CHAPTER 13

Statistical Shape Analysis for Brain

Structures
Li Shen∗,† , Shan Cong∗,‡ , Mark Inlow∗
∗ Department of Radiology and Imaging Sciences, Center for Neuroimaging, Indiana University School of
Medicine, Indianapolis, IN, USA
† Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indianapolis,
IN, USA
‡ Department of Electrical and Computer Engineering, Indiana University–Purdue University Indianapolis,
Indianapolis, IN, USA

Contents
13.1 Introduction 351
13.2 Surface Modeling and Registration 353
13.2.1 SPHARM Surface Modeling 353
13.2.1.1 Spherical Parameterization 353
13.2.1.2 SPHARM Expansion 355
13.2.2 SPHARM Surface Normalization 357
13.2.2.1 SPHARM Normalization 357
13.2.2.2 Subfield-Guided Registration 359
13.3 Statistical Inference on the Surface 362
13.3.1 Surface Atlas and Signal Processing 363
13.3.2 General Linear Model and Random Field Theory 364
13.3.3 Statistical Parametric Mapping Distribution Analysis 364
13.4 An Example Application 369
13.5 Conclusions 372
Acknowledgments 373
References 373

13.1 INTRODUCTION
Recent advances in non-invasive scanning techniques have resulted in a prominent
growth of research into the analysis of high quality 3D brain images. One fundamental
problem in brain image analysis is identifying the morphological abnormalities of the
neuroanatomy that are associated with a particular disorder to aid diagnosis and treat-
ment. One approach is volumetric analysis (e.g., [1,2]), which is based on measuring the
volume of a brain structure. Its major advantage is the simplicity; however, many struc-
tural differences may be overlooked. A newer approach, shape analysis (e.g., [3–26]),
has the potential to provide valuable information beyond simple volume measurements.
For example, it may help identify where the volume change is located, or characterize
abnormalities in the absence of volume differences.

Statistical Shape and Deformation Analysis Copyright © 2017 Elsevier Ltd

DOI: 10.1016/B978-0-12-810493-4.00016-X All rights reserved. 351
352 Statistical Shape and Deformation Analysis

Over the past decades, statistical shape analysis [27,28] has emerged as a promis-
ing new field with applications to medicine, biology and other scientific domains. The
pioneers are Kendall [29–31] and Bookstein [32–34], and their methods mostly focus
on landmark data. Computational anatomy (CA) is a prominent shape analysis model
proposed in the area of brain imaging [35–39]. The approach is based on creating and
analyzing diffeomorphisms, which are smooth invertible mappings between geometric
objects. The model consists of three steps: (1) computing deformation maps of indi-
viduals from a template, (2) computing probability laws of anatomical variations using
deformations as shape descriptors, and (3) performing inferences for diseases and anoma-
lies. Similar to the CA model, statistical shape analysis in general can be divided into two
categories: (1) to establish a statistical shape model for one group of geometric objects
by characterizing the mean and variability of the population; and (2) to identify shape
changes between two groups of objects.
Here we concentrate on the second type of statistical shape analysis and its ap-
plication to the morphometric analysis of brain structures extracted from the magnetic
resonance imaging (MRI) scans. Among many MRI morphometric techniques in brain
imaging, voxel-based morphometry (VBM) [40–42] and surface-based morphometry
(SBM) [43–48] are two widely used methods. VBM aims to compare regional differ-
ences in relative tissue concentration or deformation and has been applied to many
neuroanatomical studies. SBM can be used to quantify the amount of gray matter by
estimating the cortical thickness [44,46], where it requires a preprocessing step of seg-
menting the inner and outer cortical surfaces and the distance between the two surfaces
is defined as the cortical thickness. SBM has also been used in studying brain structures
other than cortical surfaces, such as hippocampus [10,11,48], ventricles [49], thalamus,
globus pallidus, and putamen [50]. Besides VBM and SBM, there are several other
shape models used in biomedical imaging studies. Examples include landmarks [34,51],
deformation fields mapping a template image to individual images [7,8,35,52,36–39],
distance transforms [53,13], and medial axes [54,55,23].
In this chapter, we focus on the topic of surface-based morphometry (SBM) in brain
imaging. We present typical shape analysis methods for modeling and analyzing 3D sur-
face data. We use hippocampal morphometry in Alzheimer’s disease (AD) as a test bed
to demonstrate these methods. Our goal is to identify hippocampal shape abnormalities
associated with AD or mild cognitive impairment (MCI, a prodromal stage of AD) in
order to aid early diagnosis. We first present classic spherical harmonics (SPHARM)
methods for modeling and registering 3D hippocampal surfaces [56], and then discuss
advanced methods that take into consideration hippocampal subfield information [5].
After that, we describe techniques for shape analysis of registered surface models, in-
cluding traditional general linear models (GLMs) [44,57] as well as a newly developed
statistical parametric mapping distribution analysis (SPM-DA) method for performing
 Statistical Shape Analysis for Brain Structures 353

surface-based morphometry [6]. Finally, we use an example neuroimaging application

[6] to demonstrate the effectiveness of these techniques.

13.2 SURFACE MODELING AND REGISTRATION

13.2.1 SPHARM Surface Modeling
The spherical harmonics (SPHARM) technique [56] creates a parametric surface descrip-
tion using spherical harmonics [58] as basis functions. Spherical harmonics were first
used as surface representation for radial surfaces (r (θ, φ)) [53], and later extended to
more general shapes by describing a surface using three spherical functions [56]. At
present SPHARM is a powerful surface modeling approach for arbitrarily shaped but
simply-connected objects.
This section provides a short summary of the SPHARM surface modeling tech-
nique. An input object surface is assumed to be a voxel surface, which is a square
surface mesh converted from a voxel representation; see Fig. 13.1A for an example.
Two steps are involved in obtaining a SPHARM shape description for such a voxel
surface: (1) spherical parameterization, and (2) SPHARM expansion. Below we discuss
these two steps.

13.2.1.1 Spherical Parameterization

Spherical parameterization aims to create a continuous and uniform mapping from the
object surface to the surface of a unit sphere so that each vertex on the object surface is
assigned a pair of spherical coordinates (θ, φ). As a result of spherical parameterization,
the surface of the unit sphere becomes our parameter space. To match the definition
of spherical harmonics [58], the following convention for spherical coordinates (θ, φ)
is employed (see also Fig. 13.2): θ is taken as the polar (colatitudinal) coordinate with
θ ∈ [0, π], and φ as the azimuthal (longitudinal) coordinate with φ ∈ [0, 2π). Thus, the
north pole has θ = 0 and the south pole has θ = π .
In order to create shape descriptors that can be compared across different 3D surfaces,
we require an appropriate parameterization that has the following properties.
1. Bijective mapping: each vertex on the object surface must map to exactly one point
on the sphere, and the inverse map must also be one-to-one.
2. Area preservation: each unit area on the object surface should be assigned to the
same relative amount of area in parameter space.
3. Topology preservation: each square face on the object surface should map to a
spherical quadrilateral in parameter space.
4. Minimal angular distortion: the spherical mapping of each square face should be as
close to a “spherical surface square” as possible.
To achieve the above goals, we employ the spherical parameterization approach pro-
posed by Brechbühler et al. [56]. This approach can be applied to a voxel surface, since
354 Statistical Shape and Deformation Analysis

Figure 13.1 An example SPHARM reconstruction. (A) The voxel surface of a hippocampus. (B–E)
SPHARM reconstructions of the hippocampal surface using coefficients up to degrees 1, 5, 10 and
15, respectively. (F) Spherical parameterization of the hippocampal surface shown in (A). (G) A regular
spherical mesh grid used for SPHARM reconstructions shown in (B) and (C). (H) An icosahedral subdi-
vision used for SPHARM reconstructions shown in (D) and (E).

