Page 1

MULTIVARIATE TENSOR-BASED MORPHOMETRY ON SURFACES: APPLICATION TO

MAPPING VENTRICULAR CHANGES IN HIV/AIDS

Yalin Wang1,2, Jie Zhang3, Tony F. Chan1, Arthur W. Toga2, Paul M. Thompson2

1Mathematics Department, UCLA

2Lab. of Neuro Imaging and Brain Research Institute, UCLA School of Medicine

3Statistics Department, University of Wisconsin-Madison

{ylwang,chan}@math.ucla.edu, jiezhang@stat.wisc.edu, {arthur.toga, thompson}@loni.ucla.edu

ABSTRACT

We apply multivariate tensor-based morphometry to study

lateral ventricular surface abnormalities associated with

HIV/AIDS. We use holomorphic one-forms to obtain a con-

formal parameterization of ventricular geometry, and to reg-

ister lateral ventricular surfaces across subjects. In a new

development, we computed new statistics on the Riemannian

surface metric tensors that encode the full information in the

deformation tensor fields. We applied this framework to 3D

brain MRI data, to map the profile of lateral ventricular sur-

face abnormalities in HIV/AIDS (11 subjects). Experimental

results demonstrated that our method powerfully detected

brain surface abnormalities.

statistics on the local Riemannian metric tensors, computed

in a log-Euclidean framework, detected group differences

with greater power than other surface-based statistics includ-

ing the Jacobian determinant, largest and least eigenvalue,

or the pair of eigenvalues of the Jacobian matrix. Compu-

tational anatomy studies may therefore benefit from surface

parameterization using differential forms and tensor-based

morphometry, in the log-Euclidean domain, on the resulting

surface tensors.

IndexTerms— MultivariateTensor-BasedMorphometry,

Holomorphic One-Form, Surface Modeling

Multivariate Hotelling’s T2

1. INTRODUCTION

Deformation-based morphometry (DBM) [1, 2, 3, 4] uses de-

formations obtained from the nonlinear registration of brain

images to infer local differences in brain volume or shape.

Tensor-based morphometry (TBM) [5, 6, 7] tends to examine

high-order spatial derivatives of the deformation maps regis-

tering brains to common template, constructing morpholog-

ical tensor maps such as the Jacobian determinant, torsion

This work was funded by the National Institutes of Health through the

NIH Roadmap for Medical Research, Grant U54 RR021813 entitled Center

for Computational Biology (CCB). The work was performed while the third

author was on leave at the National Science Foundation as Assistant Director

of the Directorate for Mathematics & Physical Sciences.

or vorticity. One advantage of TBM is that it derives local

derivatives and tensors from the deformation for further anal-

ysis, and statistical maps can be made to localize regions with

significant group differences or changes over time. In this pa-

per, we pursue the notion that TBM can even be applied to

surface models, making use of the Riemannian surface metric

to characterize local surface changes.

The lateral ventricles - fluid-filled structures deep in the

brain - are often enlarged in disease and can provide sensitive

measures of disease progression [8, 9, 10, 11]. Ventricular

changesreflect atrophy in surrounding structures, and ventric-

ular measures and surface-based maps can provide sensitive

assessments of tissue reduction that correlate with cognitive

deterioration in illnesses.

In this paper, we propose a new multivariate TBM frame-

work for surface morphometry, using differential forms as

the basis for the tensors that are analyzed. In an empirical

study of brain abnormalities in HIV/AIDS, we studied lat-

eral ventricular surface deformation. We found that the pro-

posed multivariate TBM detected areas of statistically signifi-

cant deformation even in a relatively small test dataset - from

11 subjects with HIV/AIDS and 8 matched healthy controls.

For comparison, we also applied other four TBM statistics to

the same dataset. The proposed multivariate TBM proved to

have more detection power by detecting consistent but more

statistically significant areas of surface deformation. This

works complements ongoing work by others, extracting mod-

els of ventricular surfaces via fluid registration, and detecting

group differences using either M-reps or support vector ma-

chines [9, 10, 11].

2. SURFACE CONFORMAL PARAMETERIZATION

WITH HOLOMORPHIC ONE FORMS

To register lateral ventricular surfaces across subjects, we ap-

plied a canonical conformal parameterization method. Holo-

morphic one-forms, a structure used in differential geometry,

can be used to generate canonical conformal parametrization

on a set of simply connected 3D surfaces [12], maximizing

Page 2

the uniformity of the induced grid over the entire domain.

