measure of how bconformalQ a
mapping is, and it is likely that experience will lead to more
useful constructs.
Exploiting bconformalQ notions in applications implies exploit-
ing nonlocal and global features, which are fellow travelers with
angle. These reside somehow in the bconformal structuresQthemselves. We will call these ensemble conformal features
(ECFs) and will discuss and illustrate several with examples in the
Illustration section. The goal in numerical conformal flat mapping
is to bring the ensemble conformal features through the surface
discretization and flattening stages of the approximation process to
the maximum extent possible. The scientific goal is then to learn
the uses, and perhaps meaning, of this conformal information.
Circle packing and its mechanics
Conformal flat mapping was initiated by our group in the
context of cortical flattening. However, our circle packing
methods lie within a much broader theory, tracing back to a
conjecture of Field’s Medalist William Thurston, circa 1985.
Circle packing has rapidly developed into a comprehensive
discrete analytic function theory that both parallels and approx-
imates the classical continuous theory. We provide some basics
on circle packing mechanics along with some images, and hope
that this will convey some of the key intuition; see the surveys
(Stephenson, 1999; Stephenson, 2002) and the forthcoming book
(Stephenson, in preparation) for details.
A circle packing is a configuration of circles with a prescribed
pattern of tangencies. Our packings lie in the geometric spaces P,
C and D and each will be univalent, meaning that its circles have
mutually disjoint interiors. The pattern of tangencies will be
prescribed by the surface triangulation T. Abstractly, T is a
simplicial 2-complex, that is, an encoding of the connectivity
involving vertices, edges, and triangular faces. A circle packing P
associated with T in a geometry G (one of P, C, or D) must have a
circle cv for each vertex v of T, so that if hv, ui is an edge of T, then
cv and cu are tangent, and if hv, u, wi is a positively oriented faceof T, then hcv, cu, cwi is a positively oriented triple of mutually
tangent circles. In other words, the triangulation in space tells us
the pattern of tangencies required of the circles in P. The bcarrierQ,carr(P), is the collection of triangles in G obtained by connecting
the centers of tangent circles of P, so these triangles necessarily
have the same pattern as T. Later, we also discuss inversive
distance packings in which tangency is replaced by a conformally
invariant notion of separation between pairs of circles. An
example of a circle packing is shown in Fig. 1.
Definition 4.1. Suppose P is a circle packing for T in one of P, C
or D. The associated discrete conformal flat map refers to the
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128S122
continuous and one-to-one simplicial map identifying each triangle
of the surface triangulation with the corresponding triangle in
carr(P).
A typical triangulation of a cerebral hemisphere involves
150,000 vertices, each of which is associated with a circle when
computing P. One may be forgiven for being skeptical that this
many circles can so tightly choreograph their sizes as to meet the
combinatorics dictates of T. However, that is the content of the
Discrete Riemann Mapping Theorem (Rodin and Sullivan, 1987;
Beardon and Stephenson, 1990).
Theorem 4.2 (DRMT). Given any triangulation T of a simply
connected surface S , there exists an associated univalent circle
packing P in G (one of P, C, or D) with the combinatorics of T.
Hence, there exists a discrete conformal flat map f:S Y G from Sto G.
Overview of packing mechanics
The process begins with the abstract combinatorics of the
given triangulation T, metric data from the 3D surface (if using
inversive distance packings), and choice of a geometry. If T is a
topological disc, then boundary conditions are set, meaning that
the radii of boundary circles or the sum of the angles in the faces
meeting at a boundary vertex are specified. The main computa-
tional task is in finding radii of the circles. This is an iterative
process driven by a packing condition that must be achieved for
all interior circles; it is nonlinear, accounting for the sometimes
length computations. Nonetheless, a typical cortical lobe may
involve 40,000 circles and takes only seconds on a PC for the
tangency packing computation. Once radii are known, a packing
P can be laid out. Spherical packings are stereographically
projected from packings in C. The software CirclePack is
available for computing circle packings of topologically correct
surfaces (Stephenson, 1992–2004), and algorithm details can be
found in Collins and Stephenson (2003). The software TopoCV
is also available (Hurdal, 2004a) for checking and correcting
surfaces with topological errors before using CirclePack.
Typical surface interaction features, such as color coding,
geodesic tracing, area computations, interactive data exchange
between the 3D and flat images, can be applied between the
original cortical surface in R3 and its flattened image. Our aim in
Fig. 1. Circle packing mechanics illustrated. Left: 3D surface. Middl
this paper is to show the mapping options and ensemble
conformal features (ECFs).