Figure 13.2 Rotational convention of spherical coordinates used in spherical harmonics. For the point
(θ, φ) on the unit sphere, θ is taken as the polar (colatitudinal) coordinate, and φ as the azimuthal
(longitudinal) coordinate.
 Statistical Shape Analysis for Brain Structures 355

it exploits the uniform quadrilateral structure of a square surface mesh. The approach
consists of two steps: (1) initialization, and (2) optimization.
Step 1. Initialization. An initial parameterization is formed by constructing a har-
monic map from the object surface to the parameter surface. For colatitude θ , two
poles are selected by finding two surface vertices with the maximum and minimum z
coordinates in the object space, respectively. Then, a Laplace equation (Eq. (13.1)) with
Dirichlet conditions (Eq. (13.2) and Eq. (13.3)) is solved for colatitude θ :

∇ 2θ = 0 (except at the poles) (13.1)

θnorth = 0 (13.2)
θsouth = π (13.3)

Given our case being discrete, we approximate Eq. (13.1) by assuming that each vertex’s
colatitude (except at the poles) equals the average of its neighbors’ colatitudes. Thus,
after assigning θnorth = 0 to the north pole and θsouth = π to the south pole, we form
a system of linear equations (one for each vertex) and obtain the solution by solving
this linear system. For longitude φ , the same approach can be employed except that
longitude is a cyclic parameter. To solve this problem, a “date line” is introduced. When
crossing the date line, longitude is added or subtracted by 2π , depending on the crossing
direction. After adjusting the linear system accordingly, the solution for longitude φ can
be obtained.
Step 2. Optimization. The initial parameterization is refined to obtain an area
preserving mapping. Brechbühler et al. [56] formulate this refinement process as a con-
strained optimization problem. They establish a few constraints for topology and area
preservation and formulate an objective function for minimizing angular distortion.
To solve this constrained optimization problem, an iterative procedure is developed
to perform the following two steps alternately: (1) satisfying the constraints using the
Newton–Raphson method [59], and (2) optimizing the objective function using a con-
jugate gradient method [59]. For more details about these steps, refer to [60,56].

13.2.1.2 SPHARM Expansion

Spherical harmonics are a natural and convenient choice of basis functions for repre-
senting any twice-differentiable spherical function [61,53,58]. They form the Fourier
basis on the sphere, including an infinite set of spherical functions that are continuous,
orthogonal, single-valued, and complete.
Using the notational convention for spherical coordinates described in Fig. 13.2,
spherical harmonics Ylm (θ, φ) of degree l and order m are defined as follows:

2l + 1 (l − m)! m
Ylm (θ, φ) = P (cos θ ) eimφ , (13.4)
4π (l + m)! l
356 Statistical Shape and Deformation Analysis

where Plm (cos θ ) are associated Legendre polynomials (with argument cos θ ), and l and
m are integers with −l ≤ m ≤ l. The associated Legendre polynomial Plm is defined by
the differential equation

(−1)m m d
l+m
Plm (x) = (1 − x2 ) 2 l+m (x2 − 1)l . (13.5)
2 l!
l dx

Any twice-differentiable spherical function r (θ, φ) can be represented by a linear

combination of spherical harmonics Ylm (θ, φ) as follows:

∞ 
 l
r (θ, φ) = aml Ylm (θ, φ), (13.6)
l=0 m=−l

where the coefficients aml are uniquely determined by (see [62])

 π  2π
aml = Ylm (θ, φ)∗ r (θ, φ) sin θ dφ dθ. (13.7)
0 0

Here Ylm (θ, φ)∗ is the complex conjugate of Ylm (θ, φ).
Spherical harmonics were first used for surface representation for radial or stellar
surfaces r (θ, φ) (e.g., [53,62]), where the radial function, r (θ, φ), encodes the distance of
surface points from a chosen origin. Brechbühler et al. [60,56] extended this spherical
harmonics expansion technique to more general shapes by representing a surface using
three spherical functions. This surface expansion technique has been referred to as the
SPHARM expansion in previous studies (e.g., [10,49]). The SPHARM expansion tech-
nique can be applied to arbitrarily shaped but simply-connected objects. It is suitable
for surface comparison and can deal with protrusions and intrusions. In this chapter, the
SPHARM expansion is employed to describe 3D closed surfaces.
The SPHARM expansion requires a spherical parameterization performed in ad-
vance, as described in Section 13.2.1.1. The spherical parameterization has the following
form:
⎛ ⎞
x(θ, φ)
⎜ ⎟
v(θ, φ) = ⎝ y(θ, φ) ⎠ , (13.8)
z(θ, φ)

where x(θ, φ), y(θ, φ), and z(θ, φ) are three functions defined on the sphere. Thus,
the object surface can be described via expanding these three spherical functions using
spherical harmonics.
 Statistical Shape Analysis for Brain Structures 357

The expansion takes the form:

⎛ ⎞ ⎛ ⎞
x(θ, φ) ∞ cm ∞ 
⎟   ⎜ xl 
l l
⎜ m ⎟ m
v(θ, φ) = ⎝ y(θ, φ) ⎠ = ⎝ yl ⎠ l
c Y (θ, φ) = cml Ylm (θ, φ), (13.9)
z(θ, φ) l=0 m=−l cmzl l=0 m=−l

where
⎛ ⎞
cm
⎜ xlm ⎟
cl = ⎝ cyl ⎠ .
m
(13.10)
czlm
The coefficients cml up to a user-specified degree can be estimated by solving three
sets of linear equations in a least squares fashion, given a pre-defined spherical param-
eterization. The object surface can be reconstructed using these coefficients, and using
more coefficients leads to a more detailed reconstruction. Fig. 13.1 provides an exam-
ple of an object surface (Fig. 13.1A) and its SPHARM reconstructions (Fig. 13.1B–E).
Thus, a set of coefficients actually form an object surface description. Note that the
degree one reconstruction is always an ellipsoid (Fig. 13.1B). Using more coefficients
in a SPHARM expansion yields a more accurate reconstruction (Fig. 13.1C–E).
In our brain imaging applications, a hippocampus is originally represented by a
voxel surface that may contain errors due to the voxel quantization and the limited
voxel resolution. Using the first few degrees of SPHARM coefficients to describe the
surface can help to smooth the object and reduce these errors, given that a hippocam-
pal surface is assumed to be relatively smooth. In Fig. 13.1, we use coefficients up to
degree 15, which can derive smooth but detailed reconstructions. Thus, for each ex-
pansion (x(θ, φ), y(θ, φ), or z(θ, φ)), there are 15
l=0
l
m=−l 1 = 256 coefficients. In total,
256 × 3 = 768 coefficients are extracted to describe a hippocampal surface.

13.2.2 SPHARM Surface Normalization

13.2.2.1 SPHARM Normalization
The previous section shows how the SPHARM expansion derives a set of SPHARM
coefficients that describe a 3D closed surface. In order to compare different surfaces,
these coefficients need to be normalized into a common reference system. We call this
step SPHARM normalization.
In this chapter, we are mainly concerned with shape information, i.e.,
all the geometrical information that remains when location, scale, and rotational effects are fil-
tered out from an object [27].
Thus, the goal of the SPHARM normalization is to create a shape descriptor for any
given surface by removing the effects of translation, rotation, and scaling. The shape
358 Statistical Shape and Deformation Analysis

descriptor is formed by a set of normalized SPHARM coefficients that are designed to

be comparable across different object surfaces.
The translation effect of a SPHARM model can be easily removed by ignoring the
degree 0 coefficient. This revision centers the SPHARM reconstruction at the origin.
The scaling effect can be removed by dividing all the coefficients by a scaling factor f .
For example, we can choose f so that the object volume is constant. Note that while the
object volume information must be removed for pure shape analysis, in general mor-
phometric analysis, the object volume can be treated as an additional feature and can
be combined together with shape features. Another strategy is to keep the size infor-
mation in the SPHARM model and perform morphometric analysis on the geometric
configuration of size and shape.
The rotation effect can be removed by first creating a surface correspondence and then
aligning the corresponding parts together in object space. This is achieved by aligning
the degree one reconstruction, which is always an ellipsoid and often called first order
ellipsoid (FOE). To establish the surface correspondence, the parameter net on the FOE (see
Fig. 13.3A) is rotated to a canonical position such that the north pole is at one end of the
longest main axis, and the crossing point of the zero meridian and the equator is at one
end of the shortest main axis (see Fig. 13.3B). Surface correspondence is then defined
by taking two points with the same parameter pair (θ, φ) on two different surfaces to
be a corresponding pair. To align the corresponding parts together in object space, each
surface is rotated in object space such that the main axes of its degree one ellipsoid
coincide with the coordinate axes, putting the shortest axis along x and longest along z
(see Fig. 13.3C).
The normalization technique described above works only if the FOE is a “real”
ellipsoid (i.e., one with three different main axes), such as our hippocampal data, but
not an ellipsoid of revolution or a sphere. In the latter case, higher degree coefficients
might need to be involved for normalization. In addition, even for a “real” ellipsoid, it
has rotational symmetry of order 2 with respect to each of its main axes. Therefore, in
practice, manual inspection is involved to avoid any excessive 180 degree rotation that
breaks the alignment.
After the above steps, a set of canonical coordinates (i.e., normalized coefficients)
can be used as a shape descriptor for each surface, and these shape descriptors are com-
parable across objects. Assuming the scaling effect is properly handled, the procedure
for removing translation and rotation effects, including establishing surface correspondence,
is also called registration. Fig. 13.3 shows an example registration procedure: (A) original
objects, (B) objects with aligned parameterization, and (C) objects registered in both
parameter and object spaces.
 Statistical Shape Analysis for Brain Structures 359

Figure 13.3 SPHARM registration using first order ellipsoids (FOEs). Each row shows one sample hip-
pocampus. Each of (A)–(C) shows the FOE on left and degree 15 reconstruction on right. Parameteri-
zation is indicated by the lines on the surface, including equator θ = π/2 and four longitudinal lines
φ = −π/2, 0, π/2, π . The north and south poles and the point (π/2, 0) are shown as dots.

