Fuzzy Connectedness and Image Segmentation
JAYARAM K. UDUPA, SENIOR MEMBER, IEEE, AND PUNAM K. SAHA, MEMBER, IEEE
Invited Paper
Image segmentation—the process of defining objects in im-
ages—remains the most challenging problem in image processing
despite decades of research. Many general methodologies have
been proposed to date to tackle this problem. An emerging frame-
work that has shown considerable promise recently is that of fuzzy
connectedness. Images are by nature fuzzy. Object regions manifest
themselves in images with a heterogeneity of image intensities
owing to the inherent object material heterogeneity, and artifacts
such as blurring, noise and background variation introduced by
the imaging device. In spite of this gradation of intensities, knowl-
edgeable observers can perceive object regions as a gestalt. The
fuzzy connectedness framework aims at capturing this notion via a
fuzzy topological notion called fuzzy connectedness which defines
how the image elements hang together spatially in spite of their
gradation of intensities. In defining objects in a given image, the
strength of connectedness between every pair of image elements is
considered, which in turn is determined by considering all possible
connecting paths between the pair. In spite of a high combinatorial
complexity, theoretical advances in fuzzy connectedness have
made it possible to delineate objects via dynamic programming
at close to interactive speeds on modern PCs. This paper gives a
tutorial review of the fuzzy connectedness framework delineating
the various advances that have been made. These are illustrated
with several medical applications in the areas of Multiple Sclerosis
of the brain, magnetic resonance (MR) and computer tomographic
(CT) angiography, brain tumor, mammography, upper airway
disorders in children, and colonography.
Keywords—Digital topology, fuzzy connectedness, image seg-
mentation, medical imaging.
I. INTRODUCTION
A. Background
Two-dimensional (2-D) digital picture processing activi-
ties started in the early 1960s [1]. Very early on, Rosenfeld
[2] recognized the need and laid the foundation for digital
counterparts of key topological and geometric concepts that
have long been available for the continuous space (point sets)
such as neighborhood, connectedness, curves, boundaries,
Manuscript received July 15, 2002; revised March 17, 2003. This work
was supported by a DHHS grant NS37172.
The authors are with the Medical Image Processing Group, Department
of Radiology, University of Pennsylvania, Philadelphia, PA 19104 USA
(e-mail: jay@mipg.upenn.edu).
Digital Object Identifier 10.1109/JPROC.2003.817883
0018-9219/03$17.00 © 2003 IEEE
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
surroundness, distance, perimeter, and area. In particular, the
concept of connectedness he developed found widespread
use in image analysis such as in separating an already seg-
mented object from other irrelevant structures, and in sepa-
rating the component parts of a disconnected object that has
also been segmented already. This concept was defined in a
hard (crisp) fashion in the sense that a set of pixels was con-
sidered to be either connected or not connected. Similar con-
cepts were also developed for the three-dimensional (3-D)
space when 3-D image processing efforts commenced in the
1970s [3]–[5]. We shall see that when this definition is gen-
eralized to fuzzy subsets so as to consider a strength of con-
nectedness value, it has implications far beyond just the fuzzy
generalization of the concept of connectedness.
Why fuzzy? Images produced by any imaging device are
inherently fuzzy. This fuzziness comes from several sources.
Because of spatial and temporal resolution limitations,
imaging devices introduce blur. They also introduce noise
and other artifacts such as background intensity variation.
In addition to these factors, objects usually have material
heterogeneity, and this introduces heterogeneity in the
images. This is especially true in the biomedical context of
the internal organs of living beings. The combined effect of
all these factors is that object regions manifest themselves
with a heterogeneity of intensity values in acquired images.
The main rationale for taking fuzzy approaches in image
analysis is to try to address and handle the uncertainties
and intensity gradations that invariably exist in images as
realistically as possible.
It appears to us that the fuzzy treatment of the geometric
and topological concepts can be made in two distinct man-
ners in image analysis. The first approach is to carry out a
fuzzy segmentation of the given image first so that we have a
fuzzy subset of the image domain wherein every image ele-
ment has a fuzzy object membership assigned to it and then to
define the geometric and topological concepts on this fuzzy
subset. The second approach is to develop these concepts
directly on the given image, which implies that these con-
cepts will have to be integrated somehow with the process of
segmentation. Taking the first approach, Rosenfeld did some
early work [6], [7] in developing fuzzy digital topological
1649
and geometric concepts. The second approach pursued by us
and others has been mainly for developing the fuzzy connect-
edness concepts. Rosenfeld [6] defined fuzzy connectedness
by using a min-max construct on fuzzy subsets of 2-D picture
domains: the strength of connectedness of any path between
any two pixels and is the smallest fuzzy pixel membership
along the path and the strength of connectedness between
and is the strength of the strongest of all paths between
and . The process of obtaining the fuzzy subset in the first
place is indeed the act of (fuzzy) segmentation. Through the
introduction of a fundamental concept — a local fuzzy re-
lation on image elements called affinity [8] — it is possible
to develop the concept of fuzzy connectedness directly on
the given image and integrate it with, and actually utilize it
for facilitating, image segmentation. As we already pointed Fig. 1. Illustration of the ideas underlying generalized fuzzy
out, image elements have a gradation of intensities within the connectedness and relative fuzzy connectedness.
same object region. In spite of this graded composition, they
seem to hang together (form a gestalt) within the same object.
We argue that this “hanging-togetherness” should be treated
in a fuzzy manner, and an appropriate mathematical vehicle
to realize this concept is fuzzy connectedness. If hard (crisp)
connectedness were to be used, then it would require a pre-
segmentation. In other words, it cannot influence the result of
segmentation, except in separating a (hard) connected com-
ponent. Fuzzy connectedness when utilized directly on the
given image (via affinity), on the other hand, can influence
the segmentation result by the spatio-topological considera-
tion of how the image elements hang together, rather than by
strategies of feature clustering that have used fuzzy clustering
techniques [9] but have ignored the important topological
property of hanging-togetherness. It is widely acknowledged
that segmentation is the most crucial and difficult problem in
image processing. Any fundamental advance that empowers
the process of segmentation will take us closer to having a
grip on this problem. We believe that fuzzy connectedness is Fig. 2. Illustration of the ideas underlying iterative relative fuzzy
one such. connectedness.

B. Outline of the Paper
We have tried to make the paper as self contained as pos-
sible, leaving out a few details that are irrelevant for the main
discussion. These details may be obtained from the cited ref-
erences. By nature, this paper is mathematical. The mathe-
matical treatment may discourage some readers who are in-
terested in only the main ideas embodied in fuzzy connect-
edness. Furthermore, in a casual reading, the mathematical
details may submerge the central ideas that motivated the
mathematics. For these reasons, in the rest of this section,
we delineate the key ideas involved in fuzzy connectedness
in an intuitive fashion utilizing Figs. 1 and 2.
There are five key ideas underlying fuzzy connectedness
— those related to affinity, scale, generalized fuzzy connect-
edness, relative fuzzy connectedness, and iterative relative
fuzzy connectedness. The latter three will be abbreviated
GFC, RFC, and IRFC, respectively, and fuzzy connectedness
will be referred to as FC for ease of reference from now on.
Affinity is intended to be a local relation between every two
image elements and . That is, if and are far apart, the
strength of this relationship is intended to be zero. If the ele-
ments are nearby, the strength of their affinity, lying between
0 and 1, depends on the distance between them, on the homo-
geneity of the intensities surrounding them, and on the close-
ness of their intensity-based features to the feature values ex-
pected for the object. In determining the homogeneity com-
ponent and the features at and , the scale values at both
and are utilized. The scale at any element is the radius of
the largest hyperball centered at that element within which
the intensities are homogeneous under some criterion of ho-
mogeneity. In determining the affinity between and , all
elements in the scale regions of both and are consulted,
the reason being to make affinity (and subsequently, FC) less
susceptible to noise and to intensity heterogeneity within the
same object region.
Contrary to affinity, FC is intended to be a global fuzzy
relation. Its strength between any two elements, such as
and in Fig. 1, is the strength of the strongest of all possible
paths between and . A path such as is a sequence of
nearby elements, and the strength of a path is the smallest
affinity of pairwise elements along the path. Imagine Fig. 1
to be a gray-level image with three object regions , , and