The extent to which our circle packing maps are bconformalQ innature and j-quasi-conformal in fact involves considerations that
we can only briefly describe here, and not all are yet fully
resolved.
(1) If the triangles forming the surface S in 3D were equilateral,
then one could define an in situ circle packing Q on S . In the
fuller discrete theory, the map f:Q Y P is a discrete
conformal map: there are known bounds on its dilatation j,and with well-established refinement procedures, one can
obtain improved maps that actually converge to a classical
(i.e., truly) conformal map of S to G. (As the first general
realization of the 150-year-old Riemann Mapping Theorem,
this is a singular accomplishment.)
(2) In point of fact, cortical triangulations are determined by
numerous factors, from image resolution to data correction
and sampling, and are never equilateral. When the data
reaches our group, the principal impacts on triangle shape
are due to surface extraction and smoothing. If the triangles
are moderately equilateral, then circle packing computations
can be adjusted by replacing tangency packings with
inversive distance packings, as described shortly. (There
are several inversive distance packings among our illustra-
tions.) The inversive distance theory is not as complete as
that for tangency packings, but in practice, inversive
distance packings are becoming routine for lobe-size data
sets and they considerably reduce the quasi-conformal
distortion in the associated conformal flat maps. More
details of inversive distances can be found in Bowers and
Hurdal (2003).
(3) With wire-frame or marching cube methods, and especially
with ad hoc smoothing, surface triangulations tend to be less
equilateral—often spectacularly so. Discrete conformal meth-
ods compensate only in part and the results have greater (but
quantifiable) distortion. A growing body of experience
indicates that the combinatorics alone—irrespective of
triangle shapes—capture many ECFs. This is in part due to
elements of randomness: noncoherent distortions tend to
cancel, leaving an ECF largely in tact. This effect is not yet
quantified and a key goal in ongoing work concerns stability
e: tangency packing, P1. Right: inverse diastance packing, P2.
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128 S123
of ECFs concerning triangle shapes, retriangulations, refine-
ments, and so forth.
Illustrations
Fig. 1 displays discrete conformal flat map images f:S Y C on
a small surface fragment S . The packing P1 is a tangency packing
(Fig. 1, middle); one sees the circles, the edges connecting tangent
circles, and the triangles forming the carrier, the flat image. The
angle sum for a circle cv in P is the sum cv of angles in the faces
meeting at its center; hv must be 2p for interior circles—this is the
packing condition we alluded to above. In this example, angle sum
hw = p was prescribed for all boundary circles cw, save hw = p/2for the four corner circles. The next packing, P2, is an inversive
distance packing (Fig. 1, right); each edge of the carrier runs
between two circle centers, and the separation between those
circles was prescribed in advance based on 3D metric data in S .The color coding in P1 and P2 reflects quasi-conformal dilatation jfrom 1 to roughly 2 (light to dark red). Loosely speaking, j = 1.50
represents 50% distortion in the local bworst-caseQ sense. Noticethat the quasi-conformal distortion is reduced considerably in the
inversive distance packing.
The least familiar of the standard geometries is that of the
hyperbolic plane, represented as the unit disc D, famous for being
bnon-euclideanQ. The packings on the left in Fig. 2 are maximal
packings because they fill the plane—their outer circles are
horocycles, internally tangent to the unit circle and of infinite
hyperbolic radius. Several features are attractive in the hyperbolic
setting, among them, a common canonical domain and manipu-
lation with Mfbius transformations, conformal mappings whose
actions are illustrated on the right.
Figs. 3 and 4 are based on a single data set. The gray matter–
CSF surface representing the left cerebral hemisphere was
Fig. 2. Hyperbolic packings and manipulations.
extracted from a 1 � 1 � 1 mm MRI scan of a normal human
adult and parcellated into lobes. The resulting triangulation T
(146,922 vertices) is a topological sphere.
We first discuss Fig. 3. Top right are the lateral and medial
views of the full triangulated surface S in 3D; the surface
extraction method gives triangles only moderately nonequilateral.
To the left are the associated discrete conformal flat maps in P,
C, and D; the latter two are based on the triangulated topological
disc L obtained from the triangulation of S by removing the
11,600 vertices of the corpus callosum and ventricle (i.e., nongray
matter). The euclidean packing of L preserves the edge lengths of
boundary edges from S . The packing in the unit disc is treated in
hyperbolic geometry; this is the essentially unique maximal
packing for L.