13.2.2.2 Subfield-Guided Registration

The hippocampus is composed of multiple subfields [63], and many hippocampal studies
have indicated that subfields play an important role in brain functions [64,63,65]. There
is an increased interest in recent literature in examining the subfields of the hippocam-
pal formation using MRI [66–69,4,3,70,71]. However, the above SPHARM modeling
framework does not take into account this critical subfield information. In particular, it
is important to use the hippocampal subfield information to guide hippocampal surface
registration.
One possible approach is to define landmarks on subfield boundaries, and use these
landmarks to help surface registration via a landmark-guided approach as proposed in
[72]. However, it becomes a challenging problem to identify anatomically meaning-
ful landmarks that are comparable across surfaces. To overcome this limitation, we
have proposed in [5] a landmark-free method. Specifically, we form a spherical label
image by assigning subfield labels to the surface parameterization on the sphere. The
subfield-aware surface registration problem can then be solved using a spherical image
registration method to align subfield label information across surfaces. The spherical
image registration method used in this chapter is the spherical demons (SD) framework
proposed in [73].
360 Statistical Shape and Deformation Analysis

As shown in Fig. 13.4A, B, the hippocampal segmentation result contains sub-

field label information, which can be mapped onto both (A) the object surface and
(B) the parameter surface. We use spherical images containing these label values (e.g.,
Fig. 13.4B) to guide the following surface registration procedure by the SD method
proposed in [73]. Note that the major goal here is to establish better surface correspon-
dence by aligning all the subfields together in the spherical parameter space.
Let F be the spherical image template, and M be the individual spherical image to
be aligned to the template,  be the desired transformation to register M to F, and γ
be intermediate hidden transformation. The SD objective function is formulated as:

(γ ∗ ,  ∗ ) = argmin  −1 (F − M ◦ )2
γ ,
1 1 (13.11)
+ 2
dist(γ , ) + 2
Reg(γ )
σx σT

while:
dist(γ , ) = γ − 2 (13.12)

Reg(γ ) = ∇(γ − Id)2 (13.13)

Here Eq. (13.12) defines the geodesic distance between the hidden transformation
γ and the optimized transformation  . Eq. (13.13) characterizes the regularization
penalization on gradient magnitude of the displacement field γ − Id of the hidden
transformation γ . Our goal is to minimize not only the loss function defined by the
first term of Eq. (13.11) but also the above geodesic distance and the regularization
term. Parameter  is a diagonal matrix that specifies the feature variability of each ver-
tex on the surface. Parameters σx and σT describe the trade-off between the similarity
measure and the regularization of the objective function.
Following the spherical vector spline interpolation theory [74], the SD algorithm
proposed in [73] implements the optimization procedure in two steps: (1) The first
step is to formulate a nonlinear least-square problem and solve the first two terms in
Eq. (13.11) by Gauss–Newton optimization. (2) The second step is to solve the last
two terms in Eq. (13.11) using a single convolution of the displacement field  with a
smoothing kernel [75]. The details of the SD algorithm are available in [73].
We employ a multi-resolution strategy at different levels in our adaptation of the SD
algorithm. We pre-define four levels of icosahedral subdivisions meshes, including 2562
vertices at level 1, 10,242 vertices at level 2, 40,962 vertices at level 3, and 163,842 ver-
tices at level 4. We start from the level with lowest resolution and apply the SD algorithm
to register the individual to the template. At each level, we linearly interpolate both the
template and the registered individual to the next higher level, and apply the SD algo-
rithm again. We repeat this procedure until the highest level is reached. The complete
 Statistical Shape Analysis for Brain Structures 361

Figure 13.4 (A–E) Example result of spherical parameterization (A–C) and spherical demons (SD) regis-
tration (D–E). (A) Original object in Euclidean space. (B–C) Original spherical mapping and its unfolded
version on the 2D plane. (D–E) Spherical mapping registered to the template subfield parameterization
(shown in (J)–(K)) and its unfolded version on the 2D plane. (F–K) SD registration procedure for creat-
ing the template subfield parameterization. (F) Mean subfield image after spherical parameterization
and FOE alignment. (G–J) Mean subfield image after 1st–4th iterations during SD registration respec-
tively. (K) Spherical image in (J) unfolded to 2D space. The mean spherical image shown in (J)–(K) is
the converging result of SD method and is chosen to be our resulting template subfield parameteriza-
tion. Green, light green and yellow colors correspond to three subfield categories respectively: cornu
ammonis (CA, including CA1, CA2 and CA3), dentate gyrus (DG), and subiculum and miscellaneous
(SUB+MISC). (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this chapter.)

algorithmic details of applying the SD method to subfield-guided hippocampal surface

registration are available in [5].
We demonstrate the performance of the above subfield-guided hippocampal sur-
face registration using a sample of 12 healthy control (HC) participants recruited
at the Indiana Alzheimer’s Disease Center (IADC). MRI scans were acquired on a
Siemens MAGNETOM Prisma 3T MRI scanner. The scanning protocols include a
T1-weighted (MPRAGE) whole-brain scan and a T2-weighted (TSE) partial-brain scan
and an oblique coronal slice orientation (positioned orthogonally to the main axis of
the hippocampus). See [70,76] for similar imaging protocols to collect high resolution
hippocampal subfield data.
Hippocampal subfields were segmented by the Automatic Segmentation of Hip-
pocampal Subfields (ASHS) software [70]. Topology fix was performed on segmentation
results to ensure a spherical topology for each hippocampus. Subfields were assigned
to the corresponding locations on the surface of hippocampal results. In this work,
we include the following three subfield groups on the surface: “Cornu Ammonis
362 Statistical Shape and Deformation Analysis

(CA, including CA1–3)”, “Dentate Gyrus (DG)”, or “Subiculum + Miscellaneous

(SUB+MISC)”. See Fig. 13.4A for an example hippocampal surface mapped with sub-
field labels.
Each hippocampal surface was modeled using the SPHARM method described be-
fore, including spherical parameterization, SPHARM expansion, and SPHARM align-
ment using FOE. Fig. 13.4A–C shows the result of spherical parameterization with
initial FOE-based alignment for an example right hippocampus.
Note that the initial spherical parameterization of each hippocampus can be rep-
resented as a spherical image. Fig. 13.4B–C shows an example spherical image and its
unfolded version on the 2D plane. Based on the spherical images of 12 subjects, we
use the following procedure to create an average spherical image as our template sub-
field parameterization: (1) Calculate the average of all the individual spherical images.
(2) Register each individual spherical image to the average image. (3) Calculate the
average of registered individual spherical images. (4) Repeat Steps 2 and 3 until the
average image converges. The resulting average image is used as our template subfield
parameterization.
Fig. 13.4F–K shows the SD registration procedure for creating the template sub-
field parameterization for the right hippocampus. Fig. 13.4F shows the mean subfield
image after spherical parameterization and FOE alignment. Here, we can see that the
boundaries between three subfields are blurred, indicating that the initial alignment is
not optimized for subfield registration.
Fig. 13.4G–J shows the mean subfield image after 1st–4th iterations during the SD
registration respectively. Fig. 13.4K shows the spherical image in (J) unfolded to 2D
space. The mean spherical image shown in Fig. 13.4J–K is the converging result of SD
method and is chosen to be our resulting template subfield parameterization. Compared
with the initial alignment shown in Fig. 13.4F, this resulting template demonstrates an
improved alignment of three subfield regions.
Fig. 13.5 shows the root mean square distance (RMSD) at each iteration for each
subject. The mean RMSDs of 12 subjects are 0.49 and 0.52 respectively for left and
right hippocampi at the beginning. They reduce to 0.32 and 0.34 after first iteration,
and then keep reducing until reaching 0.18 and 0.20 at the convergence.

13.3 STATISTICAL INFERENCE ON THE SURFACE

In this section, we first show how to extract surface signals for statistical shape analysis.
After that, we describe two methods for performing statistical inference on the surface:
(1) One which uses the classical general linear model (GLM) coupled with random field
theory (RFT) for multiple comparison correction [44,57]. (2) The other which uses a
newly proposed statistical parametric mapping distribution analysis (SPM-DA) [6].
 Statistical Shape Analysis for Brain Structures 363

Figure 13.5 SD registration performance. Root mean square distances between each individual (la-
beled from 1 to 12) and the template (i.e., the mean subfield image) at each iteration are shown for
both left and right hippocampal.

