412 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO.

3, MARCH 2008

Finsler Active Contours

John Melonakos, Student Member, IEEE, Eric Pichon, Member, IEEE,
Sigurd Angenent, Member, IEEE, and Allen Tannenbaum, Member, IEEE

Abstract—In this paper, we propose an image segmentation technique based on augmenting the conformal (or geodesic) active contour
framework with directional information. In the isotropic case, the euclidean metric is locally multiplied by a scalar conformal factor based
on image information such that the weighted length of curves lying on points of interest (typically edges) is small. The conformal factor
that is chosen depends only upon position and is in this sense isotropic. Although directional information has been studied previously for
other segmentation frameworks, here, we show that if one desires to add directionality in the conformal active contour framework, then
one gets a well-defined minimization problem in the case that the factor defines a Finsler metric. Optimal curves may be obtained using
the calculus of variations or dynamic programming-based schemes. Finally, we demonstrate the technique by extracting roads from
aerial imagery, blood vessels from medical angiograms, and neural tracts from diffusion-weighted magnetic resonance imagery.

Index Terms—Directional segmentation, Finsler metric, dynamic programming, active contours, diffusion weighted imagery.

1 INTRODUCTION

G EODESIC active contours [1], [2] have proven to be a very

useful tool for a number of segmentation tasks. Basically,
the idea is to define an active contour model based on the
data depends both upon position and direction. In other
words, for each position and direction in the domain, there
exists a unique voxel intensity. It is here that the concept of the
theory of conformal metrics and on euclidean curve-short- Finsler metric becomes crucial. In fact, if one desires to add
ening evolution. This type of curve evolution defines the directionality to the geodesic active contour framework, we
gradient direction to be that for which a curve will shrink as show that the Finsler condition is necessary to ensure that the
fast as possible relative to its euclidean arc length. One flow is well posed. Note that the Riemannian metric satisfies
multiplies the euclidean arc length by a conformal factor the Finsler conditions and is well posed. We show that there
defined by the features of interest that one wants to extract are some applications for which the Finsler metric outper-
and then one computes the corresponding gradient evolution forms the Riemannian metric, but certainly, there are others
equations. The features that one wants to capture therefore lie for which the Riemannian metric may be more desirable. For
at the bottom of a potential well to which the initial contour oriented domains, as long as the metric satisfies the Finsler
will flow. The key point is that the conformal structure defines condition, the choice of a particular metric is subject to the
a Riemannian metric in the plane for which the features of given application.
interest appear as closed geodesic curves. Flows relative to anisotropic metrics have been studied
In this paper, motivated by certain problems in pattern in the mathematics and physics literature; see [5], [6] and
detection and medical imaging, we develop a version of the references therein. A very simple directional flow was
geodesic active contours in a Finsler metric [3], [4]. See our proposed in some of our earlier work; see [7].
discussion in Section 3 below for the formal mathematical There are many applications of image data in oriented
definition. The basic idea is that we add directionality now to domains. Examples include diffusion-weighted magnetic
the active contours that allows for the segmentation of image resonance imaging in which the magnetic field is biased in
data in oriented domains. As alluded to above, isotropic several directions in order to measure the water diffusivity of
active contour models have been used to segment image data biological tissue. In this case, for each position in the tissue
in isotropic domains, meaning that the value of each voxel and for each direction of the bias field, the corresponding
depends only upon its position in the domain and not upon an image intensity provides a measure of water diffusivity at
associated direction. However, in oriented domains, image that position and direction.
Furthermore, this technique may be used in pattern
detection. Consider a small image pattern patch, which we
. J. Melonakos and E. Pichon are with Department of Electrical and
Computer Engineering, Georgia Institute of Technology, 777 Atlantic
desire to match to our image. Through translations and
Drive, Atlanta, GA 30332-0250. rotations of the patch throughout the image, we can evaluate a
E-mail: [email protected], [email protected]. measure of similarity between the patch and the image for the
. S. Angenent is with the Mathematics Department, University of given patch position and direction. Thus, this is also a problem
Wisconsin, Madison, Madison, WI 53706.
E-mail: [email protected].
in an oriented domain because for each position and rotation
. A. Tannenbaum is with the Departments of Electrical and Computer and of the patch, there exists a unique measure of similarity.
Biomedical Engineering, Georgia Institute of Technology, 777 Atlantic Geodesic active contours in the Finsler framework
Drive, Atlanta, GA 30332-0250. E-mail: [email protected]. provide a mechanism for the minimization of energy
Manuscript received 12 Aug. 2006; revised 5 Feb. 2007; accepted 25 Apr. functionals defined on oriented domains. We derive both
2007; published online 6 June 2007. the curve evolution and dynamic programming based
Recommended for acceptance by G. Sapiro.
For information on obtaining reprints of this article, please send e-mail to:
implementations for Finsler active contours. The latter is
[email protected], and reference IEEECS Log Number TPAMI-0598-0806. necessary since we will want to consider the evolution of
Digital Object Identifier no. 10.1109/TPAMI.2007.70713. open curves for which the level set methodology is not
0162-8828/08/$25.00 ß 2008 IEEE Published by the IEEE Computer Society
MELONAKOS ET AL.: FINSLER ACTIVE CONTOURS 413

