P1: GAD

Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

Annals of Biomedical Engineering, Vol. 32, No. 2, February 2004 (©2004) pp. 264–273

Three-Dimensional Geometrical Characterization

of Cerebral Aneurysms
BAOSHUN MA,1 ROBERT E. HARBAUGH,2 and MADHAVAN L. RAGHAVAN1
1
Department of Biomedical Engineering, University of Iowa, Iowa City, IA and 2 Department of Neurosurgery,
Penn State University College of Medicine, Hershey, PA

(Received 15 May 2003; accepted 15 October 2003)

Abstract—The risk of rupture of cerebral aneurysms has been ing), presurgical planning,5,6 and biomechanical analyses.22
correlated with the size of the aneurysm sac. It is conceivable that Therefore, it would be of value to develop means to quantify
geometrical shape, not just size may also be related to aneurysm
rupture potential. Further, aneurysm shape may also be a factor the salient features of cerebral aneurysm shape.
in deciding on treatment modalities, i.e., to clip or coil. How- Cerebral saccular aneurysms have various shapes and
ever, our ability to make use of available information on aneurysm sizes. However, precise quantitative information on their
shape remains poor. In this study, methods were developed to three-dimensional (3D) geometry is still quite limited. Most
quantify the seemingly arbitrary three-dimensional geometry of studies have characterized aneurysm shape with broad
the aneurysm sac, using differential and computational geometry
techniques. From computed tomography angiography (CTA) data, categories. Ujiie et al.27,28 categorized the shapes of the
the three-dimensional geometry of five unruptured human cerebral aneurysms into three types: round, dumbbell, and multilob-
aneurysms was reconstructed. Various indices (maximum diam- ular, but aspect ratio (AR) (depth/neck width) was the only
eter, neck diameter, height, aspect ratio, bottleneck factor, bulge parameter used to quantify the shape. In Hademenos et al.’s7
location, volume, surface area, Gaussian and mean curvatures, study of the morphology of aneurysms, the aneurysms were
isoperimetric ratio, and convexity ratio) were utilized to character-
ize the geometry of these aneurysm surfaces and four size-matched roughly classified into unilobular (sphere or oval) and mul-
hypothetical control aneurysms. The physical meanings of various tilobular. Parlea et al.15 analyzed the geometry of simple-
indices and their possible role as prognosticators for rupture risk lobed cerebral aneurysm on the basis of two-dimensional
and presurgical planning were discussed. (2D) angiography. In all these reports where actual pa-
tient aneurysms were studied, the geometry was obtained
from 2D angiography. But typically, these lesions have lit-
Keywords—Intracranial aneurysm, Cerebrovascular biomechan- tle symmetry that will allow for a simplification from 3D to
ics, Geometry, Curvature, Geometrical indices, Shape.
2D. Thus the measurements of geometry features can vary
with the projection obtained,2 resulting in multiple find-
INTRODUCTION
ings for the same aneurysm. In an effort to overcome the
Cerebral aneurysms are localized dilatations of the cere- above-mentioned limitations, this study extended the pre-
bral arterial wall that usually occur on or near the circle of vious work to 3D space and used anatomically accurate
Willis. Rupture of these lesions is the leading cause of sub- aneurysm geometry reconstructed from computed tomog-
arachnoid hemorrhage (SAH), which has a mortality rate of raphy angiography (CTA). On the basis of techniques in
50% and a morbidity rate of 50% for those who survive.20 differential and computational geometry, a set of global in-
One major issue is to predict the rupture potential of these dices for the size and shape of the aneurysm sac were derived
unruptured aneurysms. Clinical studies show that aneurysm and evaluated.
size is a factor.12 However, there is still controversy over the
critical size (from 4 to 10 mm).21,25,31 Recent studies sug- METHODS
gest that rupture does not depend on size alone, but also The proposed procedure of quantifying saccular
on some shape characteristics.4 Further, aneurysm shape aneurysm geometry is composed of four steps: 3D geom-
may also be a factor in choosing the appropriate treatment etry reconstruction, mesh refinement, aneurysm isolation,
modality (for example to decide between clipping and coil- and calculation of various size and shape indices.

Address correspondence to Madhavan L. Raghavan, PhD, BMES Acquisition of 3D Geometry

Member, Assistant Professor, Department of Biomedical Engineering,
1422 Seamans Center (SC), University of Iowa, Iowa City, IA 52242- CTA data were obtained from cerebral aneurysm pa-
1527. Electronic mail: [email protected] tients presenting at the Dartmouth-Hitchcock Medical
264
0900-6964/04/0200-0264/1 °
C 2004 Biomedical Engineering Society
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

Cerebral Aneurysm Geometry 265

FIGURE 1. Surface refinement and isolation of aneurysm sac for an unruptured basilar apex aneurysm (aneurysm A1). (a) Aneurysm
with surrounding vasculature; (b) Mesh before refinement; (c) Mesh after refinement; (d) Cutting plane used to isolate the aneurysm;
(e) Smoothing of boundary curve; (f) Isolated aneurysm sac used for analyses.