13.3.1 Surface Atlas and Signal Processing

To facilitate statistical shape analysis, we need to put all the shape objects into a common
reference system and make them comparable. One common practice is to create a shape
atlas and register all the individual shapes to the atlas, where the atlas can be defined by
the average of all the aligned shape objects. In the SPHARM case, we can use the FOE
approach to align the first order ellipsoid (FOE) of each individual SPHARM model to
a canonical configuration in both parameter and object spaces, as shown in Fig. 13.3.
After that, the shape atlas is simply the average of all the aligned individual SPHARM
models. In the hippocampal case, if subfield information is available, we need to further
adjust the spherical parameterization of all the individual shapes and the atlas. Following
the procedure described in Section 13.2.2.2, the template parameterization becomes
the spherical parameterization of the atlas, and the parameterization of each individual
shape needs to be registered to this template parameterization.
After atlas generation and shape registration, all the SPHARM models are aligned
to the same reference system (i.e., the atlas). This facilitates the subsequent analysis on
the surface. In order to perform statistical shape analysis, we need to define signals or
variables on the surface. Let xt be the atlas. An individual shape x can be described by
its deformation field δ(x) = x − xt relative to the atlas xt . However, for each vertex,
three components (corresponding to x, y, z coordinates) in δ(x) are used to characterize
the local shape change. For simplicity, similar to several previous studies [6,4,77,78], in
this chapter, we focus only on the deformation component along the surface normal
direction to reduce the number of variables considered for each surface location. This
deformation component is defined as the surface signal for each shape.
To perform statistical analysis in a Euclidean domain, Gaussian kernel smoothing
is often applied as a typical step for increasing the signal-to-noise ratio (SNR). Since
364 Statistical Shape and Deformation Analysis

the surface geometry of a brain structure is non-Euclidean, we cannot directly apply

Gaussian kernel smoothing. Instead, we employ heat kernel smoothing, which gener-
alizes Gaussian kernel smoothing to arbitrary Riemannian manifolds [44]. Heat kernel
smoothing is implemented by constructing the kernel of a heat equation on manifolds
that is isotropic in the local conformal coordinates. By smoothing the data on the sur-
face, the SNR increases and consequently it becomes easier to localize shape changes.

13.3.2 General Linear Model and Random Field Theory

Let us start our discussion from an example neuroimaging study. In this study, we want
to examine whether there is hippocampal shape change in late mild cognitive impair-
ment (LMCI), a prodromal stage of Alzheimer’s disease. In the analysis, we also want
to remove the effect of certain covariates (e.g., age, gender). To achieve this goal, we
consider the following general linear model (GLM):

signal = β0 + β · group + β1 · age + β2 · gender + .

To make it general, let Ni,j denote the surface signal at vertex/location j (j =

1, 2, . . . , m) for subject i (i = 1, 2, . . . , n), and we can describe the GLM as follows:

Ni,j = β0,j + βj xi + β1,j z1,i + β2,j z2,i + · · · + βp,j zp,i + i ,j ,

(13.14)
i = 1, 2, . . . , n, j = 1, 2, . . . , m,
in which x is the predictor of interest (i.e., diagnostic group in our case) and
z1 , z2 , . . . , zp are nuisance covariates (e.g., age and gender in our case). At each vertex j,
we fit Eq. (13.14) and compute the t-statistic for testing H0 : βj = 0 versus H1 : βj = 0.
The resulting m t-statistics comprise the statistical parametric map (SPM) for analyzing
the relationship between x and the hippocampal surface signals.
Random field theory (RFT) [44,57] is a widely used method for multiple com-
parison correction in general SPM analyses, and can be adapted to our surface-based
morphometry study. Specifically, RFT first estimates the smoothness of the surface data,
uses the smoothness values to determine statistical thresholds that control the family
wise error rate (FWER), and then provides corrected p-values for the result. The details
about how to apply RFT to surface-based morphometry are available in [44].
SurfStat [79] is a Matlab toolbox that implements both GLM and RFT for statisti-
cal inference on surfaces. Fig. 13.6 shows an example SurfStat result, where statistical
shape differences are identified on the hippocampal surface between healthy controls
and LMCI patients after removing the age and gender effects.

13.3.3 Statistical Parametric Mapping Distribution Analysis

As shown in Fig. 13.6, hippocampal shape change in LMCI can be detected by RFT.
However, using the same strategy, we identified no significant difference between HC
 Statistical Shape Analysis for Brain Structures 365

Figure 13.6 Effects of healthy control (HC) versus late mild cognitive impairment (LMCI). The t-map (A)
and p-map (B) of the diagnostic effect (HC-LMCI) on surface signals after removing the effects of age
and gender.

and early mild cognitive impairment (EMCI) subjects [4]. To bridge this gap, we re-
cently developed a new and powerful SPM image analysis framework, Statistical Paramet-
ric Mapping (SPM) Distribution Analysis or SPM-DA [6]. Unlike RFT and permutation
methods which focus on peak amplitude, clusters of supra-threshold statistics, or combi-
nations of the two, SPM-DA detects relationships by analyzing the information provided
by the entire distribution of the statistics comprising the SPM (t-statistics in our case),
hence the phrase Distribution Analysis. By making greater use of the SPM information,
SPM-DA can potentially achieve greater power than RFT methods. There are vari-
ous ways to capture the SPM distribution information. Here we employ a simplified
version of Lindsey’s Method in which the distribution is estimated by first constructing
a frequency histogram and then analyzing the frequencies (bin counts) using Poisson
generalized linear models [80]. A key advantage of this method is that it converts den-
sity estimation into a regression problem [81]. In more detail, the resulting SPM-DA
method can be described and implemented by the following four steps.
Step 1. Statistical parametric map (SPM). As mentioned earlier, our surface signal
Ni,j at vertex i of subject j is analyzed using the following regression model

Ni,j = β0,j + βj xi + β1,j z1,i + β2,j z2,i + · · · + βp,j zp,i + i ,j ,

(13.15)
i = 1, 2, . . . , n, j = 1, 2, . . . , m,

in which x is the predictor of interest (e.g., group indicator) and z1 , z2 , . . . , zp are nui-
sance covariates. In this step, at each vertex j, we fit the above model and compute the
t-statistic for testing H0 : βj = 0 versus H1 : βj = 0. The resulting m t-statistics comprise
the SPM for analyzing the relationship between x and the surface signals.
Step 2. Frequency histogram of the SPM t-statistics. Note that the RFT approach
described above analyzes the SPM using either peak amplitude or cluster size statistics
366 Statistical Shape and Deformation Analysis

as implemented by SurfStat [79,57]. In contrast to this, our SPM-DA method captures

approximately all the information provided by the SPM t-statistics by estimating their
distribution with a frequency histogram. To facilitate analysis the histogram is con-
structed using nh bins of unequal width: the bin boundaries are chosen so that each bin
is equally likely under the null hypothesis. More specifically, we compute frequency Fk ,
k = 1, . . . , nh , where Fk is the number of SPM t-statistics in the bin interval [Qk , Qk+1 )
and the bin boundaries Ql , l = 1, . . . , nh + 1, partition the real line into nh intervals of
equal probability with respect to the tn−(p+2) distribution.
Step 3. Regression analysis of the frequency histogram. The histogram bin frequencies
are analyzed to detect departures from count uniformity using two regression models:

Fk = βu wu,k + k , k = 1, . . . , nh , (13.16)
Fk = βl wl,k + k , k = 1, . . . , nh , (13.17)

in which Fk denotes the frequency of the k-th of nh bins. In our analyses we used
nh = 12. This choice ensured that the bin frequencies for our data were large enough
so that the Fk , and thus the k , were approximately normally distributed by the Central
Limit Theorem. For the first model in Eq. (13.16), we let

wu = (0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5)

be our predictor. Thus, the coefficient βu will be positive when positive SPM statistics
(right-tail values) are enriched in the result. This indicates a positive relationship be-
tween surface signal values and the predictor of interest. Similarly, for the second model
in Eq. (13.17), we let
wl = (5, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0)
be our predictor. Thus, the coefficient βl will be positive when negative (left-tail) values
are enriched in the result. This indicates a negative relationship. Fig. 13.7 shows an ex-
ample of the Eq. (13.16) predictor data (i.e., the solid line) and the bin counts computed
from the actual and permuted hippocampal surface data (i.e., the “◦” and “+” values
respectively).
Step 4. SPM-DA permutation test statistic. To detect whether the surface shape is
related to the predictor of interest, we test the following composite hypotheses:

H0 : βu ≤ 0 and βl ≤ 0 versus H1 : βu > 0 or βl > 0.