appropriate. Preliminary results for directional-dependent minimization of such energy functionals: conformal active
segmentations may be found in [8], [9]. contours (based on gradient descent) and dynamic program-
The idea of using Finsler type metrics for various purposes ming. The former works for closed curves, whereas the latter
is of course not original in this work. First of all, regarding method is valid for curves in which we fix seed and target
curve shortening, Gage [6] has considered curvature driven regions as well.
flows in a Minkowski space. General mean curvature flows
relative to Finsler metrics are studied in [10]. Gurtin and 2.1 Geodesic (Conformal) Active Contours
Angenent have proposed the use of anisotropic Finsler flows In the conformal (or geodesic) active contour model, a local
for problems in crystal growth in [5]. In the computer vision cost, : R2 ! Rþ , is defined based on image information
literature, such directionally dependent metrics have ap- [1], [2]. For a given curve , the total cost LðÞ is defined as
peared in [11], [12], [13], [14], [15]. In some nice work, the the integration of local costs along the curve
connections of graph cuts and such metrics have been Z
described in [13], [14], [15]. Geodesic active contours and 
L ðÞ ¼ ðÞds: ð1Þ
graph cut methods have been combined in [13], [14]. Further, 
in [14], the explicit connection between Finsler distances and This energy can be interpreted as the -weighted length of
the flux methods in [12] is considered in some detail.
the curve. Minimal curves will therefore tend to go through
This paper continues the above line of research. Here, we
regions where is small while, at the same time, constraining
describe Finsler flows in a completely continuous setting valid
the total conformal euclidean length to be as small as possible.
both for both open and closed curves embedded in a euclidean
Convergence of this flow is studied in [1], [2]. It is important to
space of any dimension. The key observation is that if one
note that s is the arc-length parameterization and, therefore,
defines a conformal active contour flow for a direction-
dependent conformal factor, then in order for the flow to be this energy is purely geometric.
If the curve is closed or has fixed end points, a partial
well defined, one needs the standard Finsler convexity
condition (see Section 3). Without this condition, the flow differential equation is obtained by calculus of variations
will be a backward heat equation. Thus, this present work that continuously deforms an initial curve ðt ¼ 0Þ in a way
extends the results in [1], [2] who consider conformally that optimally minimizes its total cost L. This can be
euclidean metrics that only depend on position and are in this interpreted as a gradient descent on the infinite dimensional
sense “isotropic.” (In this paper, we will follow the standard space of curves.
terminology of the mean curvature flow literature in which In the case of the functional (1), the PDE that deforms a
“isotropic” flows are defined relative to a Riemannian metric, given curve in order to minimize the energy as fast as
whereas “anisotropic” flows are defined relative to a direc- possible in the L2 sense is
tion-dependent Finsler metric; see [3].) It is important to note @
that while one can get “directionality” in the Riemannian ¼ ðr  NÞN þ ss ; ð2Þ
framework for image segmentation by a suitable choice of @t
metric (ellipses have directionality), nevertheless, we believe where N denotes the unit inward normal.
that the Finsler geodesic active contour approach gives a As is the standard, this may be implemented using level
natural way of performing segmentation in oriented domains. set methods [17], [18].
We now summarize the contents of this paper. In Section 2,
we review the theory of energy minimizing flows and 2.2 Dynamic Programming
geodesic active contours, as well as dynamic programming. Mortensen et al. [20] have proposed the live-wire segmenta-
Section 3 is the key part of this paper. Here, we define the tion technique that also determines optimal curves for the
notion of a “Finsler metric” and derive the geodesic active same kind of functional. Their framework is based on
contour flow relative to such a structure. In Section 4, we dynamic programming and is applicable to curves with one
describe the dynamic programming based solution and the end fixed in a given seed region S.
numerical implementation of such an approach. In Section 5, The underlying principle of dynamic programming is
we show the results of the experiments using these the principle of optimality verified by minimum-cost pro-
techniques on both MRI tractography and pattern detection blems such as (1) (assuming an optimal curve exists). The
applications. In Section 6, we draw some conclusions and principle is that any subpath p of an optimal path P is itself
describe some future research directions. Finally, we have optimal (otherwise, the P could be improved by following
included two mathematical appendices. The first justifies the another subpath p0 instead of p). This leads to the definition
use of dynamic programming in our situation in which we of the value function L , which is the minimal cost to reach
have a data driven anisotropic conformal factor, and the the seed region S from any point x of the domain
second gives another derivation of the Finsler geodesic flow,
which also captures some of its interesting properties. 
L ðxÞ ¼ minfLðÞ; ð0Þ ¼ x; ð1Þ 2 Sg:
In problems such as (1), the value function satisfies the
2 ENERGY MINIMIZING CURVES Eikonal equation jrL ðxÞj ¼ ðxÞ with boundary condition
Energy minimization approaches to image segmentation L ¼ 0 on S. This equation can be solved numerically using
have been very popular; see [16], [17], [18], [19], and the the fast marching algorithm [18], [21] or can be discretely
references therein. These approaches allow one to define a approximated using Dijkstra’s algorithm.
meaningful energy for a given application and to system- From any point in the domain, an optimal curve in the
atically construct contours that minimize the energy. In this sense of (1) can then be determined by gradient descent on
section, we describe two of the key approaches for the the scalar field L .
414 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008

3 GEODESIC ACTIVE CONTOURS IN A FINSLER Because of the homogeneity of F , that is,

METRIC F ðp; tvÞ ¼ tF ðp; vÞ for all x; v 2 Rn and t  0
In this section, we introduce the notion of direction-
dependent active contours. This will be seen to be essentially the anisotropic length is invariant under orientation
a version of active contours defined relative to a Finsler rather preserving reparametrizations of the curve [4]. However,
than a Riemannian metric. If one thinks of a Riemannian LðÞ may change if one reverses the orientation of .
metric as being defined by a continuously varying family of The extended anisotropic length function F ðp; vÞ is a never
inner products on the tangent bundle of a given manifold, a a strictly convex function of v, because it is homogeneous of
Finsler metric is given by a continuously varying family of degree 1. If F ðp; vÞ2 is strictly convex, then F defines a Finsler
Banach space norms. The strict convexity property given metric on Rn [3], [4]. A necessary and sufficient condition for
below is then an expression of the fact that these norms must this to occur is that r2 F ðp; Þ should be positive definite on
satisfy the triangle inequality. More mathematical details the subspace fv 2 Rn : v ? g. We compute this second
about Finsler flows may be also found in Appendix B. derivative at the particular vector  ¼ ð1; 0; . . . ; 0Þ in terms
For an excellent exposition of the Finsler property and of . (See [24] for a proof.)
comparisons to the Riemannian structure, we refer the
Lemma 1. If f : S n1 ! R is a C 2 function and if F ðvÞ ¼
interested reader to that in [4]. Finally, we should note that
jvjfðv=jvjÞ, then for any v 2 S n1 and any pair of tangent vectors
versions of curve shortening relative to the Finsler structure
X, Y 2 Tv S n1 , one has
have been studied in [5], [6], [22], [23].

3.1 Evolving Space Curves r2X;Y F ðvÞ ¼ ðX; Y ÞfðvÞ þ r2X;Y fðvÞ;
In this section, we set up our notation and define the notion where r2 f is the second covariant derivative of f : S n1 ! R.
of Finsler metric.
Therefore, consider families of evolving curves of the form If v 6¼ 0 is not necessarily a unit vector, then one has
 : ½0; 1  ½0; T Þ ! Rn . For any curve ðx; tÞ, we denote
1  
x @ 1 @ r2X;Y F ðvÞ ¼ ðX; Y ÞfðvÞ þ r2X;Y fðvÞ :
T¼ ; ¼ ; ds ¼ jx jdx: jvj
jx j @s jx j @x
It follows that defines a Finsler metric if and only if the
The curvature vector of  is quadratic form defined by gij þ ri rj is positive definite.
@2 3.2 First Variation
K ¼ ss ¼ :
@s2 We can now compute the first variation of our anisotropic
We say that the curve evolves normally if length functional and derive the flow for the Finsler geodesic
active contours. In this discussion, we assume that the curves
@ are closed or that the family of curves under consideration
V¼ ?T
@t has fixed end points. Assuming the curve  evolves
normally, one has
holds always. For such curve evolutions, one has
Z
  d d
@ @ @ LðÞ ¼ ð; TÞds
@t T ¼ @s V; ; ¼ ðK  VÞ ; and dt dt
@t @s @s Z
ð3Þ 
@ ¼ V  p ð; TÞ þ ð@t TÞ  v ð; TÞ
ds ¼ ðK  VÞds:
@t 
 ð; TÞK  V ds;
For any given function
where p and v denote derivatives with respect to the first
: Rn  S n1 ! Rþ ; and second variables in ðp; vÞ. The derivative with respect
to v 2 S n1 is a covariant derivative. We use (3) to conclude
we let
Z
Z Z   d 
L 1
x LðÞ ¼ V  p ð; TÞ þ Vs  v ð; TÞ
LðÞ ¼ ð; TÞds ¼ ; jx j dx; dt
0 x¼0 jx j 
 ð; TÞK  V ds
where L is the length of . The infinitesimal length function Z

is only defined on unit vectors, but one can extend it naturally ¼ V  v ð; TÞ  V  @s ð v ð; TÞÞ
to all vectors by requiring it to be positively homogeneous of 
 ð; TÞK  V ds
degree 1. We denote this extension by Z
  