1650 PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
 , and a background . With respect to , if an affinity
relation is chosen appropriately, we expect path to have a
higher strength than any path, such as , that treads across
multiple object regions.
In GFC, a fuzzy connected object of a certain strength
and containing an element is the largest set of image elements
containing that element such that the strength of connected-
ness between any two elements in that set is at least . The
theoretical framework demonstrates that this problem can be
solved via dynamic programming. In RFC, all co-objects of
importance (in Fig. 1, these are , , , ) are let to
compete for claiming membership of image elements via FC.
Again, referring to Fig. 1, let , , , be seed elements
in respective objects. An element such as is considered to
belong to object if the strength of connectedness of with
respect to ’s seed is greater than the strength of con-
nectedness of with respect to each of the other seeds. The
threshold required in GFC is thus eliminated in RFC. In
IRFC, the idea is to iteratively refine the competition rules for
RFC for different objects depending upon the results of the
previous iteration. Referring to Fig. 2, for an element such
as , its strength of FC from and from are likely to be
similar due to the blurring that takes place where and
come close together. In this case, the strongest path from to
is therefore likely to pass through the “core” of which
is indicated by the dotted curve in the figure. This core can
be detected first and then excluded from consideration in a
subsequent iteration for any path from to to pass through.
Then, we can substantially weaken the strongest path from
to compared to the strongest path from to which is still
allowed to pass through the core. This leads us to an iterative
strategy to separate and via RFC.
The organization of this paper is as follows. In Section II,
we describe some basic terms and the notations used
throughout the paper. In Section III, the theory of fuzzy
connectedness is outlined quite independent of its connec-
tion to image segmentation. In Section IV, its implications
in image segmentation are described. In Section V, a variety
of large applications that have utilized fuzzy connectedness
for image segmentation are described, and in Section VI,
our conclusions are stated with some pointers for future
work. We have tried to make this paper as self contained as
possible, leaving out only some details that may be irrelevant
for the main discussion. These details can be obtained from
the cited references.
II. BASIC NOTATIONS AND DEFINITIONS
Let be any reference set. A fuzzy subset [10] of
is a set of ordered pairs
where . is called the membership
function of in . The fuzzy subset is called nonempty
if there exists an such that ; otherwise
it is called the empty fuzzy subset of . We use to
denote an empty fuzzy subset and to denote an empty
hard set. The fuzzy intersection and fuzzy union between
two fuzzy subsets and of are defined as follows:
where
UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION
,
where .
A 2-ary fuzzy relation in is a fuzzy subset of
, where
. Since we are not interested in fuzzy
-ary relations for , we drop the qualifier “2-ary”
for simplicity. We use subscripted by the fuzzy subset
under consideration to denote the membership function
of the fuzzy subset. For hard subsets, will denote their
characteristic function. Let be any fuzzy relation in . is
said to be reflexive, if, , ; symmetric, if,
, ; transitive, if, ,
, . A fuzzy
relation is called a similitude relation in if it is reflexive,
symmetric and transitive. A binary (hard) relation in
is a subset of ; similar to a fuzzy relation, is
said to be reflexive, if, , ; symmetric,
if, , ; transitive, if,
, . A binary
relation that is reflexive, symmetric, and transitive is called
an equivalence relation. Similitudity and equivalence are
analogous concepts in fuzzy and hard subsets.
For any , let -dimensional Euclidean space be
subdivided into hypercuboids by mutually orthogonal fam-
ilies of parallel hyperplanes. Assume, with no loss of gen-
erality, that the hyperplanes in each family have equal unit
spacing so that the hypercuboids are unit hypercubes, and we
shall choose coordinates so that the center of each hypercube
has integer coordinates. The hypercubes will be called spels
(an abbreviation for “space elements”). When , spels
are called pixels, and when they are called voxels.
The coordinates of the center of a spel are represented by an
-tuple of integers, defining a point in . itself will be
thought of as the set of all spels in with the above inter-
pretation of spels, and the concepts of spels and points in
will be used interchangeably.
A fuzzy relation in is said to be a fuzzy adjacency
if it is both reflexive and symmetric. It is desirable that
be such that is a non increasing function of the dis-
tance between and . It is not difficult to see that the
hard adjacency relations commonly used in digital topology
[11] are special cases of fuzzy adjacencies. We call the pair
( ), where is a fuzzy adjacency, a fuzzy digital space.
Fuzzy digital space is a concept that characterizes the under-
lying digital grid system independent of any image-related
concepts. We shall eventually tie this with image related con-
cepts to arrive at fuzzy object related notions.
A scene over a fuzzy digital space ( ) is a pair
where for some ;
is the set of -tuples of positive integers; , called scene
intensity, is a function whose domain is , called the scene
domain, and whose range [ ] is a set of numbers (usually
integers). is a binary scene over ( ) if the range of
is {0,1}. Binary scenes usually result from a hard segmen-
tation of a given scene. When the result of segmentation is
a fuzzy subset of the scene domain, this fuzzy subset can be
equivalently represented by a scene wherein the scene inten-
sity represents the fuzzy membership value in [0, 1]. We call
1651
such scenes membership scenes over ( ). Unless spec- value along the path, and the degree of connectedness, de-
ified otherwise, whenever we refer to a scene in this paper, noted by (“ ” to denote Rosenfeld’s approach and
it encompasses all three types — gray-level scene, member- to distinguish it from more general approaches described
ship scene, and binary scene. We assume in this paper that later), as the strength of the strongest path between the two
. spels. Therefore
A nonempty path in from to is a sequence
of spels in ; is called the length (3.1)
of the path. Note that the successive spels in a path need not
be “adjacent” in the sense hard adjacency is usually defined He used the notion of the degree of connectedness to de-
in digital topology [11]. When they are indeed adjacent, we fine certain topological and geometrical entities which had
shall refer to this as a contiguous path. The set of all spels been previously defined for hard sets of spels in . Two
in the path is denoted by . An empty path in , spels are said to be connected if
denoted , is a sequence of no spels. Paths of length 2 will . He showed that this binary (hard) rela-
be referred to as links. The set of all paths in from to tion of connectedness in a fuzzy subset , denoted by ,
is denoted by . (Note that and are not necessarily is reflexive and symmetric, but not transitive, and conse-
distinct.) The set of all paths in , defined as , quently, is not an equivalence relation. Therefore, the
is denoted by . We define a binary join to operation on components defined by may not partition , i.e., they
, denoted “ ” as follows. For every two nonempty paths may have nonempty overlap. In the same paper, he defined
and the number of components and genus in a fuzzy subset and
[note that ], the membership value of a spel in a component using the no-
tions of plateaus, tops, and bottoms and associating three dif-
(2.1) ferent connected sets with each top. Also, he introduced the
(2.2) notions of separation and surroundedness in a fuzzy subset
(2.3) and showed that surroundedness describes a weak partial or-
(2.4) dering, i.e., the hard binary relation defined by surrounded-
ness is reflexive, antisymmetric, and transitive.
Note that the join to operation between and is Rosenfeld’s “degree of connectedness” [6] was further
undefined if . It is shown in [8] (Proposition 2.1) studied to understand the topological, geometrical, and mor-
that for any scene over any fuzzy digital space phological properties of fuzzy subsets [7]. However, there
( ), the following relation holds for any two spels was no indication in these works as to how this concept could
: be exploited to facilitate image segmentation. Dellepiane
and Fontana [12], [13] and Udupa and Samarasekera [8],
[14] were the first to suggest this use. Dellepiane et al.
(2.5) utilized Rosenfeld’s “degree of connectedness” to arrive at a
We define a binary relation greater than on , de- segmentation method (to be discussed in Section IV). Udupa
noted “ ”, as follows. Let and et al. [8] simultaneously introduced a different framework,
be any paths in . We say that bringing in a key concept of a local fuzzy relation called
if and only if we can find a mapping from the set of spels affinity on spels to capture local hanging-togetherness of
in to the set of spels in that satisfies all of the following spels. They showed how affinity can incorporate various
conditions: image features in defining fuzzy connectedness, presented a
1) only if . general framework for the theory of fuzzy connectedness,
2) There exists some , for which . and demonstrated how dynamic programming can be uti-
3) For all , whenever , lized to bring the otherwise seemingly intractable notion of
for some and . fuzzy connectedness into segmentation. They also advanced
Some examples follow: the theory of fuzzy connectedness considerably [15]–[19],
; . bringing in notions of relative fuzzy connectedness [17],
Trivially, every non empty path in is greater than the [19] which was further extended by Herman and De Car-
empty path in . valha [20], iteratively defined relative fuzzy connectedness
[18], [17], and addressing the virtues of the basic min–max
III. THEORY OF FUZZY CONNECTEDNESS construct used in fuzzy connectedness [16]. Affinity forms
the link between fuzzy connectedness and segmentation.
A. Early Works and History Saha et al. [15] studied the issue of how to construct ef-
The notion of the “degree of connectedness” of two spels fective affinities and the use of local scale for this purpose.
was first introduced by Rosenfeld [6], [7] in the context Aspects related to the computational efficiency of fuzzy
of studying the topology and geometry of fuzzy subsets of connectedness algorithms have also been studied [21], [22].
. He considered contiguous paths in and defined the The general framework of fuzzy connectedness based on
strength of connectedness of a contiguous path from affinity has been utilized in conjunction with other methods
to in a fuzzy subset of as the smallest membership of segmentation [23]–[28], particularly with deformable

1652 PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
models [23], [24], a Voronoi diagram based method [25], and affinities is the only valid choice for path strength (under a
level set methods [26]. The fuzzy connectedness methods set of reasonable assumptions stated in Axioms 1 – 4 below)
have been utilized for image segmentation extensively in so as to arrive at important results that allow us to compute
several applications, including Multiple Sclerosis lesion fuzzy connectedness via dynamic programming. We shall ex-
quantification [29]–[36], late life depression [37], [38], MR amine some details pertaining to this tenet now.
angiography [39], [40], CT angiography [41], [42], brain Let be a scene over a fuzzy digital space
tumor assessment [43], [44], breast density quantification ( ) and let be a fuzzy affinity in . A fuzzy -net
via mammograms [45], craniofacial surgery [46], and CT in is a fuzzy subset of with its membership function
colonography [47]. . assigns a strength to every path of
In the rest of this section, we shall delineate the general . For any spels , is called a strongest
fuzzy connectedness theory based on affinity and its exten- path from to if . The idea
sions to relative and iterative relative fuzzy connectedness. of -net is that it represents a network of all possible paths
We will present all pertinent details to understand the struc- between all possible pairs of spels in with a strength as-
ture of the theory but only state the main results as theorems. signed to each path.
For details including proofs, please refer to [8], [15]–[19]. Axiom 1: For any scene over ( ), for any affinity
and -net in , for any two spels , the strength
B. Generalized Fuzzy Connectedness of the link from to is the affinity between them; i.e.,
Fuzzy Affinity: Let be any scene over ( ). .
Any fuzzy relation in is said to be a fuzzy spel affinity Axiom 2: For any scene over ( ), for any affinity
(or, affinity for short) in if it is reflexive and symmetric. In and -net in , for any two paths ,
practice, for to lead to meaningful segmentation, it should implies that .
be such that, for any , is a function of 1) the Axiom 3: For any scene over ( ), for any affinity
fuzzy adjacency between and ; 2) the homogeneity of the and -net in , fuzzy -connectedness in is a fuzzy
spel intensities at and ; 3) the closeness of the spel inten- relation in defined by the following membership function.
sities and of the intensity-based features of and to some For any
expected intensity and feature values for the object. Further,
may depend on the actual location of and (i.e.,
is shift variant). A detailed description of a generic func-
tional form for and how to select the right affinity for a (3.2)
specific application are presented in Section IV. Throughout
this paper, with appropriate subscripts and superscripts will For fuzzy connectedness, we shall always use the upper case
be used to denote fuzzy spel affinities. form of the symbol used to represent the corresponding fuzzy
Note that affinity is intended to be a “local” relation. This affinity.
is reflected in the dependence of on the degree of Axiom 4: For any scene over ( ), for any affinity
adjacency of and . If and are far apart, will and -net in , fuzzy -connectedness in is a symmetric
be zero. How the extent of the “local” nature of is to be and transitive relation.
determined will be addressed in Section IV. We note that Axiom 1 says that, a link is a basic unit in any path, and
the concept of affinity is applicable to membership and bi- that the strength of a link (which will be utilized in defining
nary scenes also. If the scene is a membership scene, then path strength) should be simply the affinity between the two
affinity should depend perhaps only on adjacency and ho- component spels of the link. This is the fundamental way in
mogeneity of spel intensities. Affinity has meaning even in which affinity is brought into the definition of path strength.
binary scenes, wherein it should depend only on fuzzy adja- Note that, in a link in a path (unless it is a con-
cency for like-valued spels. tiguous path), and may not always be adjacent (in
Path Strength and Fuzzy Connectedness: Our intent is the sense “adjacency” is usually considered in hard digital
that fuzzy connectedness is a global fuzzy relation which as- topology) — that is, and may be far apart differing
signs a strength to every pair ( ) of spels in . It makes in some of their coordinates by more than 1. In such cases,
sense to consider this strength of connectedness to be the Axiom 1 guarantees that the strength of is deter-
largest of the strengths assigned to all paths between and mined by and not by “tighter” paths of the form
as originally suggested by Rosenfeld. (The physical analogy wherein the successive
one may consider is to think of and as being connected by spels are indeed adjacent. Since is by definition reflexive
many strings, each with its own strength. When and are and symmetric, this axiom guarantees that link strength is
pulled apart, the strongest string will break at the end, which also a reflexive and symmetric relation in . Axiom 2 guar-
should be the determining factor for the strength of connect- antees that the strength of any path changes in a non in-
edness between and .) However, it is not so obvious as creasing manner along the path. This property is sensible and
to how the strength of each path should be defined. Several becomes essential in casting fuzzy connected object tracking
measures based on the affinities of spel pairs along the path, (whether in a given membership scene or in a given scene) as
including their sum, product, and minimum, all seem plau- a dynamic programming problem. Axiom 3 says essentially
sible. Saha and Udupa [16] have shown that the minimum of that the strength of connectedness between and should be

UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION 1653

the strength of the strongest path between them. Its reason- A fuzzy -object of containing a set of seed
ableness has already been discussed. Finally, Axiom 4 guar- spels of is a fuzzy subset of whose membership function
antees that fuzzy connectedness is a similitude relation in . is
This property is essential in devising a dynamic program-
if
ming solution to the otherwise seemingly prohibitive com-
otherwise.
binatorial optimization problem of determining a fuzzy con- (3.8)
nected object. Using the above axioms, the following main is the support of , i.e., the union of the sup-
theorem was proved in [16]. ports of the fuzzy connected objects containing the individual
Theorem 1: For any scene over ( ), and seed spels of .
for any affinity and -net in , fuzzy -connectedness Several important properties of fuzzy connected objects
in is a similitude relation in if and only if have been established [8], [16]. The following theorem
gives a guidance as to how a fuzzy -object of
should be computed. It is not practical to use the definition
directly for this purpose because of the combinatorial
complexity. The following theorem provides a practical way
where is the path .
of computing .
Following the spirit of the above theorem, path strength
Theorem 2: For any scene over ( ), for
from now on will be defined as follows:
any affinity in , for any , for any objectness
function , and for any non empty set , the support
(3.3) of the fuzzy -object of containing equals

where is the path . For the remainder of this (3.9)

paper, we shall assume the above definition of path strength.
Fuzzy Connected Objects [8], [16]: Let be The above theorem says that, instead of utilizing
any scene over ( ), let be any affinity in , and let (3.5)–(3.8) that define the support of the fuzzy
be a fixed number in [0,1]. Let be any subset of . We -object containing and having to check the strength
shall refer to as the set of reference spels or seed spels and of connectedness of every pair ( ) of spels for each
assume throughout that . A fuzzy -object of seed , we need to check the maximum strength of
containing a seed spel of is a fuzzy subset of connectedness of every spel with respect to the seed
whose membership function is spels. This is a vast simplification in the combinatorial
complexity. as shown in (3.9) can be computed via
if dynamic programming, given , , , and , as described
(3.4) in Section IV.
otherwise.
The following theorem characterizes the robustness of
In this expression, is an objectness function whose do- specifying fuzzy -objects through sets of seed spels.
main is the range of and whose range is [0,1]. It maps im- Theorem 3: For any scene over ( ), for
aged scene intensity values into objectness values. For most any affinity in , for any , for any objectness func-
segmentation purposes, may be chosen to be a Gaussian tion , and for any non empty sets ,
whose mean and standard deviation correspond to the inten- if and only if and .
sity value expected for the object region and its standard de- The above theorem has important consequences in the
viation (or some multiple thereof). The choice of should practical utilization of the fuzzy connectedness algorithms
depend on the particular imaging modality that generated for image segmentation. It states that the seeds must be
and the actual physical object under consideration. (When a selected from the same physical object and at least one seed
hard segmentation is desired, (defined bellow) will must be selected from each physically connected region.
constitute the (hard) set of spels that represents the extent of High precision (reproducibility) of any segmentation algo-
the physical object and will simply be the characteristic rithm with regard to subjective operator actions (and with
function of .) is a subset of satisfying all regard to automatic operations minimizing these actions),
of the following conditions: such as specification of seeds, is essential for their practical
utility. Generally, it is easy for human operators to specify
(3.5) spels within a region in the scene corresponding to the same
physical object in a repeatable fashion. Theorem 3 guaran-
(3.6)
tees that, even though the sets of spels specified in repeated
trials (by the same or other human operators) may not be the
(3.7) same, as long as these sets are within the region of the same
physical object in the scene, the resulting segmentations will
is called the support of , i.e., a maximal be identical. It is important to point out that the intensity
subset of such that it includes and the strength of con- based connectedness method as proposed by Dellepiane et
nectedness between any two of its spels is at least . al. [12], [13] (see Section IV) fails to satisfy this robustness

1654 PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
 together among themselves relative to others. A spel such
as in the boundary between and will be grabbed by
that object with whom hangs together most strongly. This
mechanism not only eliminates the need for a threshold re-
quired in the GFC method but also offers potentially more
powerful segmentation strategies for two reasons. First, it al-
lows more direct utilization of the information about all ob-
jects in the scene in determining the segmentation of a given
object. Second, considering from a point of view of thresh-
olding the strength of connectedness, it allows adaptively
changing the threshold depending on the strength of connect-
edness of objects that surround the object of interest. The
formal definition of relative fuzzy connected objects along
with their important properties are presented in this section.
Fig. 3. Illustration used for defining relative fuzzy connectedness. For any spels in , define

(3.10)
property. Many other region growing algorithms [48] which
adaptively change the spel inclusion criteria during the The idea here is that and are typical spels specified in
growth process cannot guarantee this robustness property. “object” and “background”, respectively. Note that
Other interesting properties of fuzzy connectedness and , if .
of fuzzy connected objects have been studied in [8], [16]. It A fuzzy -object of a scene containing a spel
has been shown that the support of the fuzzy con- relative to a background containing a spel is the fuzzy
nected object monotonically decreases in size with subset of defined by the following membership function.
the strength of connectedness threshold while monotoni- For any ,
cally increases with the set of seed spels. It has also been if
shown that is connected (in the hard sense) if and (3.11)
otherwise.
only if the strength of connectedness between every two seed
spels is at least . For short, we will refer to as simply a relative fuzzy -ob-
ject of . The following proposition states the cohesiveness
C. Relative Fuzzy Connectedness [19], [17] of spels within the same relative fuzzy connected object.
Proposition 4: For any scene over ( ),
In the original fuzzy connectedness theory [8] as described
for any affinity in , and for any spels and in
in Section III-B, an object is defined on its own based on
such that ,
the strength of connectedness utilizing a threshold . The
key idea of relative fuzzy connectedness is to consider all (3.12)
co-objects of importance that are present in the scene and
to let them to compete among themselves in having spels if, and only if, .
as their members. Consider a 2-D scene composed of mul- Note that the above result is not valid if “ ” is used in
tiple regions corresponding to multiple objects as illustrated (3.10) instead of “ ”.
in Fig. 1. If is the object of interest, then the rest of the This proposition asserts the robustness of the relative fuzzy
objects , , and may be thought of as constituting the -object to the reference spel selected in the object region.
background as far as is concerned. With such a thinking, However, constancy of the -object with respect to reference
Fig. 3 shows just two objects , the object of interest, and spels specified in the background requires more constraints,
, the background, equivalent to the scene in Fig. 1. Al- as indicated by the following theorem.
though the theory for RFC can be developed for simultane- Theorem 5: For any scene over ( ), for
ously considering multiple objects [49], for simplicity, we any affinity in , and for any spels and in such
shall describe here the two-object case, but keeping in mind that
the treatment of object grouping mentioned above.
(3.13)
Suppose that the path shown in Fig. 3 represents the
strongest path between and . The basic idea in relative Note that the condition in (3.13) is sufficient (for
fuzzy connectedness is to first select reference spels and ), but not necessary. The necessary and sufficient condi-
, one in each object, and then to determine to which ob- tion is expressed in the following theorem.
ject any given spel belongs based on its relative strength of Theorem 6: For any scene over ( ), for
connectedness with the reference spels. A spel , for ex- any affinity in , and for any spels and in such
ample, would belong to since its strength of connected- that ,
ness with is likely to be greater than that with . This rela-
tive strength of connectedness offers a natural mechanism for
partitioning spels into regions based on how the spels hang (3.14)

UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION 1655

 The above two theorems have different implications in the
practical computation of relative fuzzy -objects in a given
scene in a repeatable, consistent manner. Although less spe-
cific, and therefore more restrictive, Theorem 5 offers practi-
cally a more relevant guidance than Theorem 6 for selecting
spels in the object and background so that the relative fuzzy
-object defined is independent of the reference spels.
An algorithm for determining relative fuzzy connected ob-
jects follows directly from algorithms for objects based on
generalized fuzzy connectedness as described in Section IV.
D. Iterative Relative Fuzzy Connectedness [18], [17]
The principle behind this strategy is to iteratively refine
the competition rules for different objects depending upon
the results of the previous iteration. Consider the situation il-
lustrated in Fig. 2 which demonstrates three objects ,
and . It is very likely that, for a spel such as ,
because of the blurring that takes place in those
parts where and come close together. In this case, the
strongest path from to is therefore likely to pass through
the “core” of which is indicated by the dotted curve in
the figure. This core (which is roughly defined in Sec-
tion III-C) can be detected first and then excluded from con-
sideration in a subsequent iteration for any path from to
to pass through. Then, we can substantially weaken the
strongest path from to compared to the strongest path from
to which is still allowed to pass through the core. This
leads us to an iterative strategy to grow from (and so com-
plementarily from ) to more accurately capture (and )
than if a single-shot relative connectedness strategy is used.
The phenomenon illustrated in Fig. 2 is general and may be
characterized in the following way. Most objects have a core
part, which is relatively easy to segment after specifying the
seed spel, and other diffused, subtle and fine parts that spread
off from the core, which pose segmentation challenges. Al-
though the latter seem to hang together fairly strongly with
the core from a visual perceptual point of view, because of the
ubiquitous noise and blurring, it is difficult to devise compu-
tational means to capture them as part of the same object by
using a single-shot strategy. The iterative strategy captures
these loose parts in an incremental and reinforcing manner.
This formulation is described in this section.
For any fuzzy affinity and any two spels , define
(3.15)
(3.16)
Note that is exactly the same as , defined in (3.10).
Assuming that and are defined for any positive
integer , and are defined as follows. For any
if
if or (3.17)
otherwise
(3.18)
1656
An iteratively defined fuzzy -object of a scene
containing a spel relative to a background con-
taining a spel is a fuzzy subset of defined by the mem-
bership function
if
(3.19)
otherwise.
For short, we will refer to simply as an iterative relative
fuzzy -object of . Below, we state a few important prop-
erties of iterative relative fuzzy connectedness.
Proposition 7: For any scene over ( ),
for any affinity in , for any spels and in , and for any
non negative integers , .
The above proposition states essentially that is non
contracting as iterations continue. The following proposition
states that and maintain their disjoincy at every
iteration for any and .
Proposition 8: For any scene over ( ),
for any affinity in , for any spels in , and for any
non-negative integers , .
The following theorem states an important property of it-
erative relative fuzzy -objects and shows their robustness
with respect to reference seed spels. It is analogous to The-
orem 5 on non iterative relative fuzzy -objects.
Theorem 9: For any scene over ( ), for
any affinity in , for any spels , and in such that
, and for any non negative integer , if
.
Algorithms for determining iterative relative fuzzy objects
are more complicated. See Section IV for a further discussion
on this topic.
IV. METHODS AND ALGORITHMS
A. Affinity Definition
As pointed out earlier, affinity is the link between the
topological concept of fuzzy connectedness and its utility
in image segmentation. As such, its definition determines
the effectiveness of segmentation. A detailed formulation of
affinity definition was presented in [15]. In the same paper,
it was first explicitly pointed out that, besides adjacency —
a purely spatial concept — affinity should consist of two
additional and distinct components — homogeneity-based
and object feature-based components. This formulation
can also be identified in the original affinity-based fuzzy
connectedness work [8] although this was not explicitly
stated. These two components are quite independent and
there exist certain dichotomies between them. The object
feature-based component does not capture the notion of
path homogeneity. To clarify this, consider two spels and
that are in the same object region but that are far apart.
Assume that there is a slow varying background intensity
component such as that often found in MR images due to
magnetic field inhomogeneities. Spels and are likely to
have very different scene intensities although they belong
to the same object. Nevertheless, one can find a contiguous
path from to in the scene domain such that the intensities
of each pair ( ) of successive spels in the path are very
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
similar. That is, and have a strong homogeneity-based periphery of a digital ball with center at and radius
affinity. Conversely, the homogeneity-based component is computed as follows:
alone cannot adequately capture the notion of a global
agreement to known object intensity characteristics. Again
consider a path from to along which spel intensity (4.26)
changes very slowly. The path may pass through different
object regions. Without an object feature-based component, where is a homogeneity function. As demonstrated in
this path will indicate a strong connectedness between [15], a zero-mean unnormalized Gaussian is generally a good
and and therefore will merge the different objects it choice for . The standard deviation parameter de-
passes through into one object. The following is a general pends on the overall noise level in the image. An automatic
functional form for as proposed in [8], [15] method to determine this parameter for a specific image is de-
scribed in [15]. The scale computation algorithm iteratively
(4.20) increases the ball radius by 1, starting from 1, and checks
for the fraction of the object containing that is con-
where and represent the homogeneity-based and tained in the ball. The first time when this fraction falls below
object-feature-based components of affinity, respectively. a threshold level, the ball is considered to have entered into
and may themselves be considered as fuzzy relations in a significantly different region. See [15] for details.
. The strength of relation indicates the degree of local In the rest of this section, we describe how scale is utilized
hanging togetherness of spels because of their similarities of in defining and . Let denote the digital ball de-
intensities. The strength of relation indicates the degree of fined by
local hanging togetherness of spels because of the similarity
of their feature value to some (specified) object feature. (4.27)
General constraints on the functional form of were de-
First, to define , consider any two spels such
scribed in [15] and the following examples were presented:
that . Consider any spels and
such that they represent the corresponding spels
(4.21) within and , that is . We will
(4.22) define two weighted sums and of differ-
(4.23) ences of intensities between the two balls as follows:

(4.24) if
otherwise
(4.25) (4.28)
if
A fundamental question that arises in defining affinity otherwise
(both and ) is how to determine the extent of the neigh- (4.29)
borhood of spels. (This question also arises in all image
processing operations that entail a local process.) If the and
intensities, or features derived from them, at the level of
the individual spels are considered for defining and ,
this definition will be sensitive to noise. The idea then is to
consult a neighborhood around each of and in defining
and such that the neighborhood represents
(4.30)
a region within the same object as in which or is situated.
“Scale” is a fundamental, well-established concept in
image processing [50]–[52]. The premise behind this
concept is to consider the local size of the object as the
neighborhood size in carrying out whatever local operations (4.31)
that are done on the image. Saha et al. [15] introduced a
new notion of location-specific object scale. Roughly, object and are window functions (such as a Gaussian). We
scale in at any spel is the radius of the largest note that the parameters of depend on and .
hyperball centered at which lies entirely in the same object The connection of the above equations to the homo-
region. Ironically, all this is done exactly for the purpose geneity-based affinity is as follows. There are two types
of defining the object in the first place and it appears that of intensity variations surrounding and — intra- and in-
object definition is needed first to define scale. A simple and terobject variations. The intraobject component is generally
effective algorithm is presented in [15] to estimate object random, and therefore, is likely to be near 0 overall. The
scale at each spel without explicit object definition but interobject component, however, has a direction. It either
based on the continuity of intensity homogeneity alone. To increases or decreases along the direction given by ,
estimate scale at a spel , a fraction of object on the and is likely to be larger than the intraobject variation. It is

UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION 1657

reasonable, therefore, to assume that the smaller of Finally, the intensity-based connectedness of with respect
and represents the intraobject component and the to the seed is expressed as another membership scene
other represents the combined effect of the two components. , where, for all
(Note that when the values of (respectively, ) are
small, (respectively, ) also becomes
small.) Note that, if there is a slow background component
of variation, within the small neighborhood considered,
this component is unlikely to cause a variation comparable
to the interobject component. This strategy leads us to the
following functional form for : A hard segmentation is obtained by thresholding . The
authors presented a non iterative algorithm using a tree ex-
(4.32) pansion mechanism for computing given . Note that
the above intensity based connectedness is sensitive to seed
selection as the seed intensity is used for computing , and
Note that represents the degree of subsequently for the connectedness .
local inhomogeneity of the regions containing and . Its The affinity-based approach of Udupa and Samarasekera
value is low when both and are inside an (homogeneous) [8], [14] not only overcomes this difficulty of dependence on
object region. Its value is high when and are in the vicinity seed points (Theorems 3, 5, 6, and 9) but also made it pos-
of (or across) a boundary. The denominator in (4.32) is a sible to extend the fuzzy connectedness concept to relative
normalization factor. and iterative relative fuzzy connectedness (with much rele-
In the formulation of , instead of considering directly the vance in practical image segmentation) and to demonstrate
spel intensities , we will consider a filtered version of it (Theorem 2) that the well-developed computational frame-
that takes into account the ball at defined by work of dynamic programming [53] can be utilized to carry
out fuzzy connectedness computations efficiently.
(4.33) In this section we describe two algorithms [8] for ex-
The filtered value at any is given by tracting a fuzzy -object containing a set of spels in
a given scene for a given affinity in , both based on
dynamic programming, and another for computing iterative
(4.34) relative fuzzy connected object. The first algorithm, named
( -fuzzy object extraction for multiple seeds),
extracts a fuzzy -object of of strength generated by
where (another window function) and its parameters de- the set of reference spels. In this algorithm, the value of
pend on . The functional form of is given by (4.35), is assumed to be given as input and the algorithm uses
shown at the bottom of the page, where and are in- this knowledge to achieve an efficient determination of the
tensity distribution functions for the object and background, -object. The second algorithm, named , out-
respectively. See [15] for details on how the parameters of puts what we call a -connectivity scene
these and other functions can be estimated for a given appli- of generated by the set of reference spels defined by
cation. .
Algorithm terminates faster than
B. Algorithms for two reasons. First, produces the hard set
Dellepiane et al. [12], [13] used the formulation of fuzzy based on (3.9). Therefore, for any spel , once we find a
connectedness proposed by Rosenfeld [6], [7] for image seg- path of strength or greater from any of the reference spels to
mentation. Briefly, their method works as follows. Let , we do not need to search for a better path upto , and hence,
be a given scene and let be the maximum intensity can avoid further processing for . This allows us to reduce
over the scene. A membership scene is first computation. Second, certain computations are avoided for
created, where, for all , . Let de- those spels for which .
note a seed spel; another membership scene Unlike , computes the best path
is computed from and , where, for all from the reference spels of to every spel in . Therefore,
every time the algorithm finds a better path upto , it mod-
ifies the connectivity value at and subsequently processes