Using Bonferroni, a p-value for testing these hypotheses is given by p = 2 min(pl , pu )

where pu is the p-value for testing H0 : βu ≤ 0 versus H1 : βu > 0 and pl is defined sim-
ilarly. For various reasons, in particular correlation between the bin counts Fk , pl and
pu cannot be computed via the usual regression t-tests. Therefore we compute pl and
 Statistical Shape Analysis for Brain Structures 367

Figure 13.7 Bin counts for HC vs. EMCI. Our linear model in Eq. (13.16) aims to use the values on the
solid line to predict the “+” values (for permuted data) or the “◦” values (for actual data). Note that the
significance of the group difference is driven mainly by the “◦” value on the 12th bin.

pu using permutation tests. As a result, we avoid the strict and often unmet assump-
tions of RFT methods and obtain a distribution-free method: the only requirement for
valid permutation inference is that the data are exchangeable [82]. In order to achieve
exchangeability in the presence of covariates we use the Smith (“orthogonalization”)
method in which the predictor of interest (here x) is orthogonalized with respect to all
covariates prior to permutation [83]. Thus, in more detail, we compute pl and pu as
follows:
a. Execute Steps 1–3 described above on the unpermuted data to obtain the least
squares estimates β̂u and β̂l .
b. Implement the Smith method by computing x̃ where x̃i = xi − (B̂0 + B̂1 z1,i +
B̂2 z2,i + · · · + B̂p zp,i ) and the B̂i are the least squares coefficient estimates for the
model

xi = B0 + B1 z1,i + B2 z2,i + · · · + Bp zp,i + i , i = 1, 2, . . . , n.

Note that x̃ is the zero mean orthogonalization of x with respect to the covari-
ates z1 , z2 , . . . , zp , i.e., the n-dimensional vector x̃ is perpendicular to the vectors
z1 , z2 , . . . , zp .
368 Statistical Shape and Deformation Analysis

c. Randomly permute the elements of x̃ N times to generate x̃∗k , k = 1, 2, . . . , N .

Execute Steps 1–3 described above on each x̃∗k to generate permutation coefficient
estimates β̂u∗,1 , . . . , β̂u∗,N and β̂l∗,1 , . . . , β̂l∗,N .
d. Compute the p-value pu for testing H0 : βu ≤ 0 versus H1 : βu > 0 (a positive rela-
tionship between the surface signals Ni,j and x) as follows:
• If the distribution of the β̂u∗ ’s is normal, compute the p-value using the
t-distribution,

β̂u − X̄
pu = P tN −1 ≥ √ ,
S 1 + 1/N

in which X̄ and S are the sample mean and sample standard deviation of the
√
β̂u∗ ’s. Note that S is increased by 1 + 1/N to account for using x̄ instead of a
constant in the numerator.
• If the distribution of the β̂u∗ ’s is nonnormal compute the p-value in the usual
manner,

pu = (# of β̂u∗ ’s ≥ β̂u )/N .

e. Similarly compute pl for testing H0 : βl ≤ 0 versus H1 : βl > 0 (a negative relationship

between the surface normals Ni,j and x).
Simulation studies. We perform simulation studies to compare the performance of
the proposed SPM-DA method with that of traditional RFT peak and cluster methods.
We first create a random data set of 72 subjects on a template surface with 652 vertices,
using the following model:

Si,j = β xi + i,j , i = 1, . . . , 72, j = 1, . . . , 126,

= i,j , i = 1, . . . , 72, j = 127, . . . , 652,

where Si,j represents the surface value at location j for subject i. We conduct two studies
involving simulated two-sample data with xi equal to −1 for i = 1, . . . , 36 and 1 for i =
37, . . . , 72. Corresponding to these x values and the model the signal extends across 126
contiguous locations and is constant with a magnitude determined by β . In both studies,
β takes on the values 0, 1/12, 1/6, and 1/3. In the first study, the random errors i,j
are independent normal (μ = 0, σ 2 = 1) pseudorandom numbers. In the second study,
the i,j are also independent normal (μ = 0, σ 2 = 1) but are smoothed prior to the signal
being added, where the smoothing is applied using the heat kernel smoothing method
[44]. The resulting data sets are analyzed using our SPM-DA method (implemented
in R [84]) and RFT peak and cluster statistics (implemented by SurfStat [79,57]). For
each combination of β and choice of unsmoothed or smoothed random errors, 100
simulated surface data sets are generated and analyzed.
 Statistical Shape Analysis for Brain Structures 369

Figure 13.8 Simulation results. The number of rejections (out of 100 runs) at α = 0.05 and 0.01 for the
SPM-DA, RFT Peak, and RFT Cluster methods.

The results of our simulation studies are shown in Fig. 13.8, which plots the number
of rejections (out of 100 runs) of the SPM-DA, RFT Peak, and RFT Cluster methods
for two significance levels: α = 0.05 and α = 0.01. For the null (signal strength = 0)
scenarios, all three methods have type I error rates at or below α . For all non-null
scenarios the SPM-DA method outperforms the RFT Cluster method, demonstrating
substantially greater power at all signal strengths. SPM-DA also outperforms the RFT
Peak method for all but the strongest signal case where both methods reject H0 on every
run. In particular, its power is at least eight times greater than RFT Peak for the weakest
signals.

13.4 AN EXAMPLE APPLICATION

In this section, we use a real world neuroimaging application to demonstrate the shape
modeling and statistical analysis methods described in this chapter. The hippocampus is
an important brain structure responsible for learning and memory and is widely studied
in Alzheimer’s Disease (AD). MRI-based hippocampal measures have been shown as
effective biomarkers for detecting the status of AD or LMCI [85–87,70]. In this ap-
plication [6], we demonstrate that our SPM-DA method, coupled with the SPHARM
modeling, can identify hippocampal shape differences in EMCI that are undetected by
standard RFT methods.
370 Statistical Shape and Deformation Analysis

The real data used in this study were downloaded from the ADNI database [88]. One
goal of ADNI has been to test whether a serial MRI, positron emission tomography,
other biological markers, and clinical and neuropsychological assessment can be com-
bined to measure the progression of MCI and early AD. For up-to-date information,
see www.adni-info.org.
We downloaded the relevant data consisting of 172 HC and 267 early MCI (EMCI)
participants, including their baseline 3T MRI scans and demographic and diagnostic
information. We employed FIRST [89], a surface registration and segmentation tool
developed as part of the FMRIB Software Library (FSL), to perform hippocampal seg-
mentation. We conducted topology fix on the segmented hippocampal volumes to make
sure that each hippocampal surface has the spherical topology. We used the SPHARM
method described earlier to model each surface and registered it to the atlas (the mean
of all the HC individuals) by aligning its first order ellipsoid. We computed the defor-
mations along the surface normal direction of the atlas at each surface location for each
individual shape and used these as our surface signals.
We evaluated the age and gender effects on the surface signals, using the following
two GLMs respectively:

signal = β0 + β · age +
signal = β0 + β · gender +

At each surface vertex/location, we fit either of the above models and computed the
t-statistic for testing H0 : β = 0 versus H1 : β = 0. The maps of the resulting t-statistics
are shown in Fig. 13.9A, C. The maps of the corresponding p-values corrected by RFT
are shown in Fig. 13.9B, D.
We evaluated the diagnostic effect (HC-EMCI) on surface signals, without and with
removing the effects of age and gender, using the following two GLMs respectively:

signal = β0 + β · group +
signal = β0 + β · group + β1 · age + β2 · gender +

At each surface vertex/location, we fit either of the above models and computed the
t-statistic for testing H0 : β = 0 versus H1 : β = 0. The maps of the resulting t-statistics
are shown in Fig. 13.10A, C. The map of the corresponding p-values corrected by RFT
for the first model is shown in Fig. 13.10B. The p-value map for the second model is
identical to that of the first model, and thus is not shown in Fig. 13.10.
We applied the SPM-DA method to the comparison of HC vs. EMCI after remov-
ing the effects of age and gender. Fig. 13.10D shows the SPM-DA bin value map, where
12 bins were used in the analysis. The p-value generated by SPM-DA is 0.009, indicat-
ing that the hippocampal shape difference between HC and MCI is statistically different
 Statistical Shape Analysis for Brain Structures 371

Figure 13.9 Effects of age and gender. (A–B) The t-map and p-map of the age effect on surface signals.
(C–D) The t-map and p-map of the gender effect on surface signals.