In the lower part of the figure, we have isolated the occipital
lobe from the euclidean flat packing. This is in fact an inversive
distance packing, as one sees in the blowup. Four physical
landmarks have been indicated on the boundary of the occipital
lobe and we have created a new flat mapping to a rectangle that
maps those points to the corners. This is displayed using the faces
of the carrier rather than its circles. A grid imposed on the
rectangle can be pulled back to this lobe on any of the flat maps
and to the surface as well.
In Fig. 4, we have added three orange patterns on the occipital
lobe in various of its manifestations. This could represent, for
example, an overlay of functional MRI (fMRI) data obtained on
this subject. These are pure simulations, however, chosen with
large, identifiable shapes for convenience in this demonstration.
The identical patches are shown in blowups as they appear on (1)
the occipital lobe of the 3D cortical surface; (2) the spherical map;
(3) the hyperbolic maximal packing; (4) the euclidean inversive
distance occipital map; (5) the rectangular occipital map. Our point
in this is to let the reader compare the similarity of the bshapesQ andrelative positions of the patches in the various settings. At the
bottom, we have manipulated the occipital lobe based on these
patterns: the rectangular packing from earlier, nowwith the patterns;
cutting out the largest of the islands and mapping the resulting
annulus to a round annulus, A = f ðx; yÞ: r bffiffiffiffiffiffiffiffiffiffiffiffiffiffix2 þ y2
pb Rg;
cutting out all three islands; and mapping to a rectangle with three
round holes.
Ensemble conformal features
Underlying all our work are two fundamental facts regarding
conformal maps (both classical and discrete): They exist and they
are canonical (vis-a-vis the given surface), meaning that they are
uniquely determined (based on boundary conditions and up to
standard Mfbius normalization). Conformal invariants are
surface quantities that remain unchanged under conformal maps,
such as curve intersection angles. However, as mentioned earlier,
angles are not preserved in numerical flattening (since angles are
a local infinitesimal feature); this is the reason why we are
concentrating on bensembleQ features. Due to space limitations,
we emphasize these two principal ECFs.
Extremal length of annuli
The extremal length of the round annulus A = f ðx; yÞ: r bffiffiffiffiffiffiffiffiffiffiffiffiffiffix2 þ y2
pb Rg is EL(A) = log(R/r)/2p. Annuli A and B are
conformally equivalent if and only if EL(A) = EL(B). Every
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128S124
annular subregion A of a (Riemann) surface can be mapped
conformally onto a unique (up to similarity) round annulus A;
hence, one defines EL(A) = EL(A). Extremal length, large
or small, means roughly that the annulus is fat or thin,
respectively.
Fig. 3. Circle packings of a left cerebral hemisphere. Top right: lateral and med
euclidean maps. Bottom left: inversive distance (euclidean) packing of occipital lo
Extremal length of quadrilaterals
The extremal length of the rectangle Q = [0,L] � [0,W] of
length L and width W is EL(Q) = L/W. Rectangles Q and R are
conformally equivalent (with bendsQ identified) if and only if
ial views of cerebral hemisphere. Top left: spherical, hyperbolic disc, and
be along with an enlargement. Bottom right: rectangle map of occipital lobe.
Fig. 4. Manipulations of simulated activation sites. Top: lateral view of cortical hemisphere and an enlargement of the occipital lobe containing simulated
activations. Counterclockwise from left: spherical and hyperbolic hemisphere maps with activations along with enlargements of the occipital region;
euclidean hemisphere map; euclidean and rectangle maps of occipital lobe with activations; rectangle map with activations cut out and mapped to holes;
annulus map of occipital lobe with one activation cut out.
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128 S125
Fig. 5. Cortical surfaces and rectangle maps for left ventral medial
prefrontal cortical region.
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128S126
EL(Q) = EL(R). A quadrilateral subregion Q of a surface with
designated boundary arcs as its ends can be mapped conformally
onto a unique (up to similarity) rectangle Q; hence, one defines
EL(Q) = EL(Q). Extremal length, large or small, means roughly
that the quadrilateral is long or wide, respectively, relative to the
designated ends.
These numerical extremal lengths may be treated as a form of
conformal bsizeQ due to bmonotonicityQ and bconvergenceQproperties and bquasi-invarianceQ. For instance, if B and A are
annuli and B is contained in A and separates its two boundaries,
then EL(B) b EL(A). If f:A Y B is j-quasi-conformal and A and
B are both annuli (or both quadrilaterals), thenEL Að ÞEL Bð Þ a 1=j; j½ .