if and
if (4.35)
otherwise

1658 PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
other spels which are affected by this modification. The algo- 5. find ;
rithm generates a connectivity scene of . 6. if then
Although, terminates slower, it has a practical 7. set ;
advantage. After the algorithm terminates, one can interac- 8. push all spels such that
tively specify and thereby examine various -objects and to ;
interactively select the best . The connectivity scene has in- endif;
teresting properties relevant to classification and in shell ren- endwhile;
dering and manipulation [54] of -objects. 9. output the connectivity scene ;
end
Algorithm
Input: , , , , and as defined in
Both algorithms work in an iterative fashion, and in each
Sections II & IV.
iteration in the while-do loop, they remove exactly one spel
Output: as defined in Section IV.
from the queue and check if the strength of connectedness
Auxiliary Data Structures: An -D
at that spel needs to be changed. In case a change is needed,
array representing a temporary scene
the change is passed on to the rest of the scene through the
such that corresponds to
neighbors by entering them into . It may be noted that both
the characteristic function of ,
algorithms start after initialization of strengths of connect-
and a queue of spels. We refer to
edness at the seed spels. Also, note that the strength of con-
the array itself by for the purpose
nectedness at any spel never reduces during runtime of these
of the algorithm.
algorithms. It was shown [8], [16] that both algorithms ter-
begin
minate in a finite number of steps, and when they do so, they
1. set all elements of to 0 except
produce the expected results. Recently, studies [21], [22] on
those spels which are set to 1;
the speed up of these algorithms have been reported. A pre-
2. push all spels such that for
liminary effort was made by Carvalho et al. [21] by using
some to ;
Dijkstra’s algorithm [55] in place of the dynamic program-
3. while is not empty do
ming algorithm suggested in [8]. Based on 2-D phantoms
4. remove a spel from ;
and two 2-D MR slice scenes, they suggested that a seven-to-
5. if then
eightfold speed-up in fuzzy connectedness object computa-
6. set ;
tion can be achieved. Nyul and Udupa [22] made an exten-
7. push all spels such that
sive study on the speed of a set of 18 algorithms utilizing
to ;
a variety of 3-D scenes. They presented two groups of al-
endif;
gorithms — label-correcting and label-setting — and a va-
endwhile;
riety of strategies of implementation under each group. They
8. create and output by as-
tested these algorithms in several real medical applications
signing the value
and concluded that there is no “always optimal” algorithm.
to all for which and 0 to
The optimality of an algorithm depends on the input data as
the rest of the spels;
well as on the type of affinity relations used. In general, it
end
was found that, for blob-like objects with a spread-out distri-
bution of their intensities or other properties (such as brain
Algorithm tissues in MRI), label setting algorithms with affinity-based
Input: , , and as defined Sec- hashing outperform others, while, for more sparse objects
tions II & IV. (vessels, bone) with high contrast (MRA, CTA), label setting
Output: A scene representing algorithms with geometric hashing are more speed-efficient.
the -connectivity scene of gen- Also, they demonstrated that using the right algorithms on
erated by . fast hardware (1.5 GHz Pentium PC), interactive speed (less
Auxiliary Data Structures: An -D array than one second per 3-D scene) of fuzzy object segmentation
representing the connectivity scene is achievable.
and a queue of spels. We Given , an algorithm for relative fuzzy con-
refer to the array itself by for the nectedness is easy to realize. First, the connectivity scenes
purpose of the algorithm. and are computed using
begin for seeds and , respectively. Subsequently,
1. set all elements of to 0 except a spel is included in the support of the fuzzy
those spels which are set to 1; connected object of relative to the background containing
2. push all spels such that, for if . Iterative relative fuzzy connected-
some , to ; ness is an iterative method; the connectivity scene
3. while is not empty do is computed once and in each iteration, the connec-
4. remove a spel from ; tivity scenes is computed for seed spel

UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION 1659

and affinity as defined in (3.17) in each iteration . The under the application. The second task consists of assessing
algorithm for computing iterative relative fuzzy connected- the precision, accuracy, and efficiency of the complete
ness [17], [18] is presented below. method in the application [56]. Precision here refers to the
reliability (reproducibility) of the complete method taking
Algorithm into consideration all subjective actions that enter into the
Input: , , spels and , and entire process including possible variations due to repeat
as defined in Section II. acquisition of the scenes for the same subject on the same
Output: Iteratively defined fuzzy -ob- and different scanners, and any operator help and input
ject containing relative to a required in segmentation. Accuracy relates to the agreement
background containing . of the segmented result to truth. Efficiency refers to the
Auxiliary Data Structures: The -connec- practical viability of the segmentation method. This has
tivity scene of in ; the to consider pure computational time and the time required
-connectivity scene of in producing any operator help needed in segmentation.
in . The above two tasks have been accomplished under each
begin application (except evaluative studies for the last application
0. compute using with the which are very different in their requirement from other
input as , , and ; applications). For reasons of space, we shall not present
1. set and ; here details of either the engineering or the evaluation
2. set , and ; studies, but refer the reader to references given under each
3. while is true do application. In the rest of this section, we give a cursory
4. set , , and description of each of the seven application areas outlining
; the segmentation need and methodology.
5. compute using with the
A. Multiple Sclerosis (MS)
input as , , and ;
6. set ; Multiple Sclerosis (MS) is an acquired disease of the
7. for all do central nervous system characterized primarily by multifocal
8. if then inflammation and destruction of myelin [57]. Inflamma-
9. set ; tion and edema are accompanied by different degrees of
10. increment by 1; demyelination and destruction of oligodendrocytes, and
11. for all do may be followed by remyelination, axonal loss and/or
12. set ; gliosis. The highest frequency of MS occurs in northern
endfor; and central Europe, Canada, the United States, and New
13. else set ; Zealand and South of Australia [58]. In the US, it affects
endif; approximately 350 000 adults and stands out as the most
endfor; frequent cause of nontraumatic neurologic disability in
14 if then set young and middle-aged adults. In its advanced state, the
; disease may severely impair the ambulatory ability and
15. increment by 1; may even cause death. The most commonly used scale
endwhile; to clinically measure disease progression in MS is the
15. output ; expanded disability status scale (EDSS) introduced by
end Kurtzke [59]. The clinical quantification of disease severity
is subjective and sometimes equivocal. The development of
new treatments demands objective outcome measures for
V. APPLICATIONS relatively short trials. Therefore, MR imaging has become
In this section, we will describe seven applications one of the most important paraclinical tests for diagnosing
in which the fuzzy connectedness algorithms have been MS and in monitoring disease progression in MS.
utilized and tested extensively. It is one thing to have a Various MRI protocol have been used in studying MS [34].
powerful segmentation framework and quite another matter At present they include dual-echo T2-weighted imaging,
to put it to practical use, such as in clinical radiology, for T1-weighted imaging with and without Gadolinium
routinely processing 100s and 1000s of scenes. The latter administration, magnetization transfer imaging, diffu-
usually involves considerable work in 1) engineering the sion-weighted imaging, and MR spectroscopy. The different
proper utilization of the framework, and 2) evaluating the protocols convey information about different aspects of
efficacy of the method in each application. For the fuzzy the disease. Current MRI-based research on MS is focused
connectedness framework, the first task entails the determi- on 1) characterizing the natural course of the disease and
nation of the affinity relation appropriate for the application, to distinguish among disease subgroups through images,
choosing one among generalized, relative, and iterative and 2) characterizing treatment effects to assess different
relative fuzzy connectedness strategies, and devising other treatment strategies. In achieving these goals, several image
auxiliary processing and visualization methods needed processing operations are utilized, most crucial among them

1660 PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
Fig. 4. Top row: A slice of a T2- and proton-weighted scene of an MS patient’s head and the
segmented brain parenchymal region. Below row: the segmented CSF region and the brain lesions
(which appear hyper intense in both T2 and PD scenes).
being segmentation of MR brain images in the different pro-
tocols into different tissue regions and lesions. Our efforts in
this direction utilizing fuzzy connectedness and the ensuing
clinical results are summarized in [34]. Fig. 4 shows an
example of a T2- and proton-density-weighted scene of an
MS patient’s head and the segmented tissue regions.
A somewhat similar problem from the image processing
perspective arises in the study of late-life depression [37],
[38]. The aim here is to understand the disease and its treat-
ment effects in the case of white matter lesions occurring in
elderly subjects due to age and the associated effects on the
function of their brain. In MS (and other neurological ap-
plications), a variety of MRI protocols are utilized. The ac-
tual imaging parameters used in these protocols vary among
institutions. In spite of numerous brain MR image segmen-
tation methods developed during the past 15 years, none of
them is capable of handling variations within the same pro-
tocol, and much less, the variation among protocols. What we
need is a segmentation “workshop” wherein a protocol-spe-
cific segmentation method can be quickly fabricated. In [60],
one such workshop is described utilizing the fuzzy connect-
edness framework.
B. MR Angiography
MR angiography (MRA) plays an increasingly important
role in the diagnosis and treatment of vascular diseases.
Many methods are available to acquire MRA scenes; the
most common is perhaps acquisition with an intravascular
contrast agent such as Gadolinium. An image processing
UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION
challenge in this application is how to optimally visualize
the imaged vascular tree from the acquired scene. In the
maximum intensity projection (MIP) method commonly
used for this purpose, the maximum voxel intensity encoun-
tered along each projection ray is assigned to a pixel in the
viewing plane to create a rendition as illustrated in Fig. 5.
Although, the MIP method does not require prior segmen-
tation of the scene, it obscures the display with clutter
due to high intensities coming from various artifacts as
shown in Fig. 5. By segmenting those aspects of the vessels
containing bright intensities via fuzzy connectedness, the
clutter can be mostly eliminated as illustrated in Fig. 5 [39].
Since the intensity of the image varies over the scene (due
to various reasons including magnetic field inhomogeneities
and different strengths of the contrast agent in different
parts of the vessels), simple thresholding and connectivity
analysis fail to achieve a proper segmentation of the vessels.
Recently, new imaging techniques with contrast agents
that stay in the blood much longer that cause far less extrav-
essation than Gadolinium have been devised [61]. Although,
these agents stay in the blood longer and hence provide a
more uniform enhancement of even significantly thinner ves-
sels, their side effect is that they enter into venous circula-
tion also because of the long decay time. Consequently, both
the arteries and veins enhance in the acquired scenes and it
becomes very difficult to mentally separate the arterial tree
from the venous tree in a display such as the one created by
a MIP method (see Fig. 6) [40]. Iterative relative fuzzy con-
nectedness can be effectively utilized in this situation [40] to
separate the arterial tree from the venous from a scene that is
1661