A summary of canonical conformal parametrization now

follows, abbreviatedduetothepagelimit, (see[12]foramore

detailed algorithm description).

(1). Compute exact harmonic one-forms basis;

(2). Compute closed harmonic one-forms basis;

(3). Compute holomorphic one-forms basis;

(4). Compute the canonical holomorphic one-form and

the canonical conformal parameterization.

Even though the ventricular geometry consists of sev-

eral branches that are not readily parameterized using other

methods, the holomorphic one-forms can be used to induce

a smooth conformal grid on the long and narrow horned

surfaces of the lateral ventricles, with a differentiable param-

eterization at the surface junctions. Even though ventricular

shape varies widely in normal populations, the curved junc-

tions between the horns are consistently recovered using the

differential form method, allowing us to segment the ventric-

ular surface consistently and match each of them via their

conformal coordinates.

3. MULTIVARIATE TENSOR-BASED

MORPHOMETRY

3.1. Derivative Map

Suppose φ : S1→ S2is a map from the surface S1to the sur-

face S2. For simplicity, we use the isothermal coordinates of

both surfaces for the arguments. Let (u1,v1),(u2,v2) be the

isothermal coordinates of S1and S2respectively. The Rie-

mannian metric of Sican be represented as gi= e2λi(du2

dv2

In the local parameters, the map φ can be represented

as φ(u1,v1)=(φ1(u1,v1),φ2(u1,v1)). The derivative

map of φ is the linear map between the tangent spaces,

dφ : TM(p) → TM(φ(p)), induced by the map φ. In the

local parameter domain, the derivative map is the Jacobian of

φ,

?

∂u1

Let the position vector of points on S1 be r(u1,v1). De-

note the tangent vector fields as

Because (u1,v1) are isothermal coordinates,

only differ by a rotation of π/2. Therefore, we can con-

struct an orthonormal frame on the tangent plane on S1as

{e−λ1

thonormal frame on S2as {e−λ2

The derivative map under the orthonormal frames is rep-

resented by:

?

i+

i),i = 1,2.

dφ =

∂φ1

∂u1

∂φ2

∂φ1

∂v1

∂φ2

∂v1

?

.

∂

∂u1=

∂r

∂u1,

∂

∂v1=

∂

∂u1and

∂r

∂v1.

∂

∂v1

∂

∂u1,e−λ1 ∂

∂v1}.Similarly, we can construct an or-

∂

∂u2,e−λ2 ∂

∂v2}.

dφ = eλ2−λ1

∂φ1

∂u1

∂φ2

∂u1

∂φ1

∂v1

∂φ2

∂v1

?

.

In practice, smooth surfaces are approximated by tri-

angle meshes. The map φ is approximated by a simplicial

map, which maps vertices to vertices, edges to edges and

faces to faces. The derivative map dφ is approximated by

the linear map from one face [v1,v2,v3] to another one

[w1,w2,w3].First, we isometrically embed the triangle

[v1,v2,v3],[w1,w2,w3] onto the plane R2; the planar coor-

dinates of the vertices of vi,wjare denoted using the same

symbols vi,wj. Then we explicitly compute the linear matrix

for the derivative map dφ,

dφ = [w3− w1,w2− w1][v3− v1,v2− v1]−1.

(1)

3.2. Multivariate Tensor-Based Statistics

In our work, we use multivariate statistics on deformation

tensors [13] and adapt the concept to surface tensors. Let

J be the derivative map and define the deformation tensors

as S = (JTJ)1/2. Instead of analyzing shape differences

based on the eigenvalues of the deformation tensor, we con-

sider a new family of metrics, the “Log-Euclidean metrics”

[14]. These metrics make computations on tensors easier to

perform. Tensors are first transformed via a logarithmic trans-

formation to form a vector space, and statistical parameters

may then be computed easily using standard formulae for Eu-

clidean spaces [15].