at level α = 0.01. The p-values generated by RFT peak and RFT cluster methods are
both > 0.05, and thus neither method detected significant hippocampal shape changes
in EMCI. It is encouraging that the SPM-DA method identified hippocampal shape
changes in EMCI which were not detected by standard RFT methods. This demon-
strates the promise of SPM-DA for early diagnosis of the prodromal stage of Alzheimer’s
disease.
The histogram patterns (i.e., bin counts) captured by SPM-DA are shown as ◦’s for
the actual data and +’s for an example of permuted data in Fig. 13.7. For the actual data,
it is obvious that there are trends toward an enrichment of SPM values in the upper tail
of the distribution, which matches the Eq. (13.16) predictor data (shown as the solid
line) better than the Eq. (13.17) predictor data (not shown). This pattern suggests there
is hippocampal atrophy in EMCI compared with HC.
In sum, we have used a real world neuroimaging study to demonstrate almost all
the shape modeling and statistical analysis methods described in this chapter, except
the subfield-guided registration method. Since subfield-guided registration is a newly
372 Statistical Shape and Deformation Analysis

Figure 13.10 Effects of HC-EMCI. (A–B) The t-map and p-map of the diagnostic effect (HC-EMCI) on
surface signals. (C) The t-map of the diagnostic effect (HC-EMCI) on surface signals after removing
the effects of age and gender. The corresponding p-map is identical to (B) and thus not shown here.
(D) The SPM-DA bin value map for the comparison of HC vs. EMCI after removing the effects of age and
gender. In t-maps (A, C), red/blue colors respectively indicate expansion/contraction in HC compared
with EMCI. (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this chapter.)

proposed method [5], we are currently working on applying it to the analyses of the
hippocampal data available in ADNI (used here) and other independent cohorts.

13.5 CONCLUSIONS
Statistical shape analysis is a fundamental topic in biomedical image computing, and
plays important roles in numerous applications in brain imaging. This chapter describes
a surface-based morphometry framework for modeling and analyzing 3D surface data
in brain imaging studies. These studies examine 3D brain structures, and aim to iden-
tify morphometric abnormalities associated with a particular condition in order to aid
diagnosis and treatment. We have presented a classic shape description method based on
spherical harmonics (SPHARM) for modeling arbitrarily shaped but simply connected
 Statistical Shape Analysis for Brain Structures 373

3D brain structures. We have discussed a traditional SPHARM-based shape registra-

tion method via aligning first order ellipsoids, as well as a newly developed registration
method that takes into account subfields on the surface. After that, we have described
two techniques for shape analysis of registered surface models. The first is the traditional
general linear model (GLM), coupled with random field theory (RFT) for multiple
comparison correction, to perform vertex-wise surface-based morphometry. The sec-
ond is a newly proposed and powerful analysis approach, Statistical Parametric Mapping
(SPM) Distribution Analysis, or SPM-DA, which provides a distribution-free alternative
to RFT methods. We have demonstrated these shape modeling and analysis methods
using a real world brain imaging application of identifying hippocampal shape changes
in early mild cognitive impairment. These methods successfully identified promising
shape-based imaging biomarkers for early detection of Alzheimer’s disease.

ACKNOWLEDGMENTS
This work was supported in part by NIH R01 LM011360, R01 EB022574, U01 AG024904, RC2
AG036535, R01 AG19771, P30 AG10133, R01 AG040770, UL1 TR001108, R01 AG042437, and
R01 AG046171; DOD W81XWH-14-2-0151, W81XWH-13-1-0259, and W81XWH-12-2-0012;
NCAA 14132004; and IUPUI RSFG.

REFERENCES
[1] A. Goldszal, C. Davatzikos, D.L. Pham, M.X.H. Yan, R.N. Bryan, S.M. Resnick, An image processing
system for qualitative and quantitative volumetric analysis of brain images, J. Comput. Assist. Tomogr.
22 (5) (1998) 827–837.
[2] Michael D. Nelson, Andrew J. Saykin, Laura A. Flashman, Henry J. Riordan, Hippocampal volume
reduction in schizophrenia as assessed by magnetic resonance imaging, Arch. Gen. Psychiatry 55 (1998)
433–440.
[3] Shan Cong, Maher Rizkalla, et al., Building a surface atlas of hippocampal subfields from MRI scans
using FreeSurfer, FIRST and SPHARM, in: 2014 IEEE 57th International Midwest Symposium on
Circuits and Systems (MWSCAS), IEEE, 2014, pp. 813–816.
[4] Shan Cong, Maher Rizkalla, Paul Salama, John West, Shannon Risacher, Andrew Saykin, Li Shen,
Surface-based morphometric analysis of hippocampal subfields in mild cognitive impairment and
Alzheimer’s disease, in: 2015 IEEE 58th International Midwest Symposium on Circuits and Systems
(MWSCAS), IEEE, 2015, pp. 1–4.
[5] Shan Cong, Maher Rizkalla, Paul Salama, Shannon L. Risacher, John D. West, Yu-chien Wu, Liana
Apostolova, Eileen Tallman, Andrew J. Saykin, Li Shen, ADNI, Building a surface atlas of hippocam-
pal subfields from high resolution T2-weighted MRI scans using landmark-free surface registration,
in: MWSCAS’16: The IEEE 58th International Midwest Symposium on Circuits and Systems, Abu
Dhabi, United Arab Emirates, 2016.
[6] Mark Inlow, Shan Cong, Shannon L. Risacher, John West, Maher Rizkalla, Paul Salama, Andrew
J. Saykin, Li Shen, A New Statistical Image Analysis Approach and Its Application to Hippocampal
Morphometry, Springer International Publishing, 2016, pp. 302–310.
[7] J.G. Csernansky, S. Joshi, L. Wang, J.W. Halleri, M. Gado, J.P. Miller, U. Grenander, M.I. Miller,
Hippocampal morphometry in schizophrenia by high dimensional brain mapping, Proc. Natl. Acad.
Sci. USA 95 (1998) 11406–11411.
374 Statistical Shape and Deformation Analysis

[8] J.G. Csernansky, L. Wang, D. Jones, D. Rastogi-Cruz, J.A. Posener, G. Heydebrand, J.P. Miller, M.I.
Miller, Hippocampal deformities in schizophrenia characterized by high dimensional brain mapping,
Am. J. Psychiatr. 159 (2002) 2000–2006.
[9] Rhodri H. Davies, Carole J. Twining, P. Daniel Allen, Timothy F. Cootes, Christopher J. Taylor, Shape
discrimination in the hippocampus using an MDL model, in: Christopher J. Taylor, J. Alison Noble
(Eds.), IPMI 2003: 18th International Conference on Information Processing in Medical Imaging,
Ambleside, UK, in: Lect. Notes Comput. Sci., vol. 2732, 2003, pp. 38–50.
[10] G. Gerig, M. Styner, Shape versus size: improved understanding of the morphology of brain structures,
in: Proc. MICCAI’2001: 4th International Conference on Medical Image Computing and Computer-
Assisted Intervention, Utrecht, The Netherlands, in: Lect. Notes Comput. Sci., vol. 2208, 2001,
pp. 24–32.
[11] G. Gerig, M. Styner, M. Chakos, J.A. Lieberman, Hippocampal shape alterations in schizophrenia:
results of a new methodology, in: 11th Biennial Winter Workshop on Schizophrenia, 2002.
[12] P. Golland, B. Fischl, M. Spiridon, N. Kanwisher, R.L. Buckner, M.E. Shenton, R. Kikinis, A. Dale,
W.E.L. Grimson, Discriminative analysis for image-based studies, in: Proc. of MICCAI’2002: 5th
International Conference on Medical Image Computing and Computer Assisted Intervention, Tokyo,
Japan, in: Lect. Notes Comput. Sci., vol. 2488, 2002, pp. 508–515.
[13] P. Golland, W.E.L. Grimson, M.E. Shenton, R. Kikinis, Small sample size learning for shape analysis of
anatomical structures, in: Proc. MICCAI’2000: 3rd International Conference on Medical Image Com-
puting and Computer-Assisted Intervention, Pittsburgh, Pennsylvania, USA, in: Lect. Notes Comput.
Sci., vol. 1935, 2000, pp. 72–82.
[14] P. Golland, W.E.L. Grimson, M.E. Shenton, R. Kikinis, Deformation analysis for shaped based classi-
fication, in: Proc. IPMI’2001: 17th International Conference on Information Processing and Medical
Imaging, in: Lect. Notes Comput. Sci., vol. 2082, 2001, pp. 517–530.
[15] A.J. Saykin, L.A. Flashman, T. McHugh, C. Pietras, T.W. McAllister, A.C. Mamourian, R. Vidaver,
L. Shen, J.C. Ford, L. Wang, F. Makedon, Principal components analysis of hippocampal shape in
schizophrenia, in: International Congress on Schizophrenia Research, Colorado Springs, Colorado,
USA, 2003.
[16] Li Shen, James Ford, Fillia Makedon, Laura Flashman, Andrew Saykin, Organization for human brain
mapping, surface-based morphometric analysis for hippocampal shape in schizophrenia, NeuroImage
19 (2) (2003) S1004.
[17] Li Shen, James Ford, Fillia Makedon, Andrew Saykin, Effective classification of 3D closed surfaces: ap-
plication to modeling neuroanatomical structures, in: International Conference on Computer Vision,
Pattern Recognition and Image Processing (CVPRIP) in conjunction with Seventh Joint Conference
on Information Sciences (JCIS), Association for Intelligent Machinery, Cary, North Carolina, USA,
2003, pp. 708–711.
[18] Li Shen, James Ford, Fillia Makedon, Andrew Saykin, Hippocampal shape analysis: surface-based
representation and classification, in: M. Sonka, J.M. Fitzpatrick (Eds.), Medical Imaging 2003: Image
Processing, San Diego, CA, USA, in: Proc. SPIE, vol. 5032, 2003, pp. 253–264.
[19] Li Shen, James Ford, Fillia Makedon, Andrew Saykin, A surface-based approach for classification of
3D neuroanatomic structures, Intell. Data Anal. 8 (5) (2004).
[20] Li Shen, James Ford, Fillia Makedon, Yuhang Wang, Tilmann Steinberg, Song Ye, A. Saykin, Mor-
phometric analysis of brain structures for improved discrimination, in: MICCAI: Medical Image
Computing and Computer Assisted Intervention, Montreal, Quebec, Canada, in: Lect. Notes Com-
put. Sci., vol. 2879, 2003, pp. 513–520.
[21] Li Shen, Fillia Makedon, Andrew Saykin, Shape-based discriminative analysis of combined bilateral
hippocampi using multiple object alignment, in: J.M. Fitzpatrick, M. Sonka (Eds.), Medical Imaging
2004: Image Processing, San Diego, CA, USA, in: Proc. SPIE, vol. 5370, 2004.
 Statistical Shape Analysis for Brain Structures 375