(a)
(b)
Fig. 5. A slice of an MRA scene acquired with Gadolinium
administration, a MIP rendition of this scene which shows much
clutter, and a MIP rendition of the region segmented by using fuzzy
connectedness.
created by segmenting the entire vascular tree first by using
generalized fuzzy connectedness. This is illustrated in Fig. 6.
C. CT Angiography
Currently, in the study of vascular diseases, CT angiog-
raphy [42] is a serious contender with MRA to eventually re-
place conventional X-ray angiography. A major drawback of
CTA, however, has been the obscuration of vessels by bone
in the 3-D display of CTA images. The problem is caused by
blood as well as bone assuming high and comparable inten-
sity in the CTA scene and often by vessels running close to
osseous structures. Partial volume and other artifacts make it
really difficult to suppress bone in the latter situation. This
scenario is particularly true for cerebral vessels since bone
surrounds the brain and vessels come close to the inner sur-
face of the calvarium at many sites. Interactive clipping of
bone in three dimensions is not always feasible or would
produce unsatisfactory results and slice-by-slice outlining is
1662
(a)
(b)
(c)
Fig. 6. (a) An MIP rendition of an MRA scene acquired with
the blood-pool contrast agent MS325. (b) The entire vascular
tree segmented via generalized fuzzy connectedness and volume
rendered. (c) The arteries and veins separated by using iterative
relative fuzzy connectedness. See Fig. 12(a) for color illustration.
simply not practical since most studies involve 100s of slices
comprising a CTA scene.
A method employing fuzzy connectedness has been pro-
posed in [41] for the suppression of osseous structures, and
its application in neurosurgery, particularly for managing
cerebral aneurysms, is described in [42]. The method uses
both generalized and iterative relative fuzzy connectedness
along with some morphological operations to separate the
vessels from osseous structures. Particularly challenging in
this context are the suppression of thin bones and of bones
juxtaposed close to vessels. Fig. 7 shows an example of
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003

(a)
(b)
Fig. 7. An MIP rendition of CTA of the head of a subject, and
an MIP rendition after bone has been suppressed via relative fuzzy
connectedness.
a MIP rendition of a cerebral CTA scene before and after
suppression of bones by using the method.
D. Brain Tumor
The management of patients with brain tumor (particularly
gliomas) is at present difficult in the clinic since no objective
UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION
information is available as to the morphometric measures of
the tumor. Multiprotocol MR images are utilized in the radi-
ologic clinic to determine the presence of tumor and to sub-
jectively discern the extent of the tumor [43]. Both subjective
and objective quantification of brain tumor are difficult due
to several factors. First, most tumors appear highly diffused
and fuzzy with visually indeterminable borders. Second, the
tumor pathology has different meaning in images obtained by
using different MRI protocols. Finally, a proper definition of
what is meant by “tumor” in terms of a specific pathology and
a determination of what definition is appropriate and useful
for the task on hand are yet to be determined. In the pres-
ence of these difficulties, we arrived at the following oper-
ational definition for the purpose of quantification. We seek
two measures to characterize the size of the tumor, the first
being the volume of the edema and the second being the ac-
tive part of the tumor. Edema is defined to be constituted by
the regions of high signal intensity in scenes acquired via
the FLAIR sequence. Active aspects of the tumor are con-
sidered to be formed of regions that have high signal inten-
sity in T1-weighted scenes acquired with Gadolinium ad-
ministration but not in T1-weighted scenes acquired without
Gadolinium.
Our approach to derive these measures from the MR
scenes is described in [44]. Briefly, it consists of applying
the generalized fuzzy connectedness method to the FLAIR
scene to segment the hyper-intense regions and estimating
the volume. Further, the two T1 scenes are registered, then
one scene is subtracted from the other, and the hyper-intense
region in the subtracted scene is segmented and its volume
is computed. The reason for subtracting the scenes is to
remove any postsurgical scar/blood appearing as hyper-in-
tense regions in the T1 scenes (without Gadolinium). Fig. 8
shows an example of this application.
E. Breast Density
Breast density as measured from the volume of dense
tissue in the breast is considered to indicate a risk factor
for breast cancer [62]. X-ray mammography is one of the
most commonly used imaging modalities for breast cancer
screening. Estimation of breast density from mammograms
is therefore a useful exercise in managing patients at high
risk for breast cancer. It may also be useful in assessing how
this risk factor is affected by various treatment procedures
such as hormone therapy.
In the Wolfe classification scheme [63], [64], a mammo-
grapher assigns a grade of density based on a visual inspec-
tion of the mammogram. This method, besides being very
rough and subjective, may have considerable inter and intra
reader variations and variation from one projective image to
another of the same breast. We have developed a method
utilizing generalized fuzzy connectedness to segment a dig-
itized mammogram into dense fibroglandular and fatty re-
gions and to provide several measures of density [49]. The
fuzzy connectedness framework seems to be especially suit-
able in this application since the dense regions appear fuzzy
with a wide range of scene intensities. Fig. 9 shows an ex-
ample.
1663
 (a)
(d)
Fig. 8. (a)–(c) A slice each of a FLAIR scene, a T1-weighted scene, and a T1 with Gadolinium
scene. (d)–(f) A surface rendition of brain plus edema segmented from the FLAIR scene; brain plus
edema and enhancing aspect of tumor segmented from the two T1 scenes; edema plus enhancing
tumor volume rendered. See Fig. 12(b)–(d) for color illustrations.
F. Upper Airway in Children
MRI has been used to visualize the upper airway of adult
humans. This enables imaging of the air column as well as
of the surrounding soft tissue structures including lateral
pharyngeal wall musculature, fat pads, tongue and soft
palate [65]. With detailed images of the upper airway and its
nearby structures, researchers can now use image processing
methods to efficiently segment, visualize, and measure these
structures. Measurements of upper airway geometry and the
architecture of the surrounding structures can be used to
investigate the reasons for obstructive sleep apnea (OSA),
and to assess the efficacy of medical treatment [65].
Similarly, MRI has recently been introduced to study the
upper airway structure of children during growth and devel-
opment and to assess the anatomic role in pathological con-
ditions such as OSA [66]. MRI is particularly useful in chil-
dren since it can delineate the lymphoid tissues surrounding
the upper airway very accurately. These tissues, the tonsils
and adenoids, figure predominately in childhood OSA and
are a common cause for pediatric medical visits and surgery.
Our methodology [67] uses axial T2 images. We have de-
veloped a PC-based system that runs under the Linux op-
erating system (which also runs on other Unix platforms)
for upper airway segmentation and measurement. General-
ized fuzzy connectedness is used for segmenting the upper
airway and adenoide and palatine tonsils. The system per-
forms all tasks from transforming data to giving a final result
1664
(b) (c)
(e) (f)
report. To achieve accurate quantification and assure quality,
the system allows clinicians to interact with critical data pro-
cessing steps as minimally as possible. Fig. 10 shows an ex-
ample.
G. CT Colonoscopy
An established imaging method for detecting colon
cancer has been X-ray projection imaging. A more
fool-proof method has been the endoscopic examination
through a colonoscopy that is commonly employed in
the clinic for screening process. Recently, in view of the
invasiveness and attendant patient discomfort of this latter
procedure, methods utilizing tomographic images have
been investigated [68]. Dubbed virtual colonoscopy or CT
colonoscopy, the basic premise of these methods is to utilize
the 3-D information available in a CT scene of the colon,
acquired upon careful cleaning and preparation of the colon,
to produce 3-D displays depicting views that would be
obtained by an endoscope. Since full 3-D information is
available, displays depicting colons that are cut or split open
or unraveled in other forms can also be created.
In the system we have developed [47], generalized fuzzy
connectedness is used for segmenting the colon. A 3-D
display of the entire colon is then created and the user
selects several points along the colon on the outer surface
so displayed to help the system create a central line running
through the colon. Subsequently, along this central line view
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
 (a)
Fig. 9. A digitized mammogram; the fuzzy connectivity scene of the dense region; and the
segmented binary region.
points are selected at uniform intervals, and at each such
point, three renditions are created of the inner wall. These
renditions represent the views of the inner wall of the colon
that would be obtained by viewing the wall surface enface in
three directions that are 120 apart. The speed of our surface
rendering method is high enough to produce in real-time the
set of three views that will be obtained as the view point
moves along the central line automatically. Fig. 11 shows an
example.
VI. CONCLUDING REMARKS
Although the idea of fuzzy connectedness existed in the lit-
erature since the late 1970s [6], its full-fledged development
and application to image segmentation started in the 1990s
[14]. Apart from generalizing the well known concept of hard
connectedness to fuzzy subsets, fuzzy connectedness allows
capturing the spatio-topological concept of hanging-togeth-
erness of image elements even in the presence of a grada-
tion of their intensities stemming from natural material het-
erogeneities, blurring and other imaging-phenomenonrelated
artifacts. In this paper, we have examined in a tutorial fashion
the different approaches that exist for defining fuzzy connect-
edness in fuzzy subsets of the discrete space as well as di-
rectly in given scenes. The latter, in particular, allow us to tie
the notion of fuzzy connectedness with scene segmentation,
leading to an immediate practical utilization of fuzzy con-
nectedness in scene analysis. Starting from the fundamental
UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION
(b) (c)
notion of fuzzy connectedness, we have demonstrated how
more advanced concepts such as relative and iterative rela-
tive fuzzy connectedness can be developed which have prac-
tical consequences in scene segmentation. We have also ex-
amined seven major medical application areas to which the
fuzzy connectedness principles have been applied for delin-
eating objects.
Fuzzy connectedness has several interesting character-
istics that enhance its effectiveness in scene segmentation.
The scale-based versions of fuzzy connectedness make
it less sensitive to noise. It is also possible to consider
texture in the definition of affinity to arrive at connectedness
and segmentation of regions in terms of texture. We note
that fuzzy connectedness is to some extent immune to a
slow varying component of background scene intensity as
demonstrated in [14]. This is not difficult to explain. Such
slow varying components do not affect affinities much,
as such the strengths of connectedness of paths are not
significantly affected, and therefore, so also the overall
fuzzy connectedness.
Many methods exist for fuzzy clustering of feature vec-
tors in a feature space [69], [70]. It would be worth investi-
gating if the added hanging-togetherness character of fuzzy
connectedness can lead to improved clustering strategies in
this area. Another direction for future research is to use fuzzy
connectedness, relative and iterative relative fuzzy connect-
edness on binary scenes and on membership scenes for sep-
1665
 (a)
(d)
(g)
Fig. 10. A slice of a T2 scene of the upper airway region of a child, the corresponding segmented
airway and adenoid and palatine tonsils, and a 3-D rendition of the three structures. See Fig. 12(f) for
color illustration.
arating components and for doing morphological operations
on them. The aspect of the size of the object at various spels
can be captured via scale values and separation or connected-
ness of binary objects can be analyzed via fuzzy connected-
ness of scale scenes. It should be feasible to directly obtain
medial line and skeletal representations of objects from bi-
nary scenes by using fuzzy connectedness on scale scenes.
The iterative relative fuzzy connectedness theory and algo-
rithms currently available [17], [18] are for only two objects.
Considering the fact that most scenes contain more than two
objects, for these algorithms to be useful, they and the asso-
ciated theory should be extended to multiple objects.
One of the current limitations (stemming from a theoret-
ical requirement) in the design of affinities [49] is that their
functional form should be shift-invariant over the whole
scene domain, although there are ways of tailoring affinities
1666
(b) (c)
(e) (f)
to variations in the characteristics of different object regions
(see [49]). It will be practically useful to investigate ways to
allow different functional forms for different object regions
and their theoretical and practical consequences.
A fundamental question arises in relative fuzzy connect-
edness as to how to do object groupings. Even when multiple
objects are considered as in Fig. 1, each object region may
consist of a group of objects. From the consideration of
the effectiveness of segmentation, certain groupings may
be better than others or even better than the consideration
of all object regions that may exist in the scene. Should
the object groupings be considered based on separability?
That is, the scene domain is first conceptually divided into
a set of most separable groups of objects. These groups
are then segmented via RFC (or IRFC). Next, within each
group, further conceptual groupings are made based on
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
Fig. 11. A surface rendition of the whole colon with the central line shown. Three enface
views of the colon wall obtained from a view point situated on the central line with the viewing
directions that are 120 apart.