We apply Hotelling’s T2test on sets of values in the log-

Euclidean space of the deformation tensors. We stack the log-

transformed tensor components into a vector, and given two

groups of n-dimensional vectors Si, i = 1,...,p, Tj, j =

1,...,q, we use the Mahalanobis distance M to measure the

group mean difference,

M = (log¯S − log¯T)Σ−1(log¯S − log¯T)

where¯S and¯T are the means of the two groups and Σ is the

combined covariance matrix of the two groups.

4. EXPERIMENTAL RESULTS

4.1. Automatic Lateral Ventricular Surface Registration

via Holomorphic One-Forms

The concave shape, complex branching topology and narrow-

ness of the inferior and posterior horns have made it diffi-

cult for surface parametrization approaches to impose a grid

on the entire structure without introducing significant area

distortion. To model the lateral ventricular surface, we au-

tomatically locate and introduce three cuts on each ventri-

cle. The cuts are motivated by examining the topology of

the lateral ventricles, in which several horns are joined to-

gether at the ventricular ”atrium” or ”trigone”. We call this

topological model, creating a set of connected surfaces, a

topology optimization operation. After modeling the topol-

ogy in this way, a lateral ventricular surface, in each hemi-

sphere, becomes an open boundary surface with 3 bound-

aries(Fig. 1(a)). We computed the exact harmonic one-form

Page 3

(Fig. 1(b)), its conjugate one-form (Fig. 1(c)), canonical holo-

morphic one-form (Fig. 1(d)). With the induced conformal

net (Fig. 1(e)), each lateral ventricular surface can be divided

into 3 connected pieces (Fig. 1(f)). Although surface geome-

try is widely variable across subjects, the zero point locations

of these differential forms are intrinsically determined by the

surface conformal structures, and the partitioning of the sur-

face into component meshes is highly consistent across sub-

jects. Figure 1(f) illustrates the automatic surface segmenta-

tion results for the same pair of lateral ventricular surfaces,

which is similar to the manual surface segmentation results

used in prior research [8]; even so it improves on past work

as it avoids arbitrarily chopping the surface into 3 parts using

a fixed coronal plane.

Fig. 1. Automatic lateral ventricular surface partitioning and

parameterization using holomorphic one-forms.

After surface segmentation, each lateral ventricular sur-

face is divided to three surfaces, each topologically equivalent

to a cylinder. For each piece, we applied holomorphic flow

segmentation algorithm again on it and conformally mapped

it to a rectangle. Then we register the surfaces across subjects

by pairing up corresponding coordinates in the parameter do-

main [12]. Figure 2 shows an example of a left lateral ventric-

ular surface. After segmentation, each of its three segments is

conformally mapped to a rectangle. The computed canonical

holomorphic one-form and parameterization results are also

shown.

4.2. Multivariate Tensor-Based Morphometry Study on

Lateral Ventricular Surface of HIV/AIDS

In our experiments, we compared ventricular surface models

extracted from 3D brain MRI scans of 11 HIV/AIDS indi-

viduals and 8 control subjects [8]. After surface registration,

we computed the surface Jacobian matrix and applied multi-

Fig. 2. Lateral ventricular surface registration via holomor-

phic one-forms. An example left ventricular surface is shown

in the middle, with its segmentation results. After segmenta-

tion, a new canonical holomorphic one-form is computed on

each piece and each piece is conformally mapped to a rectan-

gle. Surface registration is performed via the parameter do-

main.

variate tensor-based statistics to study differences in ventric-

ular surface morphometry. We ran a permutation test with

5000 random assignments of subjects to groups to estimate

the statistical significance of the areas with group differences

in surface morphometry. We also used a statistical threshold

of p = 0.05 at each surface point to estimate the overall sig-

nificance of the experimental results by non-parametric per-

mutation test [16]. The experimental results are shown in Fig-

ure 3(a). After fixing the template parametrization, we used

Log-Euclidean metrics to establish a metric on the surface de-

formation tensors at each point, and conducted a permutation

test on the suprathreshold area of the resulting Hotellings T2

statistics. The statistical maps are shown in Figure 3(a). The

threshold for significance at each surface point was chosen to

be p=0.05. Although sample sizes are small, we still detected

large statistically significant areas, consistent with prior find-

ings [8]. The permutation-based overall significance p val-

ues [16], corrected for multiple comparisons, were 0.0028,

for the left ventricle and p=0.0066 for the right ventricle, re-

spectively.