Center (Lebanon, NH).9,10 The scanning was performed at parameters were determined heuristically to ensure suffi-
1.25-mm collimation and 0.6-mm slice thickness. From cient mesh size and acceptable element quality [Fig. 1(c)].
these, the 3D geometry of the cerebral aneurysm along with Q av after refinement was about 0.786. Next, the aneurysm
the surrounding vasculature was reconstructed using a man- was isolated from the rest of the vasculature by using a
ually assisted/verified automated edge detection algorithm cutting plane and separating the surfaces on either side
by Medical Media Systems, Inc. (West Lebanon, NH) for of the plane. The plane was chosen by visual observa-
visualization by neurosurgeons toward patient-specific sur- tion using the 3D visualization software, Tecplot (version
gical planning. The reconstructed 3D vasculature surface 9.0, Amtec Engineering, Inc., Bellevue, WA) such that it
was represented as a surface mesh of triangular elements just about touched the branches at its neck [Fig. 1(d)].
[Figs. 1(a), and 1(b)]. The boundary curve of the isolated aneurysm sac was
smoothed to eliminate sharp corners using a nonshrinking
Mesh Refinement and Isolation of Aneurysm Sac curve and surface-smoothing algorithm26 [Fig. 1(e)]. When
the smoothing parameters are conservatively employed,
The originally reconstructed surface mesh usually con-
this method helps in removing sharp corners while re-
tains some low-quality elements, unrealistic sharp corners,
taining genuine geometric features. The isolated aneurysm
and bumps [Figs. 1(a) and 1(b)] likely because it was re-
mesh was used for all subsequent geometrical analysis
constructed only for visualization prior to surgery. The av-
[Fig. 1(f)].
erage element quality index (Q av ) of the mesh was around
0.297 for the five cerebral aneurysm surfaces studied.
Zero-Order Indices
Element quality index Q is a normalized AR of a triangular
element, The indices evaluated may be classified as zero order and
√ second order. Zero-order indices involve only the nodal po-
Q = 2 3ρ/ h max
sitions, while second-order indices are curvature-based and
where ρ is the inradius (the radius of the largest circle that involve the second derivative of the nodal positions. Most
may be inscribed inside a triangle) and h max the longest previous reports13,15,28,29 on assessing geometric charac-
edge of the triangle and 0 < Q ≤ 1 (equilateral triangle). teristics of cerebral aneurysms have used simple measures
To improve the element quality and smooth the surface, (e.g., maximum diameter, neck to height ratios, etc.) based
the original mesh was refined using a commercial algo- on angiographic projections. Although they cannot possibly
rithm Yams (version 2.1, INRIA, France). The refinement describe the entire geometry, these measures do serve some
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

266 MA et al.

purpose because they are simple in definition and have a TABLE 1. Schematic illustrations of zero-order shape in-
clear physical meaning. For example, maximum diameter dices. Although many of these indices are dependent on
each other (e.g., low CR is likely to result in high IPR), the
is one of the most commonly used quantity and does pro-
illustrations provide an approximate measure to construe
vide some measure of size, although it is debatable as to the global shape.
how good a measure it is. To overcome the limitations that
are inherent in the current implementation of these mea- Schematic
sures as discussed in the introduction, a more rigorous 3D
Shape Index Low High
approach was used to estimate these indices.

Isoperimetric ratio (IPR)

One-Dimensional (1D) Size Indices
Three 1D size indices were defined: Height (H ), Max-
imum diameter (Dmax ), and Neck diameter (Dn ). H is the
Convexity ratio (CR)
maximum perpendicular distance from each node to the
neck boundary plane. Dmax is the maximum equivalent di-
ameter for all cross sections that are parallel to the neck
boundary plane. To obtain Dmax , a series of cutting planes Aspect ratio (AR)
parallel to the neck boundary plane were used to obtain
typically 100 consecutive cross sections of the sac along its
height. Since these cross sections and the neck boundary are Bottleneck factor (BF)
typically noncircular in shape [Fig. 1(e)], the definitions for
the maximum diameter and the neck diameter were drawn
from convention in fluid mechanics:32
Bulge location (BL)
D = 4A/ p
where A is the cross-sectional area and p is the perimeter
of the cross section.
with triangular elements using a 2D Delaunay triangula-
2D Shape Indices tion approach.16 The interior volume of this closed mesh
was approximated by the vector sum of the volumes of the
Three 2D shape indices were derived based on the three
tetrahedrons formed by each element with the origin.3,14
1-D size measures: Aspect Ratio (AR), Bottleneck Fac-
tor (BF), and Bulge Location (BL). These were defined as 1 ¯¯ X ¯
V = (L i12 × L i13 ) · (X i1 + X i2 + X i3 )¯
follows: 18 all triangles
A R = H/Dn where L i12 is the vector from node 1 to node 2 of the ith
B F = Dmax /Dn triangular element, Li13 similarly the vector from node 1 to
node 3, X i j ( j = 1, 2, 3) is the position vector of the jth
B L = Hb /H node of the ith triangular element. The total surface area S
where Hb is the distance of the largest cross section (i.e., of the aneurysm sac was approximated by the sum of the
where D = Dmax ) from the neck boundary plane. areas of all triangles in the open mesh.
While AR has an obvious meaning, BF represents the
level to which the neck acts as a bottleneck to entry of 3D Shape Indices
blood during normal physiological function and to coils
during endovascular procedures. BL is a normalized height Convexity ratio (CR) and Isoperimetric ratio (IPR) are
(starting from the neck) at which Dmax is found and provides two 3D shape indices based on volume and surface area.
a measure of the position of the maximum bulge. Table 1 CR is the ratio of the volume of the closed aneurysm sac
provides a schematic illustration of the meanings of these (V ) to that of its convex hull (Vch ):
zero-order shape indices. C R = V /Vch
The convex hull of an aneurysm sac is the smallest, fully-
3D Size Indices
convex surface that encompasses the aneurysm sac.16 In
Volume and surface area are two zero order 3D size in- lay terms, the convex hull of an arbitrarily shaped closed
dices. To estimate the total enclosed volume, the aneurysm surface would resemble a tight elastic membrane around
sac was first closed by filling the open boundary at the neck that surface (see Fig. 2). The Quick-hull algorithm1 was
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