[22] M.E. Shenton, G. Gerig, R.W. McCarley, G. Szekely, R. Kikinis, Amygdala–hippocampal shape dif-
ferences in schizophrenia: the application of 3D shape models to volumetric MR data, Psychiatry Res.
Neuroimaging 115 (2002) 15–35.
[23] M. Styner, G. Gerig, J. Lieberman, D. Jones, D. Weinberger, Statistical shape analysis of neuroanatom-
ical structures based on medial models, Med. Image Anal. 7 (3) (2003) 207–220.
[24] M. Styner, G. Gerig, S. Pizer, S. Joshi, Automatic and robust computation of 3D medial models
incorporating object variability, Int. J. Comput. Vis. 55 (2/3) (2003) 107–122.
[25] M. Styner, J. Lieberman, G. Gerig, Boundary and medial shape analysis of the hippocampus in
schizophrenia, in: MICCAI 2003, Medical Image Computing and Computer Assisted Intervention,
Montreal, Quebec, Canada, 2003.
[26] S.J. Timoner, P. Golland, R. Kikinis, M.E. Shenton, W.E.L. Grimson, W.M. Wells III, Performance
issues in shape classification, in: Proc. MICCAI’2002: 5th International Conference on Medical Im-
age Computing and Computer-Assisted Intervention, Tokyo, Japan, in: Lect. Notes Comput. Sci.,
vol. 2488, 2002, pp. 355–362.
[27] I.L. Dryden, K.V. Mardia, Statistical Shape Analysis, John Wiley and Sons, New York, 1998.
[28] C.G. Small, The Statistical Theory of Shape, Springer, New York, 1996.
[29] D.G. Kendall, The diffusion of shape, Adv. Appl. Probab. 9 (1977) 428–430.
[30] D.G. Kendall, Shape manifolds, Procrustean metrics and complex projective spaces, Bull. Lond. Math.
Soc. 16 (1984) 81–121.
[31] David G. Kendall, A survey of the statistical theory of shape, Stat. Sci. 4 (2) (1989) 87–99.
[32] Fred L. Bookstein, The Measurement of Biological Shape and Shape Change, Lect. Notes Biomath.,
vol. 24, Springer-Verlag, Berlin, New York, ISBN 0387089128, 1978, viii+191 pp.
[33] Fred L. Bookstein, Morphometric Tools for Landmark Data: Geometry and Biology, Cambridge
University Press, Cambridge, New York, 1991, xvii+435 pp.
[34] F.L. Bookstein, Shape and the information in medical images: a decade of the morphometric synthesis,
Comput. Vis. Image Underst. 66 (2) (1997) 97–118.
[35] Ulf Grenander, Michael I. Miller, Computational anatomy: an emerging discipline, Q. Appl. Math.
LVI (4) (1998) 617–694.
[36] Michael I. Miller, Computational anatomy: shape, growth, and atrophy comparison via diffeomor-
phisms, NeuroImage 23 (Suppl. 1) (2004) S19–S33, http://www.sciencedirect.com/science/article/
B6WNP-4DD95D7-2/2/1464f473b6e671929e545888cb6ea8ba.
[37] Michael I. Miller, M. Faisal Beg, Can Ceritoglu, Craig Stark, Increasing the power of functional
maps of the medial temporal lobe by using large deformation diffeomorphic metric mapping, Proc.
Natl. Acad. Sci. USA 102 (27) (2005) 9685–9690, http://dx.doi.org/10.1073/pnas.0503892102,
http://www.pnas.org/cgi/content/abstract/102/27/9685.
[38] Anqi Qiu, Benjamin J. Rosenau, Adam S. Greenberg, Monica K. Hurdal, Patrick Barta, Steven
Yantis, Michael I. Miller, Estimating linear cortical magnification in human primary visual cor-
tex via dynamic programming, NeuroImage 31 (1) (2006) 125–138, http://www.sciencedirect.com/
science/article/B6WNP-4J7300B-1/2/8b52c730f16dcfa477de2cf91d4e4c68.
[39] Lei Wang, J. Philip Miller, Mokhtar H. Gado, Daniel W. McKeel, Marcus Rother-
mich, Michael I. Miller, John C. Morris, John G. Csernansky, Abnormalities of hip-
pocampal surface structure in very mild dementia of the Alzheimer type, NeuroImage
30 (1) (2006) 52–60, http://www.sciencedirect.com/science/article/B6WNP-4HCMSHW-3/2/
0ac679d0a9d5b6c5f6f94c26ff0f1b14.
[40] J. Ashburner, K.F. Friston, Voxel-based morphometry – the methods, NeuroImage 11 (2000) 805–821.
[41] J. Ashburner, K. Friston, Why voxel-based morphometry should be used, NeuroImage 14 (2001)
1238–1243.
[42] C. Davatzikos, A. Genc, D. Xu, S.M. Resnick, Voxel-based morphometry using the ravens maps:
methods and validation using simulated longitudinal atrophy, NeuroImage 14 (2001) 1361–1369.
376 Statistical Shape and Deformation Analysis