Fig. 12. (a)–(e) Color illustrations of Figs. 6(c), 8(d)–(f) and 10(g), respectively.

separability, and each such sub group is segmented via
RFC. This process is continued until the desired objects
are reached and segmented. Such a hierarchical RFC (or
IRFC) has interesting theoretical, computational, and appli-
cation-specific questions to be studied.
Fuzzy connectedness in its present form attempts to ex-
tract as much information as possible entirely from the given
scene. It, however, does not attempt to use in its framework
prior shape and appearance knowledge about object bound-
aries to be segmented. These entities have been shown to be

UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION 1667

useful in scene segmentation [71], [72]. When information is
missing in certain segments of the boundary, fuzzy connect-
edness often goes astray establishing connection with other
objects surrounding the object of interest. Ways to incorpo-
rate prior boundary shape appearance knowledge into the
fuzzy connectedness framework may mitigate this problem.
This is a challenging and worthwhile direction to pursue. The
consideration of multiple objects in the model and utilization
of such models in conjunction with multiple object relative
fuzzy connectedness opens many new avenues for segmen-
tation research.
ACKNOWLEDGMENT
The authors are grateful to Dr. T. Lei and J. Liu for Figs. 6,
8, and 10 and to Dr. H. Eisenberg for data pertaining to
Fig. 11.
REFERENCES
[1] R. L. Grimsdale, F. H. Summer, C. J. Tunis, and T. Kilburn, “A
system for automatic recognition of patterns,” Proc. Inst. Elect. Eng.,
vol. 106B, pp. 210–221, 1959.
[2] A. Rosenfeld, “Connectivity in digital pictures,” J. Assoc. Comput.
Machinery, vol. 17, pp. 146–160, 1970.
[3] E. Artzy, G. Frieder, and G. T. Herman, “The theory, design, imple-
mentation, and evaluation of a three-dimensional surface detection
algorithm,” Comput. Graph. Image Processing, vol. 15, pp. 1–24,
1981.
[4] A. Rosenfeld, “Three dimensional digital topology,” Inform. Contr.,
vol. 50, pp. 119–127, 1981.
[5] J. K. Udupa, S. N. Srihari, and G. T. Herman, “Boundary detection
in multidimensions,” IEEE Trans. Pattern Anal. Machine Intell., vol.
4, pp. 41–50, 1982.
[6] A. Rosenfeld, “Fuzzy digital topology,” Inform. Contr., vol. 40, no.
1, pp. 76–87, 1979.
[7] , “The fuzzy geometry of image subsets,” Pattern Recogn. Lett.,
vol. 2, pp. 311–317, 1991.
[8] J. K. Udupa and S. Samarasekera, “Fuzzy connectedness and object
definition: Theory, algorithms, and applications in image segmenta-
tion,” Graph. Models Image Processing, vol. 58, pp. 246–261, 1996.
[9] D. L. Pham, “Spatial models for fuzzy clustering,” Comput. Vision
Image Understanding, vol. 84, pp. 285–297, 2001.
[10] A. Kaufmann, Introduction to the Theory of Fuzzy Subsets. New
York: Academic, 1975, vol. 1.
[11] T. Y. Kong and A. Rosenfeld, “Digital topology: Introduction and
survey,” Comput. Vision, Graph. Image Processing, vol. 48, pp.
357–393, 1989.
[12] S. Dellepiane and F. Fontana, “Extraction of intensity connectedness
for image processing,” Pattern Recogn. Lett., vol. 16, pp. 313–324,
1995.
[13] S. Dellepiane, F. Fontana, and G. L. Vernazza, “Nonlinear image
labeling for multivalued segmentation,” IEEE Trans. Image Pro-
cessing, vol. 5, pp. 429–446, 1996.
[14] J. K. Udupa and S. Samarasekera, “Fuzzy connectedness and object
definition,” in Proc. SPIE: Med. Imag., vol. 2431, 1995, pp. 2–11.
[15] P. K. Saha, J. K. Udupa, and D. Odhner, “Scale-based fuzzy con-
nected image segmentation: Theory, algorithms, and validation,”
Comput. Vision Image Understanding, vol. 77, pp. 145–174, 2000.
[16] P. K. Saha and J. K. Udupa, “Fuzzy connected object delineation:
Axiomatic path strength definition and the case of multiple seeds,”
Comput. Vision Image Understanding, vol. 83, pp. 275–295, 2001.
[17] J. K. Udupa, P. K. Saha, and R. A. Lotufo, “Relative fuzzy connect-
edness and object definition: Theory, algorithms, and applications
in image segmentation,” IEEE Trans. Pattern Anal. Machine Intell.,
vol. 24, pp. 1485–1500, 2002.
[18] P. K. Saha and J. K. Udupa, “Iterative relative fuzzy connectedness
and object definition: Theory, algorithms, and applications in image
segmentation,” Proc. IEEE Workshop on Mathematical Methods in
Biomedical Image Analysis, pp. 28–35, 2000.
1668
[19] J. K. Udupa, P. K. Saha, and R. A. Lotufo, “Fuzzy connected object
definition in images with respect to co-objects,” in Proc. SPIE: Med.
Imag., vol. 3661, 1999, pp. 236–245.
[20] G. T. Herman and B. M. De Carvalho, “Multiseeded segmentation
using fuzzy connectedness,” IEEE Trans. Pattern Anal. Machine In-
tell., vol. 23, pp. 460–474, 2001.
[21] B. M. Carvalho, C. J. Gau, G. T. Herman, and T. Y. Kong, “Algo-
rithms for fuzzy segmentation,” Patt. Anal. Appl., vol. 2, pp. 73–81,
1999.
[22] L. G. Nyul, A. X. Falcao, and J. K. Udupa, “Fuzzy-connected 3D
image segmentation at interactive speeds,” Graph. Models, vol. 64,
pp. 259–281, 2002, to be published.
[23] T. N. Jones and D. N. Metaxas, “Automated 3D segmentation using
deformable models and fuzzy affinity,” in Proc. Inform. Processing
in Med. Imag., 1997, pp. 113–126.
[24] T. N. Jones, “Image-Based Ventricular Blood Flow Analysis,” Ph.D.
dissertation, University of Pennsylvania, 1998.
[25] E. D. Angelini, C. Imielinska, Y. Jin, and A. Laine, “A improving sta-
tistics for hybrid segmentation of high-resolution multichannel im-
ages,” in Proc. SPIE: Med. Imag., vol. 4684, 2002, pp. 401–411.
[26] Y. Jin, A. Laine, and C. Imielinska, “An adaptive speed term based
on homogeneity for level-set segmentation,” in Proc. SPIE: Med.
Imag., vol. 4684, 2002, pp. 383–400.
[27] N. Bonnet and S. Maegherman, “Algorithme de Segmentation Semi-
Antomatique par la Ligne de Partage des Eaux,” (in French), LERI,
University of Reims, 1998.
[28] J. Cutrona and N. Bonnet, “Two methods for semi-automatic image
segmentation based on fuzzy connectedness and watershed,” in
France-Iberic Mecroscopy Congress, Barcelona, Spain, 2001, pp.
23–24.
[29] J. K. Udupa, L. Wei, S. Samarasekera, Y. Miki, M. A. van Buchem,
and R. I. Grossman, “Multiple sclerosis lesion quantification using
fuzzy connectedness principles,” IEEE Trans. Med. Imag., vol. 16,
pp. 598–609, 1997.
[30] S. Samarasekera, J. K. Udupa, Y. Miki, and R. I. Grossman, “A
new computer-assisted method for enhancing lesion quantification
in multiple sclerosis,” J. Comput. Assisted Tomogr., vol. 21, pp.
145–151, 1997.
[31] Y. Miki, R. I. Grossman, J. K. Udupa, M. A. van Buchem, L. Wei,
M. D. Philips, U. Patel, J. C. McGown, and D. L. Kolson, “Differ-
ences between relapsing remitting and chronic progressive multiple
sclerosis as determined with quantitative MR imaging,” Radiology,
vol. 210, pp. 769–774, 1999.
[32] Y. Miki, R. I. Grossman, J. K. Udupa, L. Wei, M. Polansky, L.
Mannon, and D. L. Kolson, “Relapsing-remitting multiple sclerosis:
Longitudinal analysis of MR images – Lack of correlation between
changes in T2 lesion volume and clinical findings,” Radiology, vol.
213, pp. 395–399, 1999.
[33] Y. Ge, R. Grossman, J. K. Udupa, L. Wei, L. Mannon, M. Polansky,
and D. L. Kolson, “Brain atropy in relapsing-remitting multiple
sclerosis and secondary progressive multiple sclerosis: Longitudinal
quantitative analysis,” Radiology, vol. 213, pp. 665–670, 2000.
[34] J. K. Udupa, L. G. Nyul, Y. Ge, and R. I. Grossman, “Multiprotocol
MR image segmentation in multiple sclerosis: Experience with over
1000 studies,” Acad. Radiol., vol. 8, pp. 1116–1126, 2001.
[35] T. Alderliesten, W. Niessen, K. Vincken, J. Maintz, F. Janson, O. van
Nieuwenhuizen, and M. Viewgever, “Objective and reproducible
segmentation and quantification of tuberous sclerosis lesions in
FLAIR brain MR images,” in Proc. SPIE: Med. Imag., vol. 4322,
2001, pp. 1509–1518.
[36] R. He and P. A. Narayana, “Automatic delineation of Gd enhance-
ments on magnetic resonance images in multiple sclerosis,” Med.
Phys., vol. 29, pp. 1536–1546, 2002.
[37] A. Kumar, W. Bilker, Z. Jin, J. K. Udupa, and G. Gottlieb, “Age of
onset of depression and quantitative neuroanatomic measures: Ab-
sence of specific correlations,” Psych. Res. Neuroimag., vol. 91, pp.
101–110, 1999.
[38] A. Kumar, W. Bilker, Z. Jin, and J. K. Udupa, “Atropy and high
intensity lesions: Complementary neurobiological mechanisms in