Cerebral Aneurysm Geometry 267

The Gauss–Weingarten map for this surface patch is

à !
2a 2b
A=
2b 2c
The principal curvatures and principal directions are the
eigenvalues (k1 , k2 ) and eigenvectors (v10 , v20 ) of the matrix
A, respectively. They can be expressed explicitly as
p
k1 = a + c + (a − c)2 + 4b2
FIGURE 2. Illustration of the convex hull of an aneurysm sac p
(aneurysm A3). Note that the convex hull resembles a tight k2 = a + c − (a − c)2 + 4b2
elastic membrane encompassing the aneurysm ³ ´
p
v10 = 2b, a − c + (a − c)2 + 4b2
³ p ´
v20 = 2b, a − c − (a − c)2 + 4b2
used to compute the convex hull. CR is a measure of
the level of convexity in the surface and 0 < CR ≤ 1 (a After normalization
fully convex surface). Thus a surface with many collapsed
regions will have a lower CR than one that is mostly convex v1 = v10 /kv10 k
(Table 1). v2 = v20 /kv20 k
IPR is a nondimensionalized ratio of the surface area
to volume of a closed surface and represents a measure the two principal directions are converted from the local
of the degree of folding of the surface.23 For example, coordinate system (P, Q, N) to the global coordinate system
for a given volume, a sphere has the smallest IPR among by
all possible shapes since it has the smallest surface area.   
Px Q x N x v j1
Therefore, IPR is a measure of deviation from spherical P Q
Vj =  y y N y  v j2 
 
, j = 1, 2.
surface. In this study, IPR is defined for saccular aneurysm
as Pz Q z Nz 0

IPR = S/V 2/3 where Px , Q x , etc. are direction cosines of P, Q, and N in

the global coordinate system, v j1 and v j2 are coordinates of
where S is the open surface area, and V the enclosed volume. vj in the local coordinate system.
A hemisphere has a smaller IPR than a half ellipsoid of the Using the above approach, the principal curvatures
same volume (Tables 1 and 3). and principal directions at all nodes of the triangu-
lated aneurysm surface were determined (Fig. 3). Be-
cause curvatures are second-order quantities based on zero-
Second-Order Curvature-Based Indices order quantity (position), they were quite noisy, even af-
Surface curvature distributions have been used in earlier ter surface smoothing.17−19 Small variations in the sur-
reports to study the right ventricular free wall,17 abdominal face will lead to large variations in the principal curva-
aortic aneurysms,18,24 trabecular bone microarchitecture11 , tures, making them impractical to interpret. The contex-
and the surface of human brain.23 In this study, differen- tual peer-review technique17−19 was used to refine the cur-
tial geometry techniques proposed by Hamann8 were ap- vature distribution. Briefly, the principal curvatures, prin-
plied to estimate the distribution of local curvatures, based cipal directions, and normal vectors at each node were
on which global indices were defined. At node p on the made more consistent with those at their neighborhood
surface, a local orthogonal coordinate system (P, Q, N) nodes. The size of the neighborhood would control the
is established. P and Q are unit vectors in the tangent level of smoothing. In this paper, the adjoining nodes
plane at node p. N is the unit normal vector at node p. of each node were chosen as the contextual neighbor-
It is initially approximated by the weighted average of unit hood. The curvatures at the neck boundary points were
normals of all adjoining elements (elements that node p ignored.
is connected to), with the element areas as the weights. The Gaussian (K g = k1 k2 ) and mean (K m = [k1 +
Node p and its adjoining nodes (directly connected nodes) k2 ]/2) curvatures at each node are indicative of the local
are fit to a quadratic surface patch expressed by a bivariate shape of the surface. Table 2 provides an overview of lo-
polynomial f (u, v) of degree two in the local coordinate cal surface characteristics associated with various combi-
system nations of K g and K m . K g and K m averaged or integrated
over the entire surface provide some global measures of the
f (u, v) = au 2 + buv + cv 2 entire surface. At least two different approaches have been
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

268 MA et al.

FIGURE 3. Curvature refinement for aneurysm A1. The contour levels of the initial estimations (from −1.7 to 3.4 for Km and from
−11.1 to 10.2 for Kg ) were reset to be the same as those of the refined to allow for convenient comparisons. The contour values
outside of the range were set to be the closest extreme value in the range.