The annular region A (see Fig. 4, lower right) is an example; this
bsizeQ of EL(A) has nothing to do with area, diameter, or how
contorted A is within the parent surface. It represents intrinsic
information that is not otherwise accessible, for example, visually
or intuitively. The same is true for the size of quadrilateral
regions in our examples.
Note that extremal lengths are numerical quantities readily
approximated using circle packing—independently of whether the
packings themselves are to be displayed or not. In processing
cerebral hemispheres, for example, one could routinely compute
(no visualization involved) EL(A), where A is the annular region,
say, between the occipital lobe and the corpus callosum; this is an
intrinsic measure of the bdistanceQ between these two regions.
Concerning visualization, Fig. 4 is intended to show the quasi-
stability of bshapeQ under conformal flat mappings. The appearance
of features that are deep in the interior of flattened images is largely
independent of the parent region or of the mode of flattening. There
are as yet only rudimentary methods for quantifying this, but it is
clear that the effect is at least visually helpful. The mathematics of
shape is an intense area of investigation in mathematics, though
largely in studies of visualization and recognition problems.
There are many other types of ECFs accessible via circle
packing. (These are amenable to PDE methods only in limited
special cases.) Extremal lengths associated with four-connected
regions (for example, discs with three holes as shown in Fig. 4,
lower right) are sets of numbers; likewise extremal lengths for point
distributions (discs with, say, n distinguished points), extremal
lengths of regions with slits, etc. Undoubtedly, experience, now that
we have computable flat maps, will suggest other intrinsic
quantities. Related conformal concepts such as capacities, harmonic
measures, and more general extremal lengths are candidates.
Conformal mapping and its applications in neuroscience
Our circle packing methods have been applied to a number of
different cortical regions. With Dr. Michael Miller’s laboratory (see
http://www.cis.jhu.edu), we are using conformal flat mapping to
look at small cortical regions including the plenum temporal
(Ratnanather et al., 2003), the medial prefrontal cortex (Hurdal et
al., 2003), as well as the superior temporal gyrus and the occipital
lobe. Focusing on smaller regions of cortex enables us to use
cortical flat mapping to identify and demarcate specific regions of
interest, and also impose local coordinate systems, while reducing
the larger distortions that result from flattening a cortical hemi-
sphere. In some of these studies, we are exploiting the ECF
extremal length metrics to quantify anatomical differences across
subjects. Fig. 5 illustrates sample results from the ventral medial
prefrontal cortex. The 3D surface representing the ventral medial
prefrontal cortex from young adult female twins is shown along
with rectangular euclidean flat maps. The surfaces are colored
according to mean curvature.
In collaboration with Dr. David Rottenberg’s group (see http://
www.neurovia.umn.edu), we are using conformal flattening to
localize functional activity on the cerebellum, as well as the
cortical hemispheres (Hurdal et al., 1999; Ju et al., 2004).
There is, in fact, one well-known example of conformal
mapping in neuroscience in which the angle-preserving property
is the target. This example is the retinotopic mapping between the
visual cortex and the retina p visual field. A number of papers in
the literature approximate the retinotopic mapping in monkeys and
humans by a conformal map and suggest mapping functions to
model experimental observations (see for example Drasdo, 1991;
Ermentrout, 1984; Horton and Hoyt, 1991; Hurdal, 1998;
Schwartz, 1977; Sereno et al., 1995).
Summary
Conformal mapping as applied to cortical data of the human
brain is in its infancy. However, the extensive mathematical history
and theory of conformal mapping offers many potential features
that can be exploited in a neuroscientific context. A unique
advantage of conformal methods over metric-based approaches is
that conformal mappings are canonical and hence mathematically
unique. In addition, by the RMT, conformal mappings exist and
they can be computed in the euclidean, hyperbolic, and spherical
geometries. Perhaps what is most exciting, and potentially useful,
are the ensemble conformal features. Conformal invariants, such as
the extremal length of an annulus or a rectangle, can provide new
bsizeQ and bshapeQ metrics for comparing cortical features across
subjects. Currently, only the circle packing approach allows these
kinds of conformal metrics to be computed and may provide new
ways of analyzing functional and structural data of the human
brain.