  ¼  V  @s v ð; TÞ þ ð; TÞK  p ð; TÞ ds
v
F ðp; vÞ ¼ jvj x;
jvj
so that you get the steepest descent with
so that the anisotropic length of  is ?
Z 1 V ¼ @s ð v ð; s ÞÞ  p ð; s Þ þ ð; s Þss : ð5Þ
Lðð; tÞÞ ¼ F ð; x Þ dx: ð4Þ ?
Here, X denotes the component of X that is perpendicular
0
to s ¼ T.
MELONAKOS ET AL.: FINSLER ACTIVE CONTOURS 415

Note that v ð; vÞ 2 Tv S n1 is a vector perpendicular to v This means that if an optimal trajectory x ðÞ is found such
since it is the gradient of a function on S n1 at the point that x ð0Þ ¼ x0 and x ð1Þ 2 S, then for any r 20; 1½, the
v 2 S n1 . If you expand the derivative @s ð v ð; s ÞÞ you will subtrajectories xj½0;r and xj½r;1 are also optimal. See [25] for a
get two terms, one of which contains second derivatives of detailed proof.
, namely, vv ð; s Þ  ss . In our case, using xt ¼ u and xð0Þ ¼ x0 , the following
The steepest descent flow, then leads to the following Hamilton-Jacobi-Bellman equation is obtained:
quasi-linear PDE
0 ¼ inf fLðx0 ; u ð0ÞÞ þ rJ  ðx0 Þ  u ð0Þg: ð7Þ
? u ð0Þ
V ¼ s  rvp ð; s Þ  p ð; s Þ
ð6Þ In general, the value function may not be differentiable. In
þ ð; s Þ þ vv ð; s Þ ss :
that case, the differential (7) holds in the sense of viscosity
Here, vv ð; TÞ is the linear map on Tv S n1 defined by the theory. See [26].
second covariant derivative of ð; vÞ. Thus, for any pair of Under our above assumptions, this can be applied to the
vectors X, Y 2 Tv S n1 , one has by definition Finsler cost functional
 Z L
r2X;Y ðp; vÞ ¼ X; vv  Y :
LðÞ ¼ ððsÞ; s ðsÞÞds
One sees that (6) is a parabolic equation exactly when 0
Z 1 Z 1 ð8Þ
defines a Finsler metric. This equation defines our model for ¼ F ð; x =jx jÞjx jdx ¼ F ð; x Þdx;
the Finsler geodesic active contours. 0 0
The above derivation works for closed curves. In the
where s is arclength and L is length of the curve.
planar case, one may implement such a flow using level set
techniques. We, however, are also in interested direction- The resulting Hamilton-Jacobi-Bellman equation is
dependent flows for curves in which we fix seed and target  x ð0Þ
regions, and for this, we will propose (in Section 4) the use 0 ¼ inf ð0Þ; jx ð0Þj þ rL ðð0ÞÞ  x ð0Þ
of dynamic programming. This is essential for diffusion x ð0Þ jx ð0Þj
tensor imaging in which we want to discover white matter and, finally,
tracts starting from some point in the image. 8
Finally, in Appendix B below, we derive the first variation < 0 ¼ inf f ðpp; d^Þ þ rL ðppÞ  d^g;
of the Finsler functional in terms of the homogeneous d^ 2S n1 ð9Þ
:
extension F , which leads to another numerical scheme. L ðsÞ ¼ 0 for s 2 S;
where anticipating our discussion in Section 5 for images,
4 DIRECTION-DEPENDENT DYNAMIC PROGRAMMING we denote the given voxel location (that is, point in Rn ) as p
In this section, we show how dynamic programming can be and direction as d^.
used to determine optimal curves. The Finsler metric
4.2 Numerics
condition on the anisotropic factor will be assumed
throughout this discussion (so that optimal paths will This equation can be solved numerically in a straightforward
manner. Several numeric approaches may be used, such as
indeed exist).
those given in [27], [28], [29]. We use the Fast Sweeping
4.1 Optimal Control and the Principal of Optimality approach proposed in [29]. From any point p 0 2 Rn , an
Consider the optimal control problem of determining a optimal path in the sense of (8) can then be determined by
trajectory x : ½0; 1 ! Rn that is optimal with respect to the following locally the vector d^ for which the minimum is
functional attained in (9).
Z 1 Algorithm 1. Sweeping algorithm to solve the Hamilton-
JðxðÞ; u ðÞÞ ¼ LðxðtÞ; u ðtÞÞdt: Jacobi-Bellman (9); see [28]
0 Require: seed region S, direction-dependent local cost
We assume in the discussion below that L is homogeneous 1: Initialize L ðÞ þ1, except at starting points s 2 S,
of degree 1 in the u variable. The control u ðÞ is defined by where L ðsÞ 0
2: repeat
_ ¼ u ðtÞ:
xðtÞ 3: sweep through all voxels p , in all possible grid
directions
For any given starting point x0 , define the value function as 4: d^0 arg mind2SSn1 fL ; ðpp; d^Þ
the minimum cost for reaching a seed region S  Rn from x0 5: if fL ; ðpp; d^0 Þ < L ðppÞ then L ðppÞ fL ; ðpp; d^0 Þ and
d^ ðppÞ

d^ end if
0
J  ðx0 Þ ¼ inf JðxðÞ; u ðÞÞ:
u ðÞ;xð0Þ¼x0 ;xð1Þ2S 6: end sweep
7: until convergence of L
When an optimum exists, it may be found using Bellman’s
The algorithm sweeps through all points p in search of the
principle of optimality [25]. Basically, this states that if x ðÞ is an
least expensive direction. The cumulated cost to reach p from
optimal trajectory, then all subpaths are also optimal. This
direction d^ is
can be expressed by the following relation:
! !
Z r X
n1 X
n1
 
J  ðx0 Þ ¼ inf LðxðtÞ; u ðtÞÞdt þ J  ðxðrÞÞ : fL ; ðpp; d^Þ ¼ k L ðpp þ  k Þ þ ðpp; d^Þ k ;
u ðÞ;xð0Þ¼x0 ;xð1Þ2S 0 k¼0 k¼0
416 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008

where the n neighbors p þ  0 ; . . . ; p þ  n1 of p in direction d^

are interpolated using the components of the vector

 ¼ ½0 j . . . jn1 1 d^. Thus, in three dimensions, this would
be n ¼ 3 neighbors among 26. If we take then, for example,
d^ ¼ ð0:912; 0:228; 0:342Þt , one could choose the three neigh-
bors  0 ¼ ð1; 0; 0Þt ,  1 ¼ ð1; 0; 1Þt ,  2 ¼ ð1; 1; 1Þt , and the
corresponding weights would be  ¼ ð0 ; 1 ; 2 Þ ¼ ð0:228;
0:114; 0:570Þ. One interpretation is that the value for reaching
p from direction d^ will be influenced most by the value at
p þ  2 , which is the neighbor as much in direction d^ as the grid
allows. Since, in general, it is not exactly in that direction, the
final result will also be interpolated using the two other most
aligned neighbors p þ  0 and p þ  1 .
One sees that if the continuous direction d^ is exactly
defined by one of the neighboring voxels, that is, d^ ¼  k =kk k,
then f ¼ L ðpp þ  k Þ þ ðpp; d^Þkk k, which is the cost for
reaching voxel p from voxel p þ  k . This same quantity would
be computed in Dijkstra’s algorithm. Unlike Dijkstra’s
algorithm, however, the search for the optimal direction is
not restricted to discrete grid directions, and the minimiza-
tion is performed continuously over the sphere S n1 . In our
implementation, the minimization is performed over 100 di-
rections sampled uniformly on the sphere,1 and the coeffi-
cients ðd^Þ are precomputed. Mathematical details and a
convergence proof are available in [28].
Note that the number of iterations (where one iteration is
defined as one set of all possible directional sweeps)
required for convergence depends upon the number of
turns in the curve. Each iteration, therefore, can flow
information through one complete turn of the curve. In the
Applications and Simulations in Section 5, we found all the
curves in three iterations or less.