used in the literature in attempting to derive global geometry of the ith element. Since the curvatures at the bound-
measures from localized K g and K m distributions–area- ary are not defined, the sum is over all elements whose
averaging11 and L2-norm.23 The area-averaged Gaussian nodes are not on the boundary. 4π is used here such that
and mean curvatures are defined as11 GLN will be 1 for a sphere. On the basis of the defini-
X ± X tions above, four curvature-based global indices were ob-
GAA = K gi Si Si
all triangles all triangles
tained: GAA, MAA, GLN, and MLN. GAA and MAA
X ± X have units of [L−2 ] and [L−1 ], respectively, and depend on
MAA = K mi Si Si both size and shape of the aneurysm surface, while GLN
all triangles all triangles and MLN are both nondimensionalized, depend on sur-
The L2-norm of the Gaussian and mean curvatures are de- face shape only and are measures of irregularities on the
fined as23 surface.
s X X ¡
1 ¢
GLN = Si · 2
K gi Si Sensitivity to Random Noise
4π all triangles all triangles
s X Because of the limited resolution of the current CTA im-
1 ages and errors resulting from the software and software
MLN = 2
K mi Si
4π operators, uncertainties exist in the nodal positions of the
all triangles
reconstructured surface mesh. To evaluate the effect of these
where K gi and K mi are the Gaussian and mean curva- uncertainties on different indices, sensitivity analysis was
tures associated with the ith triangular element (defined performed on one of the aneurysm (aneurysm A1). Random
as the averages of its three nodes), Si the surface area noises of given amplitude were added to the nodal positions

TABLE 2. Characteristics of local surface shape for various combinations of mean and
Gaussian curvatures.

Km > 0 Km < 0 Km = 0

Kg > 0 Elliptical (convex) Elliptical (concave) Not possible

Kg < 0 Hyperbolic (more convex) Hyperbolic (more concave) Hyperbolic
Kg = 0 Cylindrical (convex) Cylindrical (concave) Planar
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

Cerebral Aneurysm Geometry 269

of the original CTA-reconstructed surface mesh. Multiple

such surfaces were created at 11 different noise amplitudes
(from 0.05 to 0.5 mm). The resulting noise-added meshes
simulate the surfaces that may result from possible errors
in the reconstruction process. Further, at each given noise
amplitude, 54 random generations were performed to ac-
count for randomness-induced differences. Thus, a total of
54×11 = 594 variations of aneurysm A1 were created. The
entire process of mesh refinement, aneurysm isolation, and
calculation of various indices were repeated on all the 594
surfaces (Fig. 5). With the original, noise-free case as the
standard, percent errors at each noise amplitude and their
95% confidence intervals were calculated.

RESULTS

A suite of linked algorithms in Matlab (version 6.1,

Mathworks, Inc., Natick, MA) were developed for all as-
pects of the geometric analysis including mesh refinement,
isolation of aneurysm sac, and determination of all size and
shape indices. All algorithms were provided sufficient user-
FIGURE 4. Shaded view of all hypothetical and real aneurysms
control to allow for modifications where necessary. (to scale).
The 3D surface geometry of five CTA-reconstructed hu-
man cerebral aneurysms and four hypothetical aneurysms
has been used to evaluate the various indices. The four hypo- identical smoothing protocols. All size and shape measures
thetical axisymmetric aneurysms are hemisphere, half ellip- described earlier were calculated and tabulated in Table 3.
soid, three-fourth sphere, and three-fourth ellipsoid (Fig. 4). Figure 5 demonstrates the effectiveness of the mesh re-
The volumes of all four hypothetical aneurysms are equal finement algorithms in removing the random noise and re-
to the average volume of the five real aneurysms (67.64 covering the original surface. Figure 6 shows the sensitivity
mm3 ), thus allowing them to serve as size-matched con- of various second-order and zero-order indices to random
trols. The ellipsoid has a major-to-minor axis ratio of 2:1 noises in mesh. The random noise amplitudes were from
with the major axis along the height of the aneurysm sac. 0.05 to 0.5 mm. At all noise levels up to 0.5 mm, all zero-
The analyses were conducted on all nine aneurysms, using order indices, except for BL, have errors of less than 2%;

TABLE 3. Geometrical indices of hypothetical and real aneurysms.

Indices

Zero-order indices Second-order indices

Aneurysm H Dmax Dn AR BF BL V S CR IPR GAA MAA GLN MLN

Hypothetical
Hemisphere 3.18 6.36 6.36 0.50 1.00 0.00 67.64 63.59 1.00 3.84 0.09 0.31 0.44 0.19
1/2 Ellipsoid 5.06 5.05 5.05 1.00 1.00 0.00 67.64 68.51 1.00 4.13 0.09 0.32 0.62 0.21
3/4 Sphere 4.01 5.34 4.70 0.85 1.14 0.33 67.64 67.17 1.00 4.05 0.13 0.36 0.66 0.23
3/4 Ellipsoid 6.37 4.24 3.70 1.72 1.15 0.33 67.64 75.70 1.00 4.57 0.10 0.35 0.80 0.24
Real
A1 5.72 6.99 6.66 0.86 1.05 0.25 151.56 116.81 0.98 4.11 0.04 −0.25 1.65 0.30
A2 4.43 5.22 5.22 0.85 1.00 0.00 72.81 72.34 0.98 4.15 0.04 −0.29 1.16 0.26
A3 5.67 4.96 4.69 1.21 1.06 0.19 76.46 84.29 0.88 4.68 0.04 −0.32 2.25 0.32
A4 3.65 3.08 2.06 1.77 1.50 0.48 19.68 33.60 0.97 4.61 0.05 −0.47 2.71 0.33
A5 2.81 3.22 3.10 0.91 1.04 0.36 17.71 29.19 0.96 4.30 0.14 −0.47 1.22 0.26