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128 S127
Software
The software TopoCV is available from http://www.math.
fsu.edu/~mhurdal (Hurdal, 2004a) for checking and correcting
surfaces with topological errors. It can be read in and output
surfaces in a variety of file formats (including byu, obj, vtk,
CARET, CirclePack and FreeSurfer formats) and can be used
to convert existing data formats into the format used by
CirclePack. The software CirclePack is available from
http://www.math.fsu.edu/~kens (Stephenson, 1992–2004) for
computing discrete conformal maps via circle packings. TopoCV
was used to process the data used in this paper and
CirclePack was used to generate the flat map images in
this paper.
Acknowledgments
This work has been supported in part by an NSF Focused
Research Group grant DMS-0101329 and NIH grant P20
EB02013. The authors would like to thank David Rottenberg,
Kelly Rehm, and Lili Ju (U. Minnesota) for providing the MRI data
used in Figs. 2 and 3, and to Michael Miller (Johns Hopkins U.)
and Kelly Botteron (Washington U. School of Medicine) for
providing the MRI data in Fig. 5 (under NIH grants R01
MH62626-01 and P41-RR15241). In addition, they would like to
thank Charles Collins (U. Tennessee, Knoxville) for assistance
with Fig. 3.
References
Angenent, S., Haker, S., et al., 1999. On the Laplace–Beltrami operator and
brain surface flattening. IEEE Trans. Med. Imag. 18, 700–711.
Beardon, A.F., Stephenson, K., 1990. The uniformization theorem for circle
packings. Indiana Univ. Math. J. 39, 1383–1425.
Bowers, P.L., Hurdal, M.K., 2003. Planar conformal mappings of piecewise
flat surfaces. In: Hege, H.-C., Polthier, K. (Eds.), Visualization and
Mathematics III. Springer Verlag, Berlin, pp. 3–34.
Collins, C., Stephenson, K., 2003. A circle packing algorithm. Comput.
Geom. 25, 21–34.
Drasdo, N., 1991. Neural substrates and threshold gradients of peripheral
vision. In: Kulikowski, J.J., Walsh, V., et al., (Eds.), Limits of Vision,
Vision and Visual Dysfunction, vol. 5. Macmillan Press, Basingstoke,
pp. 251–265. General editor J.R. Cronly-Dillon.
Drury, H.A., Van Essen, D.C., et al., 1996. Computerized mappings of the
cerebral cortex: a multiresolution flattening method and a surface-based
coordinate system. J. Cogn. Neurosci. 8, 1–28.
Ermentrout, G.B., 1984. Mathematical models for the patterns of
premigrainous auras. In: Holden, A.V., Winlow, W. (Eds.), Neuro-
biology of Pain. Manchester Univ. Press, Manchester, pp. 267–283.
Fary, I., 1948. On straight lines representation of planar graphs. Acta Sci.
Math. Szeged. 11, 229–233.
Fischl, B., Sereno, M.I., et al., 1999. Cortical surface-based analysis II:
inflation, flattening, and a surface-based coordinate system. Neuro-
Image 9, 195–207.
Fischl, B., Sereno, M.I., et al., 1999. High-resolution intersubject averaging
and a coordinate system for the cortical surface. Hum. Brain Mapp. 8,
272–284.
Goebel, R., 2000. A fast automated method for flattening cortical surfaces.
NeuroImage 11, S680. (Part 2).
Gu, X., Yau, S., 2002. Computing conformal structures of surfaces.
Commun. Inform. Syst. 2, 121–146.
Gu, X., Wang, Y., et al., 2003. Genus zero surface conformal mapping and
its application to brain surface mapping. Inf. Process. Med. Imag.,
Ambleside, UK, 172–184.
Han, X., Xu, C., 2002. Topology correction in brain cortex segmentation
using a multiscale, graph-based algorithm. IEEE Trans. Med. Imag. 21
(2), 109–121.
Horton, J.C., Hoyt, W.F., 1991. The representation of the visual field in
human striate cortex: a revision of the classic holmes map. Arch.
Opthalmol. 109, 816–824.
Hurdal, M.K., 1998. Mathematical and computer modelling of the human
brain with reference to cortical magnification and dipole source
localisation in the visual cortex. PhD thesis, Queensland University of
Technology, Brisbane, Australia.
Hurdal,M.K., 2004a. TopoCVsoftware. http://www.math.fsu.edu/~mhurdal.
Hurdal, M.K., 2004b. Automated topological correction of cortical surfaces.