5 APPLICATIONS AND SIMULATIONS

Having developed the theory behind Finsler active contours,
we now illustrate these via several experiments. First, we
demonstrate curve shortening with respect to certain
anisotropic conformal factors as opposed to isotropic curve
shortening. Second, we provide a synthetic example that Fig. 1. Synthetic 2D example. These three different local costs depend
demonstrates a particular case where Finsler active contours only on direction. They are represented as polar plots (first row). The
corresponding deforming shapes are presented on the following rows
capture a corner in directional data. Third, these methods are
(black). The initial curve (gray dashed) is bean shaped. See text.
applied to a pattern detection problem, specifically to detect
roads and vessels in 2D imagery. Fourth, we show 3D results
of these techniques applied to diffusion-weighted magnetic The first cost is isotropic. In that case, the global cost of the
resonance imagery for white matter brain tractography. curve is its euclidean length, and the minimizing flow is the
euclidean curvature flow [30], [31]. This flow shrinks any
5.1 Closed Curves Evolving According to the planar shape to a circular-shaped point. This is illustrated on
Finsler Flow the first column in Fig. 1. The second and third costs are
In this section, in order to compare the proposed direction- defined using direction information. In particular, the second
dependent framework to the isotropic framework, we cost favors portions of the curve that are either horizontal or
examine the evolution of a closed bean-shaped curve with vertical. The third costs does exactly the opposite and favors
respect to three separate conformal factors. portions of the curve that are diagonal. The corresponding
In order to isolate the effect of directional information, evolutions can be observed on the second and third columns
we study local costs that do not depend on position but only
in Fig. 1. The influence of direction information is very visible
on the direction N ¼ ½n1 n2 t (we use the unit normal
instead of the unit tangent in defining the conformal factors; in these figures.
for planar curves, this is clearly equivalent): 5.2 Simulated Example
1. ¼ 1. Finsler active contours extend the isotropic geodesic active
2. ¼ maxðp1ffiffi2 jn1 þ n2 j; p1ffiffi2 jn2  n2 jÞ3 =0:75. contours by adding directionality to the distance functional.
3. ¼ maxðjn1 j; jn2 jÞ3 =0:75. Therefore, in oriented domains where directionality is
important, Finsler active contours capture important direc-
1. For the algorithm to initialize properly, discrete grid directions have to tional information unavailable to isotropic geodesic active
be present. contours.
MELONAKOS ET AL.: FINSLER ACTIVE CONTOURS 417

active contour approach implemented using the Fast March-

ing Method [17], [18]. At each point in the image, the strongest
diffusion value (without respect for directionality) was
chosen as the scalar to be used in the isotropic approach.
Since, by construction, the strongest diffusion value at each
point is the same (that is, the magnitude of red arrows is the
same as the magnitude of green arrows), it is obvious that the
optimal path for the isotropic geodesic active contour is a
straight line connecting the yellow markers, as shown by the
blue line in Fig. 2b. This example illustrates the need for
directionality, and the difficulties that arise from attempting
to discard the directionality via a pointwise scalar function (in
this case, the maximum diffusion at each point).
Second, we compared the Finsler active contour approach
to the Riemannian active contour approach. In this simula-
tion, we created tensors from the directional data using the
Stejskal-Tanner equation, as shown by the zoomed-in view of
the corner in Fig. 2d. This figure shows a zoomed-in view of
the Riemannian image for nine pixels surrounding the
corner. The resulting Riemannian optimal path is shown in
green in Fig. 2b. This figure reveals the smoothing effect,
which the Riemannian metric tends to have on the result.
This is due to the directional averaging, which occurs in the
construction of the tensors by imposing the elliptical
diffusion profile on the data.
These simulations reveal that, for this particular case, the
Fig. 2. A simulated 2D example. (a) The synthetic baseline “L” shaped Finsler active contour is desirable due to the fact that it is
corner. (b) Finsler results (red), Riemannian results (green), and capable of capturing a sharp corner in an oriented domain.
Isotropic results (blue). (c) A zoomed-in view showing the directional In other cases, the scalar geodesic active contour or
data of nine pixels surrounding the corner. (d) A zoomed-in view Riemannian active contour approaches may have attributes,
showing the Riemannian image of nine pixels surrounding the corner,
which render them more suited to the particular task.
created by applying the Stejskal-Tanner equation to the directional data.
5.3 Curve Detection in Imagery
Furthermore, Finsler active contours extend Riemannian Given a sample image I for which a portion of a curve  is
active contours by minimizing with respect to the more given (for example, by a human expert), imagine sliding a
general Finsler metric. It will be shown that, in some cases, small rectangular window along the curve in such a way
minimization with respect to a Riemannian metric will yield a that the center of the window is always on the curve, and
smoothed version of the result obtained via the Finsler metric.
the long axis of the window is aligned with the tangent to
In order to analyze these properties of Finsler active
the curve. Computing the average value of image intensities
contours, we have devised a synthetic simulation. We at each point inside the window as the window slides along
constructed a simple 2D (64  64 pixel) image with an “L” the curve, one obtains an average pattern of what the image
shaped corner, as seen in Fig. 2a. The goal of this simulation looks like locally around the curve  . A location and
is to find an open curve that extends from the top yellow direction-dependent pattern detector can then be defined by
marker to the bottom yellow marker and that follows the translating and rotating the average pattern and determin-
“L” shape. We performed this simulation using the ing how well it matches the image.
dynamic programming numerical scheme. The yellow This protocol was applied to a road detection task. Fig. 3
markers are given as known inputs to the algorithm. shows the portion of the road that was used to learn the
The synthetic directional data was created by randomly pattern detector. The dimension of the window along its long
drawing samples from the uniform distribution on the axis (that is, in the direction of the road) was chosen to be four
interval [0, 0.5] for evenly spaced directions on the unit circle. times the width of the road, and the dimension of the window
Then, we added the signal by giving directions of high along its short axis (that is, normal to the road) was chosen to
diffusion slightly stronger values, as shown in Fig. 2c. This be two times the width of the road. The pattern detector was
figure shows a zoomed-in view of the directional image for then obtained for any position and direction by translating
nine pixels surrounding the corner. The red arrows corre- and rotating the average window and computing the sum of
spond to the high-diffusion directions (that is, the signal), the the square of the difference between the intensity of the image
green arrows correspond to a diffusion process contrary to and that of the average window. Curves that will be minimal
the signal flow, and the blue arrows are the randomly drawn for this metric will then be those for which the image locally
background samples. matches the pattern of a road.
Using the Finsler active contour framework, we are able to Curves were deformed using a straightforward particle-
capture the “L”-shaped corner accurately, as pictured in red based approach. Fig. 4 shows two different initial curves
in Fig. 2b. We will now compare this to two other methods. converging to the same portion of the road. Fig. 4 shows the
First, we show that adding directional information is evolution of a self-intersecting initial curve. Finally, Fig. 5
critical to capturing the corner. We proceed by comparing the illustrates the use of dynamic programming. Note that for
Finsler active contour approach to the isotropic geodesic very low SNR, the dynamic programming fails.
418 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008

Fig. 5. Results of road detection on noisy images using dynamic

programming. (a) Recovered curve (original). (b) Recovered curve
ðSNR ¼ 0:50Þ. (c) Recovered curve ðSNR ¼ 0:25Þ.