[43] M.K. Chung, K.M. Dalton, L. Shen, et al., Weighted Fourier series representation and its application
to quantifying the amount of gray matter, IEEE Trans. Med. Imaging 26 (4) (2007) 566–581.
[44] M.K. Chung, S. Robbins, K.M. Dalton, R.J. Davidson, A.L. Alexander, A.C. Evans, Cortical thick-
ness analysis in autism via heat kernel smoothing, NeuroImage 25 (2005) 1256–1265.
[45] M.K. Chung, K.J. Worsley, S. Robbins, T. Paus, J. Taylor, J.N. Giedd, J.L. Rapoport, A.C. Evans,
Deformation-based surface morphometry applied to gray matter deformation, NeuroImage 18 (2003)
198–213.
[46] B. Fischl, A.M. Dale, Measuring the thickness of the human cerebral cortex from magnetic resonance
images, Proc. Natl. Acad. Sci. USA 97 (2000) 11050–11055.
[47] M.I. Miller, A.B. Massie, J.T. Ratnanather, K.N. Botteron, J.G. Csernansky, Bayesian construction of
geometrically based cortical thickness metrics, NeuroImage 12 (2000) 676–687.
[48] Li Shen, James Ford, et al., A surface-based approach for classification of 3D neuroanatomic structures,
Intell. Data Anal. 8 (6) (2004) 519–542.
[49] G. Gerig, M. Styner, D. Jones, D. Weinberger, J. Lieberman, Shape analysis of brain ventricles using
SPHARM, in: Proc. IEEE Workshop on Mathematical Methods in Biomedical Image Analysis, 2001,
pp. 171–178.
[50] A. Kelemen, G. Szekely, G. Gerig, Elastic model-based segmentation of 3-D neuroradiological data
sets, IEEE Trans. Med. Imaging 18 (10) (1999) 828–839.
[51] T.F. Cootes, C.J. Taylor, D.H. Cooper, J. Graham, Active shape models – their training and application,
Comput. Vis. Image Underst. 61 (1995) 38–59.
[52] S.C. Joshi, M.I. Miller, U. Grenander, On the geometry and shape of brain sub-manifolds, Int. J. Pat-
tern Recognit. Artif. Intell. 11 (8) (1997) 1317–1343 (Special Issue on Magnetic Resonance Imaging).
[53] D.H. Ballard, C.M. Brown, Computer Vision, Prentice-Hall, NJ, 1982.
[54] S.M. Pizer, D.S. Fritsch, P. Yushkevich, V. Johnson, E. Chaney, Segmentation, registration, and mea-
surement of shape variation via image object shape, IEEE Trans. Med. Imaging 18 (10) (1999)
851–865.
[55] M. Styner, G. Gerig, S. Pizer, S. Joshi, Automatic and robust computation of 3D medical models
incorporating object variability, Int. J. Comput. Vis. 55 (2–3) (2003) 107–122.
[56] Ch. Brechbühler, G. Gerig, O. Kubler, Parametrization of closed surfaces for 3D shape description,
Comput. Vis. Image Underst. 61 (2) (1995) 154–170.
[57] K.J. Worsley, S. Marrett, P. Neelin, A.C. Vandal, K.J. Friston, A.C. Evans, A unified statistical approach
for determining significant signals in images of cerebral activation, Hum. Brain Mapp. 4 (1996) 58–73.
[58] Eric W. Weisstein, Spherical harmonic, in: MathWorld – A Wolfram Web Resource, http://
mathworld.wolfram.com/SphericalHarmonic.html.
[59] William H. Press, Numerical Recipes in C: The Art of Scientific Computing, second ed., Cambridge
University Press, Cambridge, New York, 1992.
[60] Ch. Brechbühler, Description and Analysis of 3-D Shapes by Parametrization of Closed Surfaces,
Ph.D. thesis, IKT/BIWI, ETH, Zurich, 1995.
[61] George B. Arfken, Mathematical Methods for Physicists, third ed., Academic Press, Orlando, 1985.
[62] David Ritchie, Graham Kemp, Fast computation, rotation, and comparison of low resolution spherical
harmonic molecular surfaces, J. Comput. Chem. 20 (1999) 383–395.
[63] Susanne G. Mueller, L.L. Chao, et al., Evidence for functional specialization of hippocampal subfields
detected by MR subfield volumetry on high resolution images at 4T, NeuroImage 56 (3) (2011)
851–857.
[64] Thorsten Bartsch, Juliane Döhring, et al., CA1 neurons in the human hippocampus are critical for
autobiographical memory, mental time travel, and autonoetic consciousness, Proc. Natl. Acad. Sci.
USA 108 (42) (2011) 17562–17567.
[65] Marcus Rössler, Rosemarie Zarski, Jürgen Bohl, Thomas G. Ohm, Stage-dependent and sector-
specific neuronal loss in hippocampus during Alzheimer’s disease, Acta Neuropathol. 103 (4) (2002)
363–369.
 Statistical Shape Analysis for Brain Structures 377

[66] Julie Winterburn, Jens C. Pruessner, et al., High-resolution in vivo manual segmentation protocol for
human hippocampal subfields using 3T magnetic resonance imaging, J. Vis. Exp. 105 (2015) e51861.
[67] Paul A. Yushkevich, Robert S.C. Amaral, et al., Quantitative comparison of 21 protocols for label-
ing hippocampal subfields and parahippocampal subregions in in vivo MRI: towards a harmonized
segmentation protocol, NeuroImage 111 (2015) 526–541.
[68] Bernd Merkel, Christopher Steward, et al., Semi-automated hippocampal segmentation in people with
cognitive impairment using an age appropriate template for registration, J. Magn. Reson. Imaging
42 (6) (2015) 1631–1638.
[69] Michael R. Hunsaker, David G. Amaral, A semi-automated pipeline for the segmentation of rhesus
macaque hippocampus: validation across a wide age range, PLoS ONE 9 (2) (2014) e89456.
[70] Paul A. Yushkevich, John B. Pluta, et al., Automated volumetry and regional thickness analysis of
hippocampal subfields and medial temporal cortical structures in mild cognitive impairment, Hum.
Brain Mapp. 36 (1) (2015) 258–287.
[71] Juan Eugenio Iglesias, Jean C. Augustinack, et al., A computational atlas of the hippocampal forma-
tion using ex vivo, ultra-high resolution MRI: application to adaptive segmentation of in vivo MRI,
NeuroImage 115 (2015) 117–137.
[72] Li Shen, Sungeun Kim, Andrew J. Saykin, Fourier method for large-scale surface modeling and reg-
istration, Comput. Graph. 33 (3) (2009) 299–311.
[73] B.T. Thomas Yeo, Mert R. Sabuncu, et al., Spherical demons: fast diffeomorphic landmark-free surface
registration, IEEE Trans. Med. Imaging 29 (3) (2010) 650–668.
[74] Joan Glaunès, Marc Vaillant, et al., Landmark matching via large deformation diffeomorphisms on the
sphere, J. Math. Imaging Vis. 20 (1–2) (2004) 179–200.
[75] Pascal Cachier, Eric Bardinet, et al., Iconic feature based nonrigid registration: the PASHA algorithm,
Comput. Vis. Image Underst. 89 (2) (2003) 272–298.
[76] Susanne G. Mueller, Norbert Schuff, et al., Hippocampal atrophy patterns in mild cognitive impair-
ment and Alzheimer’s disease, Hum. Brain Mapp. 31 (9) (2010) 1339–1347.
[77] J. Wan, L. Shen, K.E. Sheehan, et al., Shape analysis of thalamic atrophy in multiple sclerosis,
in: MIAMS 2009: MICCAI Workshop on Medical Image Analysis on Multiple Sclerosis, 2009,
pp. 93–104.
[78] Li Shen, Andrew J. Saykin, et al., Morphometric analysis of hippocampal shape in mild cognitive im-
pairment: an imaging genetics study, in: IEEE 7th Int. Symp. on Bioinformatics and Bioengineering,
Boston, MA, 2007, pp. 211–217.
[79] K.J. Worsley, SurfStat, http://www.math.mcgill.ca/keith/surfstat.
[80] J.K. Lindsey, Construction and comparison of statistical models, J. R. Stat. Soc. B 36 (1974) 418–425.
[81] Bradley Efron, Large-Scale Inference: Empirical Bayes Methods for Estimation, Testing, and Pre-
diction, Institute of Mathematical Statistics Monographs, Cambridge University Press, Cambridge,
ISBN 9780511761362, 2010, xii+263 pp.
[82] Thomas E. Nichols, Andrew P. Holmes, A unified statistical approach for determining significant
signals in images of cerebral activation, Hum. Brain Mapp. 15 (2001) 1–25.
[83] A.M. Winkler, G.R. Ridgway, M.A. Webster, S.M. Smith, T.E. Nichols, Permutation inference for
the general linear model, NeuroImage 92 (2014) 381–397.
[84] R Core Team, R: A Language and Environment for Statistical Computing, R Foundation for Statistical
Computing, Vienna, Austria, 2014, http://www.R-project.org/.
[85] John Pluta, Paul Yushkevich, et al., In vivo analysis of hippocampal subfield atrophy in mild cognitive
impairment via semi-automatic segmentation of T2-weighted MRI, J. Alzheimer’s Dis. 31 (1) (2012)
85–99.
[86] Li Shen, Andrew J. Saykin, Sungeun Kim, Hiram A. Firpi, John D. West, Shannon L. Risacher, Brenna
C. McDonald, Tara L. McHugh, Heather A. Wishart, Laura A. Flashman, Comparison of manual and
automated determination of hippocampal volumes in MCI and early AD, Brain Imaging Behav. 4 (1)
(2010) 86–95.
378 Statistical Shape and Deformation Analysis

[87] C. Testa, M.P. Laakso, et al., A comparison between the accuracy of voxel-based morphometry and
hippocampal volumetry in Alzheimer’s disease, J. Magn. Reson. Imaging 19 (3) (2004) 274–282,
http://dx.doi.org/10.1002/jmri.20001, http://www.ncbi.nlm.nih.gov/pubmed/14994294.
[88] M.W. Weiner, D.P. Veitch, et al., The Alzheimer’s Disease Neuroimaging Initiative: a review of pa-
pers published since its inception, Alzheimer’s Dement. 9 (5) (2013) e111–e194, http://dx.doi.org/
10.1016/j.jalz.2013.05.1769, http://www.ncbi.nlm.nih.gov/pubmed/23932184.
[89] Brian Patenaude, Stephen M. Smith, David N. Kennedy, Mark Jenkinson, A Bayesian model of shape
and appearance for subcortical brain segmentation, NeuroImage 56 (3) (2011) 907–922.