Note. Locations of real aneurysms: A1: Basilar apex; A2: Right middle cerebral artery; A3: Anterior communicating artery; A4: Left middle
cerebral artery; A5: Right middle cerebral artery. H : height (mm); Dmax : maximum diameter (mm); Dn : neck diameter (mm); AR: aspect
ratio; BF: bottleneck factor; BL: bulge location; V : volume (mm3 ); S: surface area (mm2 ); CR: convexity ratio; IPR: isoperimetric ratio; GAA:
area-averaged Gaussian curvature (mm−2 ); MAA: area-averaged mean curvature (mm−1 ); GLN: L2-norm of Gaussian curvature; MLN:
L2-norm of mean curvature.
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

270 MA et al.

DISCUSSION

The ability to quantify the seemingly arbitrary geome-

try of a human cerebral aneurysm sac can be immensely
valuable in such areas as predicting rupture risk, operative
risk, design of endovascular prostheses and embolization
coils, establishment of standards for patient selection, and
in presurgical planning. The 3D nature of geometry imposes
a minimum requirement that any size or shape measure take
into account the 3D asymmetry and arbitrariness that are in-
FIGURE 5. Random noise added to original CTA-reconstructed
mesh and the resulting mesh refined. The noise amplitude of
herent to the aneurysm sac morphology. While there have
0.156 mm used in the above illustration is the maximum noise been studies that attempted to study brain aneurysm shape,
that may result from the CTA resolution. they mostly took a 2D perspective based on angiographic
projections. Few have studied the 3D cerebral aneurysm
shape with all its true complexity. We utilized data from
for BL, error is less than 10%. Among the curvature-based CTA and developed 3D computer models and have quanti-
indices, MAA has the smallest error, which is less than 2% fied what are likely to be important size and shape features
at all noise levels, and GLN has the largest error, which is al- of the cerebral aneurysm sac and demonstrated the method-
most 60% at noise amplitude of 0.5 mm. To understand the ology on actual human lesions and hypothetical controls.
implication of a particular noise amplitude, it is necessary to Although the four hypothetical aneurysms have the same
compare it with the typical resolution of a CTA image. The volume, the other 1-D and 2D indices are quite different.
interpixel distance in the original CT images is 0.312 mm, According to the clinical practice based on maximum diam-
which means that the maximum noise amplitude due to eter alone, the three-fourth sphere is more dangerous than
CT image resolution is 0.156 mm. At noise amplitude of the three-fourth ellipsoid, but if one were to take AR into
0.156 mm, all zero-order indices except for BL have errors account, as some recent studies28,29 suggest, it is just the
less than 0.7%; for BL, error is within 4%; For second-order, opposite. It is unlikely therefore that any one quantity alone
curvature-based indices (GAA, MAA, GLN, MLN), errors could be a reliable index of rupture risk. Rather, the com-
are less than 9%. Analysis on a second aneurysm showed plete geometry should be considered. The current study is
similar results, so the results are not aneurysm specific. In an effort to provide a set of indices for the whole geometry.
summary, the zero-order indices are quite robust and almost It should be noted that a rigorous approach is needed in
unaffected by random noises in the original mesh, except for determining even the simple quantities such as maximum
BL which may have a few percent errors. The second-order diameter, because the cross sections are usually highly non-
indices are not as stable as zero-order indices as expected, circular [see Figs. 1(e) and 4] and hence 2D projections can
but here too the scope of error is within 10%. be quite unreliable. Consequently, this study makes a case

FIGURE 6. Percent errors and their 95% confidence interval due to random noise in the nodal locations of the original CTA-
reconstructed mesh.
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