Fig. 3. Road image and manually determined curve used for learning the
pattern detector.

Fig. 6. Curve evolution on a real image. The local cost is determined

using a pattern detector. (a) Initial curve. (b) Evolving curve. (c) Steady
state.

Fig. 7. Vessel detection using dynamic programming. The procedure

was run independently for two seed points (large discs) and several
target points (small discs). On the right, noise was added. This changes
the recovered curves as one of the branches at the bottom is no longer
visible. (a) Original image. (b) Recovered curves. (c) Recovered curves
ðSNR ¼ 0:25:Þ.

locally affects the Brownian motion of water molecules and

will be reflected in the DT-MRI diffusion measurements. In
classical theory, diffusion follows a Gaussian process which
can be described locally by a second order tensor.
A simple and effective method for tracking nerve fibers
Fig. 4. Particle-based curve evolution with different initializations. using DT-MRI is to follow the direction of maximum diffusion
(a) Initial curve 1. (b) Initial curve 2. (c) Initial curve 3. (d) Evolving
at each voxel [33], [34], [35], [36]. Although this method is
curve 1. (e) Evolving curve 2. (f) Evolving curve 3. (g) Steady state 1.
(h) Steady state 2. (i) Steady state 3. wide spread and used in various ways, the fiber trajectory is
based solely on local information, which makes it very
The same experiments were performed on a medical sensitive to noise. Moreover, the major direction of diffusion
image to track blood vessels. As before, the metric was can become ill defined, for example, at fiber crossings.
defined by an initial manual segmentation step. Fig. 6 shows As an application of our framework, tractography is set
the result of the curve evolution approach. Fig. 7 shows the in a continuous minimum cost framework. This is different
result of the dynamic programming approach. In that case, from that in [37], [38]. Indeed, in these works, the authors do
not propose variational (cost minimizing) techniques. Local
noise was artificially added.
costs are defined for every direction on the unit sphere
5.4 High Angular Diffusion MRI Tractography based on high angular resolution diffusion imagery.
Diffusion Tensor Magnetic Resonance Imaging (DT-MRI) Equivalently, this can be considered a minimum arrival
measures the diffusion of water in biological tissue [32]. The time framework in which the speed of fictitious particles
utility of this method stems from the fact that tissue structure would be the inverse of the cost.
MELONAKOS ET AL.: FINSLER ACTIVE CONTOURS 419

Fig. 8. Cost per unit length of end points of optimal curves for different
b-values is a validity index. Best results are achieved for the highest
b-value. (a) b ¼ 500. (b) b ¼ 1; 000. (c) b ¼ 1; 500.
Fig. 9. Fiber tracking from high angular resolution data set
5.4.1 Constructing the Direction-Dependent Cost ðb ¼ 1; 500s=mm2 Þ.
Most front propagation techniques for diffusion tensor
tractography use some ad hoc function f of the quadratic form
d^t Dd^, where D is the diffusion tensor. If the Gaussian
assumption holds, the diffusion weighted images follow
   
S p; d^ ’ Sðpp; 0Þ exp b d^t DðppÞd^ : ð10Þ

Tensor-based techniques can formally be extended to

high angular resolution diffusion data sets by setting (see
[8] and references therein for details)
!!
^  1 Sðpp; d^Þ
ðpp; d Þ ¼ f  log : ð11Þ
b Spp; 0 Þ
Fig. 10. Proposed technique on high angular resolution data (blue)
However, in the experiments below, we employed the compared with streamline technique on tensor field (red)
following metric: ðb ¼ 1; 500s=mm2 Þ.
0 13
Sðpp ; ^Þ
d sequence with 31 different gradient directions with b-values

ðpp; d^Þ ¼ @R vÞ
Sðpp;^
A : ð12Þ (see (10)) of 500, 1,000, and 1,500s=mm2 , on a 1.5 Tesla GE
v^?d^ Sðpp;00Þ d^
v Echospeed system. The data was acquired with different
b-values to enable comparisons of the results. Traditional
This quantity will be small if there is diffusion in eigenvector based tractography is normally carried out in
direction d^ (numerator small) and limited diffusion in data with b-values in the range of 700-1,000 s=mm2 . Higher
directions normal to d^ (denominator large). The main b-values give data with higher angular contrast but at the
advantage of this formulation is that several data points expense of more noise.
are used to compute the denominator that improves the Cost per unit length, which can be interpreted as a
signal to noise ratio. We chose fðxÞ ¼ x3 experimentally to validity index for the putative tracts was determined for all
accentuate the anisotropy of the data. Because, experimen- b-values, as shown in Fig. 8.
tally, only a few dozen directions are used for acquisition, All curves are optimal given their starting point. The cost
interpolation was also performed. per unit length is a measure of the likelihood that a tract
It is very important to note that the anisotropic from the seed region passes through a given point in the
conformal factor is constructed from the data, and for domain. The best contrast (corresponding to the most
example, in the DW-MRI case, we have no proof that the coherent set of “superoptimal” tracts for a given seed point
corresponding F 2 is always strictly convex. However, in posterior of the corpus callosum) was obtained at the
Appendix A below, we include for completeness a standard highest b-value available. This could indicate that the
argument that shows that using a scheme such as fast algorithm was able to take advantage of the higher angular
sweeping, one computes the optimum relative to the contrast in spite of the lower SNR. Tract results for several
convexification of F 2 . This type of convexification argument user-defined seed points are presented in Fig. 9.
is well known in the optimal control literature (see the Finally, the proposed technique was compared to a
classical text [39] for details). Thus, we are computing in fact streamline technique (see [33], [34], [35], [36]), which requires
geodesic active contours relative to the Finsler metric the computed tensor field, as shown in Fig. 10.
defined by the convexification of the defining function. Although validation is a very challenging task due to the
lack of ground truth, it can be noted that both algorithms
5.4.2 Results give similar results even though their inputs are different.
Here, we show results obtained by applying the methodology The tracts of the proposed technique tend to be more
described in the above sections to diffusion weighted data coherent as any noise in the data might set the streamline
sets acquired using a single-shot diffusion-weighted EPI off course, whereas the proposed technique is more global.
420 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008

5.5 A Note on Timings Let : S n1 ! R be a function and suppose one defines
Here, we present a note on the timings for each of the the cost of a curve  to be
experiments. All of the experiments were performed on a Z
common PC. We used a Dell Optiplex GX270 with an Intel CðÞ ¼ ðTÞ ds:

Pentium 4 single core chip and 2 GB of memory. Each of the
n
experiments above was conducted using Matlab code with Define F : R ! R to be the homogeneous extension of
C mex functions for the Fast Sweeping implementation. degree one of , so
Moreover, although this code was sufficiently fast for our  
v
purposes, we are in the process of porting the code to the F ðvÞ ¼ jvj :
jvj
freely available Insight Toolkit (ITK) [40].
All of the particle-based approaches, on 2D roads and This function need not be convex. We define its convex hull
vessels, converged quickly and in negligible time. Also, all of to be
the Fast Sweeping approaches converged in 3 iterations or  
less (where one iteration consists of all of the possible F~ðvÞ ¼ sup a  vþb : a 2 Rn ; b 2 R; 8x a  xþb F ðxÞ : ð13Þ
directional sweeps through the image), see Section 4.2 for Furthermore, we define
more discussion on Fast Sweeping convergence. For example,  
it took 0.13 seconds for the experiment in Section 5.2 to E ¼ v 2 Rn : F ðvÞ ¼ F~ðvÞ :
converge on a 64  64 grid, and it took 20 seconds for the
This is the set of extreme points.
experiment in Section 5.3 to converge on a 787  787 grid.
The cost CðÞ of any parameterized curve  : ½0; 1 ! Rn
Also, with simple masking of irrelevant voxels, the time to
is given by
converge for the 3D DWMRI experiment in Section 5.4 was
under 5 minutes. The time to compute a path from a target Z 1
point back to the seed point is negligible compared to the time CðÞ ¼ F ð0 ðÞÞ d: ð14Þ
0
required to run the Fast Sweeping portion of the algorithm.
One can also define the relaxed cost as
Z 1
6 CONCLUSIONS ~
CðÞ ¼ F~ð0 ðÞÞ d: ð15Þ
In this paper, we proposed a natural approach for adding 0
directionality to the conformal active contour technique. The Clearly, one always has
cost of a curve is defined as the length of the curve weighted
by some position and direction-dependent local costs based ~
CðÞ CðÞ ð16Þ
on image information. This allows for the asymmetric
processing of information based on direction. The local costs since F~ðvÞ F ðvÞ for all v 2 R . n

can be defined from a direction-dependent pattern detector, For any given set   Rn and point p 2 Rn n  one
which can be obtained after a learning step. defines the cost to get to  from p as
The techniques described in the paper are very general  
C ðpÞ ¼ inf CðÞ : ð0Þ ¼ p; ð1Þ 2  : ð17Þ
and could be used to extract information from many
different types of imagery. They have been applied mostly Here, the infimum is taken over all curves from p to some
to medical imaging data sets and, in particular, to images of point in .
the brain. In fact, it was the problem of extracting white One can also define
matter tracts that initially motivated this line of research. In  
the medical area, it could be also be applied to the ~
C~ ðpÞ ¼ inf CðÞ : ð0Þ ¼ p; ð1Þ 2  ; ð18Þ
extraction of blood vessels from various imaging modalities
where the infimum is again taken over all curves from p to
such as magnetic resonance or computed tomography.
some point in .
Finally, we have only described the Finsler framework in
the case of curves. One can derive and study a similar flow Lemma 2. For any curve  : ½0; 1 ! Rn and any " > 0, there
for surfaces. This will be the topic of our future research in ~ : ½0; 1 ! Rn with the
exists a piecewise linear (PL) curve 
studying directional-based segmentation methods. same endpoints for which one has

CðÞ ~ Þ
~ ¼ Cð ~ ~
CðÞ þ ":
APPENDIX A
FINSLER COMPUTATION FOR NONCONVEX This lemma says that any curve from p to  can be replaced
FUNCTIONS by a curve with the same endpoints whose velocities are
extreme points for F (that is, F~ð0 ðÞÞ ¼ F ð0 ðÞÞ for all ),
As we noted above, there may be problems in the nonconvex without increasing the cost by more than ".
case for our directional segmentation scheme. Nevertheless, An immediate consequence of the lemma is:
as we will indicate in this appendix, the fast-sweeping type
numerical approach will automatically capture an approx- Lemma 3. C~ ðpÞ ¼ C ðpÞ.
imation for the convexification of the functional (In the sense It follows that any (correct) method that computes C by
to be made precise below.) We abstract the situation to be propagating the front @ outwards with velocities given by
studied as follows. F will actually compute C~ .
MELONAKOS ET AL.: FINSLER ACTIVE CONTOURS 421

Proof of Lemma 2. First, it is clear that one can approximate One could also choose Gðp; vÞ ¼ F ðp; vÞ, which would
the given curve  by a PL curve 1 for which F~ð01 ðÞÞ result in
F~ð0 ðÞÞ þ " holds for 0  1. Thus, Cð ~ 1 Þ CðÞ
~ þ ". 1  ?
Next, the PL curve 1 is linear (that is, 01 ðÞ is t ¼ ðFv Þx  Fp :
F ð; x Þ
constant) on each interval ði1 ; i Þ from some partition
0 ¼ 0 < 1 <    < n ¼ 1. Let vi be the constant value of
01 on ði1 ; i Þ. If vi 62 E (recall that E is the set of extreme B.2 Some Identities Involving F
points), then vi is a convex combination of certain Since F ðp; tvÞ ¼ tF ðp; vÞ, for all t  0, one has
v~1i ; . . . ; v~ni 2 E, that is
Fv ðp; tvÞ ¼ Fv ðp; vÞ ð8t > 0Þ; ð24Þ
X
vi ¼ 1 v~1i þ    þ n v~ni ; i  0; i ¼ 1; ð19Þ Fv ðp; vÞ  v ¼ 0: ð25Þ
whereas For the second derivative Fvv ðp; vÞ, which we regard as a
symmetric linear transformation on Rn , this implies that
F~ðvi Þ ¼ 1 F~ðv1i Þ þ    þ n F~ðvni Þ: ð20Þ
Fvv ðp; vÞ  v ¼ 0 ð26Þ
Now, define a PL curve 2 that has
and, hence,