To explore whether our multivariate statistics provided

extra power when running TBM on the surface data, we

also conducted four additional statistical tests using different

tensor-based statistics derived from the Jacobian matrix. The

other statistics we studied were: (1) the pair of eigenvalues of

the Jacobian matrix, treated as a 2-vector; (2) the determinant

of Jacobian matrix; (3) the largest eigenvalue of Jacobian ma-

trix; and (4) the smallest eigenvalue of Jacobian matrix. For

statistics (1) we used Hotelling’s T2statistics to compute the

group mean difference. In cases of (2), (3) and (4), we applied

a Students t test to compute the group mean difference at each

surface point. For these four new statistics, their calculated

statistical maps are shown in Figure 3(b)-(e), respectively.

Page 4

Full Matrix J

0.0028

0.0066

Determinant of J

0.0330

0.0448

Largest EV of J

0.0098

0.0120

Smallest EV of J

0.0240

0.0306

Pair of EV of J

0.0084

0.0226

Left Vent Surface

Right Vent Surface

Table 1. Permutation-based overall significance levels, i.e. corrected p values, are shown, after analyzing various different

surface-based statistics (J is the Jacobian matrix and EV stands for Eigenvalue). To detect group differences, it was advanta-

geous to use the full tensor, or its two eigenvalues together; with simpler local measures based on surface area, group differences

were less powerfully detected.

For each statistic, we also computed the overall p-values (see

Table 1). Areas of surface abnormalities detected by different

tensor-based surface statistics were highly consistent. The

experiments also strongly suggested that the newly proposed

multivariate TBM method has more detection power in terms

of effect size (and the area with suprathreshold statistics),

probably because it captures more directional and rotational

information when measuring geometric differences.

5. CONCLUSION AND FUTURE WORK

In this paper, we presented a multivariate tensor-based mor-

phometry framework for analyzing parametric surfaces. The

empirical experimental results demonstrated our method out-

performed TBM methods based on simpler univariate surface

measures. In future, we will apply this multivariate TBM

framework to additional 3D MRI datasets to study brain sur-

face morphometry. We plan to apply other holomorphic dif-

ferentials to study related surface regularization, parameteri-

zation and registration problems.

6. REFERENCES

[1] J. Ashburner, et al. Hum Brain Mapp, vol. 6, no. 5-6, pp. 348–357, 1998.

[2] M. K. Chung, et al. Neuroimage, vol. 14, pp. 595–606, 2001.

[3] L. Wang, et al. NeuroImage, vol. 20, no. 2, pp. 667 – 682, 2003.

[4] M. K. Chung, et al. NeuroImage, vol. 18, no. 2, pp. 198 – 213, 2003.

[5] C. Davatzikos, et al. J Comput Assist Tomogr, vol. 20, no. 1, pp. 88–97, 1996.

[6] P. M. Thompson, et al. Nature, vol. 404, no. 6774, pp. 190–193, March 2000.

[7] M.K. Chung, et al. IEEE TMI, vol. 27, no. 8, pp. 1143–1151, Aug. 2008.

[8] P. M. Thompson, et al. NeuroImage, vol 31, no. 1, pp. 12-23, 2006.

[9] O.T. Carmichael, et al. Biomedical Imaging: Nano to Macro, 2006. 3rd IEEE

International Symposium on, pp. 315–318, April 2006.

[10] Luca Ferrarini, et al. NeuroImage, vol. 32, no. 3, pp. 1060 – 1069, 2006.

[11] Y. Chou, et al. NeuroImage, vol. 40, no. 2, pp. 615 – 630, 2008.

[12] Y. Wang, et al. IEEE TMI, vol. 26, no. 6, pp. 853–865, June 2007.

[13] N. Lepor´ e, et al. IEEE TMI, vol. 27, no. 1, pp. 129–141, Jan. 2008.

[14] V. Arsigny, et al. in Magn. Reson. Med., 2006, vol. 56, pp. 411–421.

[15] Y. Wang, et al. 2nd MICCAI Workshop on Mathematical Foundations of Compu-

tational Anatomy, pp. 36–47, 2008.

[16] P. M. Thompson, et al. J. Neurosciende, pp. 4146–4158, 2005.

Fig. 3. Comparison of various tensor-based morphometry

results on a group of lateral ventricular surfaces from 11

HIV/AIDS patients and 8 matched control subjects.