Cerebral Aneurysm Geometry 271

for CTA as a better modality than digital subtraction an- the absolute values of MAA are larger for aneurysm A4
giography (DSA) in the characterization of brain aneurysm and A5 than for other aneurysms, which is in accordance
geometry.30 with the fact that these two aneurysms are much smaller
The calculated zero-order indices for all the hypotheti- than others. Therefore the relative values of MAA can be
cal aneurysm sacs agreed quite well with visual observation. a measure of size in cases where the shapes do not differ
The spherical shaped models have lower aspect ratios (AR) too much. The Gaussian curvature has a clear meaning at a
than their ellipsoidal shaped counterparts. The three-fourth local region (Table 2). K g > 0 indicates a surface with an
models have smaller necks and so the bottleneck factors elliptic region while K g < 0 indicates a saddle region (also
(BF) are higher than the corresponding half models. The called gooseneck) which is susceptible to mechanical stress
largest cross sections occur at one-third of the height for the concentration. GAA for all real and hypothetical aneurysms
three-fourth models according to the values of BL, which are positive, but the values for real aneurysms are generally
is exactly how the models are defined. The CR has a clear smaller than the hypothetical models of comparable size,
physical meaning. The greater the CR, the less collapsed which means that real aneurysm surfaces contain some sad-
the aneurysm sac. The CR equal 1 for all the hypothetical dle regions. Similarly, a positive mean curvature in a local
models since they are perfectly convex everywhere. The region indicates that the region is more convex and when
IPR are smaller for the spherical shaped models than the negative, more concave. Thus, MAA contains information
ellipsoidal shaped ones, since this parameter is a measure about concave and convex regions on the aneurysm surface.
of the deviation from spherical shapes. Therefore, the pro- While MAA for the hypothetical aneurysms are all posi-
posed indices can reliably quantify the intended size and tive as expected, they are all negative for real aneurysms.
shape characteristics of the hypothetical aneurysms. This means that concave regions exist on all the five real
The appropriateness of the various indices in quantifying aneurysm surfaces. However, since different regions con-
the real aneurysms (A1–A5) may be assessed by comparing tribute differently depending on local curvatures it is dif-
these objective quantitative indices (see Table 3) to subjec- ficult to tell the percentage of saddle regions and concave
tive visual evaluations of their relative shape characteristics regions on the surface based on these two global indices
(see Fig. 4). Further, their morphologies may also be classi- alone. The L2-norm of curvatures, GLN, and MLN are di-
fied by comparing these indices to those of the hypothetical mensionless indices that are measures of irregularities on
models. All real aneurysms have CR smaller than 1, there- the surface. Therefore, the relative values of these parame-
fore they all have some collapsed regions on their surfaces. ters can tell something about the irregularity of the surface.
Aneurysm A3 clearly appears to have the most collapsed For example, aneurysms A3 and A4 have higher GLN and
regions and expectedly so, its CR is lower (CR = 0.88) than MLN values compared with other aneurysms, because they
all other aneurysms (CR = 0.97 on average). Aneurysm A1 are more irregular in shape than others. The physical mean-
more or less resembles the three-fourth sphere, according ings of these second-order indices need to be explored fur-
to its AR, BL, CR, and IPR values. Aneurysms A2 and ther for improving our ability to use them to infer physically
A3 have approximately the same size as the hypothetical relevant shape characteristics.
aneurysms, but the shapes are quite different. Aneurysm There are some limitations in our calculations and defi-
A2 has a large neck and overall resembles a half ellipsoid, nitions that are worth noting. Because of the limited clinical
while aneurysm A3 has a small neck and is more irregular in and diagnostic data, we ignored the fusiform aneurysms in
shape. By aspect ratio (AR), CR, and IPR values, aneurysm this analysis and concentrate on the morphology of saccular
A4 resembles a three-fourth ellipsoid, but is wider at about (berry) aneurysms. The isolation of the aneurysm sac from
half the height position. Visually, A4 appears to exhibit the the surrounding vasculature was based on visual observa-
greatest bottle-neck effect (i.e., it has a small neck compared tion and on the assumption that the neck is in a plane and lies
to its size) and this is in accordance with the fact that its exactly above the closest artery surface. It is unclear where
bottle-neck factor (BF = 1.50) is substantially higher than the aneurysm really begins and hence this issue needs to
all other aneurysms (BF = 1.04 on average). Aneurysm A5 be addressed better. Clinically, the smallest cross-sectional
resembles a three-fourth sphere by AR and BL values, but area closer to the vasculature is considered to be the neck,
has a larger neck and contains some collapsed regions. It although by our definition it is not. Fortunately though, our
can be seen that these shape indices quantitatively capture studies reveal that both these definitions result in more or
subjective observations on their shape features, while also less the same neck diameter with only about 1% difference
providing a means for categorizing their morphology under on average. As a future step, we are considering predicting
predefined categories. the normal vasculature of an aneurysmal one on the basis of
Compared with the zero-order indices, the second-order statistical morphology data and then subtracting it from the
indices are more complicated and their physical meanings aneurismal vasculature to obtain the aneurysm sac. Another
are not as clear. Since GAA and MAA are dependent on source of error is in the reconstruction process. Owing to
both size and shape, the values alone are difficult to inter- the limitations in the scanner’s resolution, the reconstructed
pret, but the relative values are of importance. For example, surfaces had some sharp corners and bad quality elements.
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