02 ðÞ ¼ vki for  2 ij1 ; ij ; ð21Þ Fvv ðp; vÞ  w ? v ð8w 2 Rn Þ: ð27Þ
where ij
¼ i1 þ j ði  i1 Þ. Thus, we replace the We may also regard Fvp ðp; vÞ as a linear transformation
segments of 1 whose velocity are not in the extreme on Rn , and in this case, we have
set E of the function by PL zigzag curves with the same  
begin and end points whose velocities are in E. v  Fpv ðp; vÞ  Fp ðp; vÞ ? v: ð28Þ
With this definition, one has Indeed, in tensor notation, this amounts to
2 ði Þ  2 ði1 Þ ¼ 1 ði Þ  1 ði1 Þ: vi Fpi vj ðp; vÞvj ¼ Fpi ðp; vÞvi ;
Hence, if one sets 2 ð0Þ ¼ 1 ð0Þ ¼ p, then one ends up which one obtains by differentiating the Euler identity
with 2 ð1Þ ¼ 1 ð1Þ 2 .
~ 1 Þ.
~ 2 Þ ¼ Cð Fvi ðp; vÞvi ¼ F ðp; vÞ
Using (20), one can easily see that Cð
0
Since 2 ðÞ 2 E for all  one also has Cð2 Þ ¼ Cð~ 2 Þ. with respect to p in the direction of v.
~
Hence, 2 is a curve from p to  with Cð2 Þ CðÞ þ ".t u
B.3 Steepest Descent with F
We continue with (23)
APPENDIX B  ?
jx j t ¼ Fvv ð; x Þxx þ x  Fpv ð; x Þ  Fp ð; x Þ :
FINSLER FLOW IN TERMS OF F
In this section, we describe the Finsler flow in terms of the By (27), the combined terms on the right are already
homogeneous extension F and derive some of its properties. perpendicular to x . We therefore find that (23) is equivalent
with
B.1 First Variation Using F
1  
Instead of writing LðÞ in terms of , we can also write LðÞ as t ¼ Fvv ð; x Þ  xx þ x  Fpv ð; x Þ  Fp ð; x Þ : ð29Þ
jx j
in (4). The first variation is then given by the usual Euler-
Lagrange equation More generally, one gets the equation
Z 1 1  
d   t ¼ Fvv ð; x Þ  xx þ x  Fpv ð; x Þ  Fp ð; x Þ :
LðÞ ¼ Fp  ðFv Þx  t dx: ð22Þ Gð; x Þ
dt 0
ð30Þ
If one looks for a normal evolution equation (that is,
t ? x ), then one is led to an equation of the form No matter which G one chooses, this equation fails to be
parabolic since Fvv always has a zero eigenvalue, namely,
 ?
t ¼ ðFv Þx  Fp Fvv ðp; vÞv ¼ 0.

for some positive scalar . If one additionally wants the B.4 Parabolic Equation
equation to be invariant under reparametrization, then the The right-hand side in (30) is invariant under reparame-
only possible choice for  is  ¼ Gð; x Þ in which Gðp; vÞ is trizations, that is, if ðx; tÞ ¼ ðyðx; tÞ; tÞ, then  satisfies
positively homogeneous of degree one in v 2 Rn . A possible 1
choice would be Gð; x Þ ¼ jx j, which leads us to the t ¼ Fvv ð; y Þyy þ yt y :
Gð; y Þ
evolution equation
Choose the parametrization so that
1  ?
t ¼ ðFv Þx  Fp : ð23Þ yt ðx; tÞ ¼ ½ðy  yy Þy
jx j y¼yðx;tÞ

This equation is equivalent with (5). for some scalar  > 0, which can depend on  and y .
422 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 30, NO. 3, MARCH 2008

The resulting equation for  is then REFERENCES

[1] S. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum, and A.
Fvv ð; y Þ Bð; y Þ Yezzi, “Conformal Curvature Flows: From Phase Transitions to
t ¼ þ y y  yy þ ; ð31Þ
Gð; y Þ Gð; y Þ Active Vision,” Archive of Rational Mechanics and Analysis, vol. 134,
no. 3, pp. 275-301, 1996.
where by definition [2] V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic Active
Contours,” Int’l J. Computer Vision, vol. 22, no. 11, pp. 61-79, 1997.
Bðp; vÞ ¼ v  Fpv ðp; vÞ  Fp ðp; vÞ: [3] D. Bao, S.-S. Chern, and Z. Shen, Introduction to Riemann-Finsler
Geometry. Springer, 2000.
As long as one chooses ð; y Þ > 0, and as long as Fvv is [4] M. Spivak, Introduction to Differential Geometry, vol. 2. Publish of
positive definite on fvg? , this equation is strictly parabolic. Perish, 1974.
[5] S. Angenent and M. Gurtin, “Anisotropic Motion of a Phase
A particular choice for G and  would be Interface. Well-Posedness of the Initial Value Problem and
Qualitative Properties of the Interface,” J. Reine und Angewandte
G ¼ jy j;  ¼ jy j2 ; Mathematik, vol. 446, pp. 1-47, 1994.
[6] M. Gage, “Evolving Plane Curves by Curvature in Relative
which leads to Geometries,” Duke J. Math., vol. 72, pp. 441-466, 1993.
[7] E. Pichon, G. Sapiro, and A. Tannenbaum, Segmentation of
  yy Diffusion Tensor Imagery, “Directions in Math. Systems Theory
t ¼ Fvv ð; TÞ þ T T  þ Bð; TÞ; ð32Þ
jy j2 and Optimization,” Lecture Notes in Computer Science, no. 286,
pp. 239-247, 2003.
where T ¼ y =jy j is the unit tangent vector. [8] E. Pichon, C. Westin, and A. Tannenbaum, “Hamilton-Jacobi-
Bellman Approach to High-Angular Resolution Diffusion Tracto-
graphy,” Proc. Int’l Conf. Medical Image Computing and Computer-
B.5 Numerical Scheme Assisted Intervention, pp. 180-187, 2005.
We should note that using the above, a simple approach can [9] E. Pichon and A. Tannenbaum, “Pattern Detection and Image
be employed to (32). For completeness, we sketch this here. Segmentation with Anisotropic Conformal Factor,” Proc. Int’l
Conf. Image Processing, 2005.
We set [10] G. Bellettini, “Anisotropic and Crystalline Mean Curvature Flow,”
A Sampler of Riemann-Finsler Geometry, pp. 49-82, Cambridge Univ.
ðjx; ktÞ ¼ jk Press, 2004.
[11] R. Kimmel and A. Bruckstein, “Regularized Zero-Crossings as
and discretize (32) as follows: Optimal Edge Detectors,” Int’l J. Computer Vision, vol. 53, no. 3,
pp. 225-243, 2003.
jkþ1  jk n o [12] A. Vasilevsky and K. Siddiqi, “Flux Maximizing Geometric
¼ Akj jþ1
kþ1
 2jkþ1 þ j1
kþ1
þ Bkj ð33Þ Flows,” IEEE Trans. Pattern Analysis and Machine Intelligence,
t vol. 24, no. 12, pp. 1565-1579, Dec. 2002.
in which the n  n matrices Akj are defined by [13] Y. Boykov, V. Kolmorgorov, D. Cremers, and A. Delong, “An
Integral Solution to Surface Evolution Pdes via Geo-Cuts,” Proc.
IEEE European Conf. Computer Vision, pp. 409-422, 2006.
Fvv ðjk ; Tkj Þ þ Tkj Tkj [14] V. Kolmorgorov and Y. Boykov, “What Metrics Can Be Approxi-
Akj ¼ 4 k k 2 mated by Geo-Cuts or Global Optimization of Length/Area and
jjþ1  j1 j
Flux,” Proc. IEEE Int’l Conf. Computer Vision, 2003.
[15] Y. Boykov and V. Kolmorgorov, “Computing Geodesics and
and one could define the unit tangents Tkj by Minimal Surfaces via Graph Cuts,” Proc. IEEE Int’l Conf. Computer
k k
Vision, pp. 26-33, 2003.
jþ1  j1 [16] J.-M. Morel and S. Solimini, Variational Methods for Image
Tkj ¼ k k
: Segmentation. Birkhauser, 1994.
jjþ1  j1 j
[17] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit
The vectors Bkj could be discretized by Surfaces. Springer, 2003.
[18] J. Sethian, Level Set Methods and Fast Marching Methods. Cambridge
Univ. Press, 1999.
Bkj ¼ Tkj  Fpv ðjk ; Tkj Þ  Fp ðjk ; Tkj Þ: [19] G. Sapiro, Geometric Partial Differential Equations and Image
Analysis. Cambridge Univ. Press, 2001.
The system of equations (33) is a tridiagonal vector- [20] E. Mortensen, B. Morse, W. Barrett, and J. Udupa, “Adaptive
valued system of equations. If one puts the components of Boundary Detection Using Live-Wire Two-Dimensional Dynamic
each vector jk in one long vector, then (33) can be written as Programming,” IEEE Proc. Computers in Cardiology, pp. 635-638,
1992.
a finite band system of equations, which can be solved very [21] J.N. Tsitsiklis, “Efficient Algorithms for Globally Optimal Trajec-
efficiently (OðnÞ operations per time step). tories,” IEEE Trans. Automatic Control, vol. 50, no. 9, pp. 1528-1538,
1995.
[22] S. Angenent, “Parabolic Equations for Curves on Surfaces. II.
ACKNOWLEDGMENTS Intersections, Blow-Up and Generalized Solutions,” Ann. Math.
(2), vol. 133, no. 1, pp. 171-215, 1991.
Allen Tannenbaum was supported by a Marie Curie Grant [23] J.A. Oaks, “Singularities and Self-Intersections of Curves Evolving
through the EU. This work was supported in part by grants on Surfaces,” Indiana Univ. Math. J., vol. 43, no. 3, pp. 959-981, 1994.
[24] M. DoCarmo, Riemannian Geometry. Springer, 2003.
from NSF, AFOSR, ARO, MURI, MRI-HEL as well as by a [25] W. Fleming and H. Soner, Controlled Markov Processes and Viscosity
grant from NIH (NAC P41 RR-13218) through Brigham and Solutions. Springer, 2003.
Women’s Hospital. This work is part of the National Alliance [26] H. Soner, “Dynamic Programming and Viscosity Solutions,” Proc.
Ann. AMS Meeting, 1999.
for Medical Image Computing (NAMIC), funded by the US [27] J. Sethian and A. Vladimirsky, “Ordered Upwind Methods for
National Institutes of Health through the NIH Roadmap for Static Hamilton-Jacobi Equations: Theory and Applications,”
Medical Research, Grant U54 EB005149. Information on the SIAM J. Numerical Analysis, vol. 41, no. 1, pp. 325-363, 2003.
[28] C. Kao, S. Osher, and Y. Tsai, “Fast Sweeping Methods for Static
National Centers for Biomedical Computing can be obtained Hamilton-Jacobi Equations,” Technical Report 03-75, CAM, Univ.
from http://nihroadmap.nih.gov/bioinformatics. of California, Los Angeles, 2003.
MELONAKOS ET AL.: FINSLER ACTIVE CONTOURS 423