NeuroImage 22 (Supplement 1, CD-Rom Abstract WE-197).
Hurdal, M.K., Bowers, P.L., et al., 1999. Quasi-conformally flat
mapping the human cerebellum. In: Taylor, C., Colchester, A.
(Eds.), Medical Image Computing and Computer-Assisted Interven-
tion-MICCAI’99, Lecture Notes in Computer Science, vol. 1679.
Springer, Berlin, pp. 279–286.
Hurdal, M.K., Lee, A., et al., 2003. Investigating the medial prefrontal
cortex with cortical flat mappings. NeuroImage 22 (Supplement 1, CD-
Rom Abstract 856).
Ju, L., Stern, J., et al., 2004. Cortical surface flattening using least square
conformal mapping with minimal metric distortion. Proc. IEEE
Biomed. Imaging, 77–80.
Levy, B., Petitjean, S., et al., 2002. Least squares conformal maps for
automatic texture Atlas generation. Computer Graphics (Proceedings of
SIGGRAPH 02), Addison Wesley.
Lorensen, W.E., Cline, H., 1987. Marching cubes: a high resolution 3D
surface construction algorithm. Comput. Graph. 21, 163–169.
MacDonald, D., Kabsni, N., et al., 2000. Automated 3-D extraction of
inner and outer surfaces of cerebral cortex from MRI. NeuroImage 12,
340–355.
Mangin, J.-F., Frouin, V., et al., 1995. From 3D magnetic resonance
images to structural representations of the cortex topography
using topology preserving deformations. J. Math. Imaging Vis. 5,
297–318.
Massey, W.S., 1967. Algebraic Topology: An Introduction. Springer-
Verlag, New York.
Polya, G., 1968. Mathematical Discovery vol. 2. Wiley, New York.
Ratnanather, J.T., Barta, P.E., et al., 2003. Dynamic programming
generation of boundaries of local coordinatized submanifolds in the
neocortex: application to the planum temporale. NeuroImage 20,
359–377.
Riemann, B., 1876. Grundlagen fqr eine allgemeine theorie der functionen
einer ver7nderlichen complexen grfsse, inauguraldissertation,
gfttingen, 1851. Gesammelte mathematische Werke und wissenschaft-
licher Nachlass. 2nd ed, pp. 3–48. Teubner, Leipzig.
Rodin, B., Sullivan, D., 1987. The convergence of circle packings to the
Riemann mapping. J. Differ. Geom. 26, 349–360.
Schwartz, E.L., 1977. Spatial mapping in the primate sensory projection:
analytic structure and relevance to perception. Biol. Cybern. 25,
181–194.
Sereno, M.I., Dale, A.M., et al., 1995. Borders of multiple visual areas in
humans revealed by functional magnetic resonance imaging. Science
268, 889–893.
Stephenson, K., 1992–2004. CirclePack software. http://www.math.utk.
edu/~kens.
Stephenson, K., 1999. Approximation of conformal structures via circle
packing. In: Papamichael, N., Ruscheweyh, S. (Eds.), Computational
Methods and Function Theory 1997, Proceedings of the Third CMFT
Conference, World Scientific, vol. 11, pp. 551–582.
Stephenson, K., 2002. Circle packing and discrete analytic function theory.
In: Kqhnau, R. (Ed.), Handbook of Complex Analysis, Geometric
Function Theory, vol. 1. Elsevier, Amsterdam.
M.K. Hurdal, K. Stephenson / NeuroImage 23 (2004) S119–S128S128
Stephenson, K. Introduction to circle packing and the theory of discrete
analytic functions. Camb. Univ. Press. (in press).
Stern, J., Schaper, K., et al., 2001. Automatic surface extraction by discrete,
topologically-controlled, region growing. NeuroImage 13, S157.
Van Essen, D.C., Drury, H.A., 1997. Structural and functional analyses of
human cerebral cortex using a surface-based atlas. J. Neurosci. 17,
7079–7102.
Van Essen, D.C., Drury, H.A., et al., 1998. Functional and structural
mapping of human cerebral cortex: solutions are in the surfaces. Proc.
Natl. Acad. Sci. U. S. A. vol. 95, 788–795.
Wagner, K., 1936. Bemerkunge zum Vierfarbenproblem. Jahresber. Dtsch.
Math.-Ver. 46, 26–32.
Wandell, B.A., Chial, S., et al., 2000. Visualization and measurement of the
cortical surface. J. Cogn. Neurosci. 12, 739–752.