272 MA et al.

When these are being smoothed, much care was taken to en- Analysis on a large database of ruptured and unruptured
sure that the genuine geometric features were not affected. aneurysms will likely reveal if the indices discussed in this
Yet, there is no clear red line between a genuine geometric study are correlated with risk of rupture and/or elective
feature and an artifact. Such issues need to be heuristically surgery. Indeed our future goal is to collect a large database
established using phantoms, something we plan to do in the of ruptured and unruptured aneurysms and retroactively
future. assess statistically significant differences in geometric in-
The curvature-refining technique we adopted was based dices to identify reliable prognosticators for rupture risk.
on a report by Sander et al.,19 but differed slightly from it. Besides shape quantification, a purely physics-based ap-
In the original method, the contextual “neighborhood” of a proach can also be valuable. Three-dimensional stress anal-
given node (where curvatures are to be refined) was defined ysis and computational fluid mechanical studies on cerebral
as the nodes within a sphere of user-defined radius from that aneurysms can also provide useful quantities (e.g., peak
node. The radius of the sphere in-effect represents a con- stress, peak wall pressure) that may serve to be prognostic
trol over the level of smoothing. It needs to be determined indicators. However, such efforts require a better under-
heuristically based on phantoms, but even that may not be standing of the wall thickness and mechanical properties of
perfectly reliable. In this paper we took a very conserva- the aneurysm sac wall, which have been difficult to attain
tive approach by simply defining the neighborhood as the so far. For now, the geometry approach appears to be a real
nodes that the particular node is connected to (the adjoining alternative and the current work has laid the foundation for
neighborhood). This approach needs to be assessed further. further effort in that direction.
It should also be noted that the refinement procedure should
not be performed beyond a certain number of iterations. ACKNOWLEDGMENTS
Understandably, each refinement will decrease the absolute
curvature values, corresponding to an enlarging effect of The authors acknowledge the services of Medical Media
the geometry. This is because for the surface determined by Systems, Inc. (West Lebanon, NH) in the 3D reconstruction
the bivariate polynomial, the maximum absolute principal of cerebral aneurysm models used for analysis in this work.
curvatures are at the dome, whose curvatures are to be re-
fined, so estimations from the neighborhood points will only NOMENCLATURE
underestimate them. Using phantoms of known geometry,
Sacks et al.17 found that two iterations would be appropri- SAH subarachnoid hemorrhage
ate. In this study, we also used two iterations to achieve CTA computerized tomography angiography
satisfactory smoothing effect while avoiding oversmooth- Q element quality index of a triangular element
ing. Ultimately, there are bound to be unavoidable errors in Qav average element quality index of all elements in
the reconstructed and refined geometry. Our analysis of the the mesh
sensitivity of the size and shape indices was an attempt at ρ inradius—the radius of the largest circle that
quantifying the resulting error. As seen in Fig. 6, we found may be inscribed inside a triangle
that all zero-order indices are robust enough to be insensitive H height (mm)
to reasonable errors. However, we did note that the second- D diameter at any cross-section
order indices are more sensitive. There too, when the noise Dmax maximum diameter (mm)
magnitude matches the CT resolution, MAA, GAA, MLN Dn neck diameter (mm)
are within 5% error, while GLN is within 9% error. A cross-sectional area
Ultimately, it is reasonable to wonder what good it would p perimeter of cross-section
be to quantify characteristics that may be observed visu- AR aspect ratio
ally anyway. For example, one does not need CR values to BF bottleneck factor
know that aneurysm A3 has more collapsed regions than BL bulge location
aneurysm A1. Just visual observation is sufficient. Yet, Hb distance of the largest crosssection (i.e., where
quantities like CR provide us something that visual observa- D = Dmax ) from the neck boundary plane
tion just cannot—a measure of the level to which the surface V volume of the aneurysm sac
is collapsed. Armed with such numbers, future efforts can Vch volume of the convex hull
concentrate on statistical comparisons to find if the level of S surface area
surface collapse is correlated with rupture risk or with op- CR convexity ratio
erative risk, thus developing reliable and clinically usable IPR isoperimetric ratio
prognosticators for such applications. This precisely is how k1 and k2 1st and 2nd principal curvatures
quantification becomes an indispensable step toward un- Kg Gaussian curvature
derstanding the roles played by shape and size of aneurysm Km mean curvature
sacs. We believe that ours is a significant first step in that GAA area-averaged Gaussian curvature
direction. MAA area-averaged mean curvature
P1: GAD
Annals of Biomedical Engineering [AMBE] pp1098-ambe-479330 January 20, 2004 15:48 Style file version 14 Oct, 2003