[29] C. Kao, S. Osher, and J. Qian, “Lax-Friedrichs Sweeping Scheme John Melonakos received BS degree in elec-
for Static Hamilton-Jacobi Equations,” J. Computational Physics, trical engineering from Brigham Young Univer-
vol. 196, no. 1, pp. 367-391, May 2004. sity in 2004. He is currently working toward the
[30] M. Gage and R. Hamilton, “The Heat Equation Shrinking Convex PhD degree in electrical and computer engineer-
Plane Curves,” J. Differential Geometry, vol. 23, pp. 69-96, 1986. ing at the Georgia Institute of Technology. He is
[31] M. Grayson, “The Heat Equation Shrinks Embedded Plane Curves currently researching on mathematical methods
to Round Points,” J. Differential Geometry, vol. 26, pp. 285-314, 1987. for computer vision and image processing with
[32] P. Basser, J. Mattiello, and D. LeBihan, “MR Diffusion Tensor Professor Allen Tannenbaum. He has won
Spectroscopy and Imaging,” Biophysical J., vol. 66, pp. 259-267, 1994. several fellowships, including the US National
[33] S. Mori, B. Crain, V. Chacko, and P. van Zijl, “Three-Dimensional Science Foundation sponsored Technological
Tracking of Axonal Projections in the Brain by Magnetic Innovation: Generating Economic Results Fellowship, the NSF spon-
Resonance Imaging,” Annals of Neurology, vol. 45, no. 2, pp. 265- sored Hybrid Neural Microsystems Fellowship, the Sam Nunn Security
269, Feb. 1999. Program Fellowship, and the Georgia Tech President’s Fellowship. He
[34] T. Conturo, N. Lori, T. Cull, E. Akbudak, A. Snyder, J. Shimony, R. is a student member of the IEEE.
McKinstry, H. Burton, and M. Raichle, “Tracking Neuronal Fiber
Pathways in the Living Human Brain,” Proc. Nat’l Academy of Eric Pichon received the Diplome d’Ingenieur
Sciences, pp. 10422-10427, Aug. 1999. degree from Ecole Superieure d’Electricite (Su-
[35] C.-F. Westin, S.E. Maier, B. Khidhir, P. Everett, F.A. Jolesz, and R. pelec) in 2001 and the PhD degree in electrical
Kikinis, “Image Processing for Diffusion Tensor Magnetic and computer engineering from the Georgia
Resonance Imaging,” Proc. Int’l Conf. Medical Image Computing Institute of Technology in 2005. His research
and Computer-Assisted Intervention, pp. 441-452, 1999. interest includes partial differential techniques
[36] P. Basser, S. Pajevic, C. Pierpaoli, J. Duda, and A. Aldroubi, “In for computer vision with an emphasis on medical
Vivo Fiber Tractography Using DT-MRI Data,” Magnetic Resonance imaging. He is currently with GE Healthcare. He
in Medicine, vol. 44, pp. 625-632, 2000. is a member of the IEEE.
[37] P. Hagmann, T.G. Reese, W.-Y.I. Tseng, R. Meuli, J.-P. Thiran, and
V.J. Wedeen, “Diffusion Spectrum Imaging Tractography in
Complex Cerebral White Matter: An Investigation of the Centrum
Semiovale,” Proc. Int’l Soc. Magnetic Resonance in Medicine, 2004. Sigurd Angenent received the PhD degree in
[38] J.S. Campbell, “Diffusion Imaging of White Matter Fibre Tracts,” mathematics from the University of Leiden in
PhD dissertation, McGill Univ., 2004. 1986. He has been in the faculty of the
[39] L.C. Young, Calculus of Variations and Optimal Control Theory. department of mathematics at the University of
W.B. Saunders, 1969. Wisconsin since 1988 and works on nonlinear
[40] L. Ibanez, W. Schroeder, L. Ng, and J. Cates, “The ITK Software heat equations in differential geometry, biology,
Guide. Kitware,” technical report, http://www.itk.org/ItkSoft and image processing. He is a member of the
wareGuide.pdf, 2003. Royal Netherlands Academy of Sciences. He is
a member of the IEEE.

Allen Tannenbaum is a faculty member of the

School of Electrical and Computer Engineering
and the School of Biomedical Engineering of the
Georgia Institute of Technology. His research
interests include image processing, control, and
computer vision. He is a member of the IEEE.

. For more information on this or any other computing topic,

please visit our Digital Library at www.computer.org/publications/dlib.