late-life major depression,” Neuropsychopharmacology, vol. 22, pp.
264–274, 2000.
[39] B. L. Rice Jr. and J. K. Udupa, “Clutter-free volume rendering for
magnetic resonance angiography using fuzzy connectedness,” Int. J.
Imag. Syst. Technol., vol. 11, pp. 62–70, 2000.
[40] T. Lei, J. K. Udupa, P. K. Saha, and D. Odhner, “Artery-vein separa-
tion via MRA – An image processing approach,” IEEE Trans. Med.
Imag., vol. 20, pp. 689–703, 2001.
PROCEEDINGS OF THE IEEE, VOL. 91, NO. 10, OCTOBER 2003
 [41] P. K. Saha, J. K. Udupa, and J. M. Abrahams, “Automatic bone-free
rendering of cerebral aneurysms via 3D-CTA,” in Proc. SPIE: Med.
Imag., vol. 4322, San Diego, CA, 2001, pp. 1264–1272.
[42] J. M. Abrahams, P. K. Saha, R. W. Hurst, P. D. LeRoux, and J. K.
Udupa, “Three-dimensional bone-free rendering of the cerebral cir-
culation using computed tomographic angiography and fuzzy con-
nectedness,” Neurosurgery, vol. 51, pp. 264–269, 2002.
[43] J. Liu, J. K. Udupa, D. C. Hackney, and G. Moonis, “Estimation
of tumor volume with fuzzy connectedness segmentation of MRI,”
in Proc. SPIE: Med. Imag., vol. 4322, San Diego, CA, 2001, pp.
1455–1465.
[44] G. Moonis, J. Liu, J. K. Udupa, and D. Hackney, “Estimation of
tumor volume using fuzzy connectedness segmentation of MR im-
ages,” Amer. J. Neuroradiol., vol. 23, pp. 356–363, 2002.
[45] P. K. Saha, J. K. Udupa, E. F. Conant, D. P. Chakraborty, and D.
Sullivan, “Breast tissue density quantification via digitized mammo-
grams,” IEEE Trans. Med. Imag., vol. 20, pp. 792–803, 2001.
[46] J. K. Udupa, J. Tian, D. Hemmy, and P. Tessier, “A pentium
PC-based craniofacial 3D imaging and analysis system,” J. Cranio-
facial Surg., vol. 8, pp. 333–339, 1997.
[47] J. K. Udupa, D. Odhner, and H. C. Eisenberg, “New automatic mode
of visualizing the colon via CT,” in Proc. SPIE: Medical Imaging,
vol. 4319, San Diego, CA, 2001, pp. 237–243.
[48] N. R. Pal and S. K. Pal, “A review of image segmentation tech-
niques,” Pattern Recogn., vol. 26, pp. 1277–1294, 1993.
[49] P. K. Saha and J. K. Udupa, “Relative fuzzy connectedness among
multiple objects: Theory, algorithms, and applications in image
segmentation,” Comput. Vision Image Understanding, vol. 82, pp.
42–56, 2001.
[50] T. Lindeberg, Scale-Space Theory in Computer Vision. Boston,
MA: Kluwer, 1994.
[51] S. M. Pizer, D. Eberly, D. S. Fritsch, and B. S. Morse, “Zoom-in-
variant vision of figural shape: The mathematics of cores,” Comput.
Vision Image Understanding, vol. 69, no. 1, pp. 55–71, 1998.
[52] M. Tabb and N. Ahuja, “Multiscale image segmentation by inte-
grated edge and region detection,” IEEE Trans. Image Processing,
vol. 6, no. 5, pp. 642–655, 1997.
[53] T. Cormen, C. Leiserson, and R. Rivest, Introduction to Algo-
rithms. New York: McGraw-Hill, 1991.
[54] J. K. Udupa and D. Odhner, “Shell rendering,” IEEE Comput.
Graph. Appl., vol. 13, no. 6, pp. 58–67, 1993.
[55] E. Dijkstra, “A note on two problems in connection with graphs,”
Num. Math. , vol. 1, pp. 269–271, 1959.
[56] J. K. Udupa, V. R. LaBlanc, H. Schmidt, C. Imielinska, P. K. Saha, Y.
Zhuge, P. Molholt, and Y. Jin, “Methodology for evaluating image
segmentation algorithms,” in Proce. SPIE: Medical Imaging, vol.
4684, San Diego, CA, 2002, pp. 266–277.
[57] J. M. Charcot, Lectures on the Diseases of the Nervous
System. London: The New Sydenham Society, 1877.
[58] J. F. Kurtzke, “Geographic distribution of multiple sclerosis: An up-
date with special reference to Europe and the mediterranean region,”
Acta Neurologica Scandinavica, vol. 62, pp. 65–80, 1980.
[59] , “Rating neurologic imapairment in multiple sclerosis: An
expanded disability status scale (EDSS),” Neurology, vol. 33, pp.
1444–1452, 1983.
[60] L. G. Nyul and J. K. Udupa, “Protocol-independent brain MRI seg-
mentation method,” in Proc. SPIE: Medical Imaging, vol. 4684, San
Diego, CA, 2002, pp. 1588–1599.
[61] R. B. Lauffer, D. J. Parmelee, S. U. Dunham, H. S. Ouellet, R. P.
Dolan, S. Witte, T. J. McMurry, and R. C. Walovitch, “MS-325: Al-
bumin-targeted contrast agent for MR angiography,” Radiology, vol.
207, pp. 529–538, 1998.
[62] E. Warner et al., “The risk of breast-cancer associated with mam-
mographic parenchymal patterns: A meta-analysis of the published
literature to examine the effect of method of classification,” Cancer
Detection and Prevention, vol. 16, pp. 67–72, 1992.
[63] J. N. Wolfe, “Breast pattern as an index of risk for developing breast
cancer,” AJR, vol. 126, pp. 1130–1139, 1976.
[64] , “Breast parenchymal patterns and their changes with age,” Ra-
diol., vol. 121, pp. 545–552, 1976.
[65] R. J. Schwab, K. B. Gupta, W. B. Gefter, E. A. Hoffman, and
A. I. Pack, “Upper airway soft tissue anatomy in normals and
patients with sleep disordered breathing: Significance of the lateral
pharyngeal walls,” Amer. J. Respiratory Crit. Care Med., vol. 152,
pp. 1673–1689, 1995.
UDUPA AND SAHA: FUZZY CONNECTEDNESS AND IMAGE SEGMENTATION
[66] R. Arens, J. M. McDonough, A. M. Corbin, E. M. Hernandez, G.
Maislin, J. R. Schwab, and A. I. Pack, “Linear dimensions of the
upper airway structure during development: Assessment by mag-
netic resonance imaging,” Amer. J. Respiratory Crit. Care Med., vol.
165, pp. 117–122, 2002.
[67] J. G. Liu, J. K. Udupa, D. Dewey, J. M. McDonough, and R. Arens,
“A system for upper airway segmentation and measurement via
MR imaging and fuzzy connectedness,” Acad. Radiol., vol. 10, pp.
13–24, 2003.
[68] D. J. Vining, “Virtual colonoscopy,” Gastrointest. Endosc. Clin.
North America, vol. 7, pp. 285–291, 1997.
[69] J. C. Bezdek, “Fuzzy Mathematics in Pattern Classification,” Ph.D.
dissertation, Cornell University, Ithaca, NY, 1973.
[70] E. Backer and A. C. Jain, “A clustering performance measure based
on fuzzy set decomposition,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 3, pp. 66–74, 1981.
[71] T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham, “Active
shape models-their training and application,” Comput. Vision Image
Understanding, vol. 61, pp. 38–59, 1995.
[72] T. F. Cootes, G. J. Edwards, and C. J. Taylor, “Active appearance
models,” IEEE Trans. Pattern Anal. Machine Intell., vol. 23, pp.
681–685, 2001.
Jayaram K. Udupa (Senior Member, IEEE)
received the Ph.D. degree in computer science
from the Indian Institute of Science, Bangalore,
in 1976 with a medal for best research.
During the past 25 years, he worked in the
areas of biomedical image and signal processing,
pattern recognition, biomedical computer
graphics, 3-D imaging, visualization, and their
biomedical applications. He has published over
120 journal papers in these areas, over 130 full
conference papers, edited two books on medical
3-D imaging, written over 20 book chapters and editorials, given over
120 invited talks (since 1986), offered consultancy to several industries,
organized conferences, seminars, and workshops, and codeveloped the
first-ever software package for medical 3-D imaging and developed and
widely distributed large software systems for 3-D imaging. Presently, he is
the Chief of the Medical Imaging Section and Professor of Radiological
Sciences in the Department of Radiology at the University of Pennsylvania,
Philadelphia.
Dr. Udupa is a Fellow of the American Institute of Medical and Biological
Engineering.
Punam K. Saha (Member, IEEE) received the
Bachelor’s and Master’s degrees in computer
science and engineering from Jadavpur Uni-
versity, India, in 1987 and 1989, respectively.
In 1997, he received the Ph.D. degree from the
Indian Statistical Institute, which he joined as a
faculty member in 1993.
In 1997, he joined the University of Penn-
sylvania, Department of Radiology, Medical
Imaging Section, as a Postdoctoral Fellow,
where he is currently an Assistant Professor. His
present research interests include biomedical imaging problems and the
application of their solutions, scale, segmentation, filtering, image registra-
tion, inhomogeneity correction, digital topology and their applications, and
estimation of trabecular bone strength from MR images. He has published
over 40 papers in international journals.
Dr. Saha received a Young Scientist award from the Indian Science Con-
gress Association in 1996. He is a member of the International Association
for Pattern Recognition and a member of the Governing body of the Indian
Unit for Pattern Recognition and Artificial Intelligence.
1669