Cerebral Aneurysm Geometry 273

GLN L2-norm of Gaussian curvature 15

Parlea, L., R. Fahrig, D. W. Holdsworth, and S. P. Lownie. An
MLN L2-norm of mean curvature analysis of the geometry of saccular intracranial aneurysms.
A1 to A5 aneurysms 1–5 in the study population AJNR Am. J. Neuroradiol. 20:1079–1089, 1999.
16
Preparata, F. P., and M. I. Shamos. Computational Geometry: An
Introduction, Vol. XII. New York: Springer-Verlag, 1985, p. 390.
REFERENCES 17
Sacks, M. S., C. J. Chuong, G. H. Templeton, and R. Peshock.
In vivo 3-D reconstruction and geometric characterization of the
1 right ventricular free wall. Ann. Biomed. Eng. 21:263–275, 1993.
Barber, C. B., D. P. Dobkin, and H. Huhdanpaa. The quickhull
18
algorithm for convex hulls. ACM Trans. Math. Softw. 22:469– Sacks, M. S., D. A. Vorp, M. L. Raghavan, M. P. Federle, and
483, 1996. M. W. Webster. In vivo three-dimensional surface geometry of
2 abdominal aortic aneurysms. Ann. Biomed. Eng. 27:469–479,
Black, S. P., H. L. Leo, and W. L. Carson. Recording and mea-
suring the interior features of intracranial aneurysms removed at 1999.
19
autopsy: Method and initial findings. Neurosurgery 22:40–44, Sander, P. T., and S. W. Zucker. Inferring surface trace and dif-
1988. ferential structure from 3-D images. IEEE Trans. Pattern Anal.
3 Machine Intell. 12:833–854, 1990.
Desbrun, M., M. Meyer, P. Schroder, and A. H. Barr (1999).
20
Implicit fairing of irregular meshes using diffusion and curvature Schievink, W. I. Intracranial aneurysms. N. Engl. J. Med.
flow. SIGGRAPH, Los Angeles, CA. 336:28–40, 1997.
4 21
Forget, T. R., Jr., R. Benitez, E. Veznedaroglu, A. Sharan, Schievink, W. I., D. G. Piepgras, and F. P. Wirth. Rupture of pre-
W. Mitchell, M. Silva, and R. H. Rosenwasser. A review of size viously documented small asymptomatic saccular intracranial
and location of ruptured intracranial aneurysms. Neurosurgery, aneurysms. Report of three cases. J. Neurosurg. 76:1019–1024,
49:1322–1325; Discussion 1325–1326, 2001. 1992.
5 22
Guglielmi, G., F. Vinuela, J. Dion, and G. Duckwiler. Elec- Seshaiyer, P., and J. D. Humphrey. On the potentially protective
trothrombosis of saccular aneurysms via endovascular approach. role of contact constraints on saccular aneurysms. J. Biomech.
Part 2: Preliminary clinical experience. J. Neurosurg. 75:8–14, 34:607–612, 2001.
23
1991. Smith, A. D. C. The folding of the human brain: From shape
6 to function (PhD dissertation). London: University of London,
Guglielmi, G., F. Vinuela, I. Sepetka, and V. Macellari. Elec-
trothrombosis of saccular aneurysms via endovascular approach. 1999.
24
Part 1: Electrochemical basis, technique, and experimental re- Smith, D. B., M. S. Sacks, D. A. Vorp, and M. Thornton. Surface
sults. J. Neurosurg. 75:1–7, 1991. geometric analysis of anatomic structures using biquintic finite
7 element interpolation. Ann. Biomed. Eng. 28:598–611, 2000.
Hademenos, G. J., T. F. Massoud, F. Turjman, and J. W. Sayre.
25
Anatomical and morphological factors correlating with rupture Solomon, R. A., and J. W. Correll. Rupture of a previously docu-
of intracranial aneurysms in patients referred for endovascular mented asymptomatic aneurysm enhances the argument for pro-
treatment. Neuroradiology 40:755–760, 1998. phylactic surgical intervention. Surg. Neurol. 30:321–323, 1988.
8 26
Hamann, B. Curvature approximation for triangulated surfaces. Taubin, G. Signal processing approach to fair surface design.
In: Geometric Modeling, edited by G. Farin, H. Hagen, and H. Proc 22nd Ann ACM Conf Comp Graph Interactive Tech Los
Noltemeier. New York. Springer-Verlag, 1993, pp. 139–153. Angeles, CA, ACM, New York, 9–11 Aug. 1995.
9 27
Harbaugh, R. E., D. S. Schlusselberg, R. Jeffery, S. Hayden, Ujiie, H., K. Sato, H. Onda, A. Oikawa, M. Kagawa, K. Takakura,
L. D. Cromwell, and D. Pluta. Three-dimensional computerized and N. Kobayashi. Clinical analysis of incidentally discovered
tomography angiography in the diagnosis of cerebrovascular dis- unruptured aneurysms. Stroke 24:1850–1856, 1993.
28
ease. J. Neurosurg. 76:408–414, 1992. Ujiie, H., H. Tachibana, O. Hiramatsu, A. L. Hazel, T.
10 Matsumoto, Y. Ogasawara, H. Nakajima, T. Hori, K. Takakura,
Harbaugh, R. E., D. S. Schlusselberg, R. Jeffery, S. Hayden, L. D.
Cromwell, D. Pluta, and R. A. English. Three-dimensional com- and F. Kajiya. Effects of size and shape (aspect ratio) on the
puted tomographic angiography in the preoperative evaluation of hemodynamics of saccular aneurysms: A possible index for sur-
cerebrovascular lesions. Neurosurgery 36:320–326; Discussion gical treatment of intracranial aneurysms. Neurosurgery 45:119–
326–327, 1995. 129; Discussion 129–130, 1999.
11 29
Jinnai, H., H. Watashiba, T. Kajihara, Y. Nishikawa, M. Ujiie, H., Y. Tamano, K. Sasaki, and T. Hori. Is the AR a re-
Takahashi, and M. Ito. Surface curvatures of trabecular bone liable index for predicting the rupture of a saccular aneurysm?
microarchitecture. Bone 30:191–194, 2002. Neurosurgery 48:495–502; Discussion 502–503, 2001.
12 30
Juvela, S., M. Porras, and O. Heiskanen. Natural history of un- Villablanca, J. P., N. Martin, R. Jahan, Y. P. Gobin, Frazee, G.
ruptured intracranial aneurysms: A long-term follow-up study. Duckwiler, J. Bentson, M. Hardart, D. Coiteiro, J. Sayre, and
J. Neurosurg. 79:174–182, 1993. F. Vinuela. Volume-rendered helical computerized tomography
13 angiography in the detection and characterization of intracranial
Kyriacou, S. K., and J. D. Humphrey. Influence of size, shape
and properties on the mechanics of axisymmetric saccular aneurysms. J. Neurosurg. 93:254–264, 2000.
31
aneurysms. J. Biomech. 29:1015–1022, 1996. Weir, B. Unruptured intracranial aneurysms: A review. J. Neu-
14 rosurg. 96:3–42, 2002.
Lien, S., and J Kajiya. A symbolic method for calculating the
32
integral properties of arbitrary nonconvex polyhedra. IEEE Com- White, F. M. Viscous Fluid Flow, Vol. XXI, New York: McGraw-
puter Graphics and Applications 4:35–41, 1984. Hill, 1991. p. 614.