2 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO.

1, JANUARY 2001

Three-Dimensional Modeling for Functional Analysis

of Cardiac Images: A Review
Alejandro F. Frangi*, Student Member, IEEE, Wiro J. Niessen, Associate Member, IEEE, and Max A. Viergever

Invited Paper

Abstract—Three-dimensional (3-D) imaging of the heart is a CT Computed tomography.

rapidly developing area of research in medical imaging. Advances CVD Cardiovascular disease.
in hardware and methods for fast spatio-temporal cardiac imaging DOFs Degrees of freedom.
are extending the frontiers of clinical diagnosis and research on
cardiovascular diseases. DSR Dynamic spatial reconstructor.
In the last few years, many approaches have been proposed to Green’s strain tensor.
analyze images and extract parameters of cardiac shape and func- EBCT Electron beam computed tomography.
tion from a variety of cardiac imaging modalities. In particular, EDV End diastolic volume.
techniques based on spatio-temporal geometric models have re- EF Ejection fraction.
ceived considerable attention. This paper surveys the literature of
two decades of research on cardiac modeling. The contribution of ESV End systolic volume.
the paper is three-fold: 1) to serve as a tutorial of the field for both Deformation gradient tensor.
clinicians and technologists, 2) to provide an extensive account of FE Finite element.
modeling techniques in a comprehensive and systematic manner, FFD Free-form deformation.
and 3) to critically review these approaches in terms of their per- Shape spectrum.
formance and degree of clinical evaluation with respect to the final
goal of cardiac functional analysis. From this review it is concluded GCG Geometric cardiogram.
that whereas 3-D model-based approaches have the capability to GDT Geometrically deformable template.
improve the diagnostic value of cardiac images, issues as robust- Mean curvature.
ness, 3-D interaction, computational complexity and clinical vali- HARP Harmonic phase.
dation still require significant attention. HR Heart rate.
Index Terms—Cardiac imaging, functional analysis, ICP Iterative closest point (algorithm).
model-based image analysis. Gaussian curvature.
Principal curvatures.
NOMENCLATURE KLT Karhunen-Loeve transform.
LV Left ventricle.
-D -dimensional, . LVM Left ventricular mass.
DE -dimensional echocardiography. LVV Left ventricular volume.
BA Biplane angiography. MF Wall/tissue motion field.
Curvedness. MRI Magnetic resonance imaging.
CDT Continuous distance transform. MTI Model tag intersections.
CFM Color flow (Doppler) mapping. NN Neural network.
CI Cardiac index. NURBS Nonuniform rational B-spline.
CO Cardiac output. RV Right ventricle.
CSG Constructive solid geometry. RVV Right ventricular volume.
Manuscript received April 25, 2000; revised October 5, 2000. This work Shape index.
was supported by the Netherlands Ministry of Economic Affairs under Project SA Strain analysis.
IOP Beeldverwerking IBV97009 and by EasyVision Advanced Development,
Philips Medical Systems BV, Best, The Netherlands. The Associate Editor re- SPAMM Spatial modulation of magnetization.
sponsible for coordinating the review of this paper and recommending its pub- SPECT Single photon emission computed tomography.
lication was A. Amini. Asterisk indicates corresponding author. SSP Similar shape patches.
*A. F. Frangi is with the Image Sciences Institute (ISI), University Medical
Center, Rm E.01.334, Heidelberglaan 100, 3584 CX, Utrecht, The Netherlands SV Stroke volume.
(e-mail: [email protected]). SVI Stroke volume index.
W. J. Niessen and M. A. Viergever are with the Image Sciences Institute (ISI), Local stretching factor.
University Medical Center, Heidelberglaan 100, 3584 CX, Utrecht, The Nether-
lands. US Ultrasound (imaging).
Publisher Item Identifier S 0278-0062(01)00799-6. WT Wall thickening.

0278–0062/01$10.00 © 2001 IEEE

FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 3

I. INTRODUCTION

T he number one cause of death in the United States since

1900 in every year but one (1918) has been CVD. More
than 2600 Americans die each day of CVD; an average of one
death every 33 s [1]. CVD claims more lives each year than
the next seven leading causes of death combined. According
to the most recent computations of the Centers for Disease
Control and Prevention of the National Center for Health
Statistics (CDC/NCHS), if all forms of major CVD were
eliminated, life expectancy would rise by almost ten years
while with elimination of all forms of cancer the gain would
be three years [1].1
Nowadays, there is a multitude of techniques available for
cardiac imaging which provide qualitative and quantitative in-
formation about morphology and function of the heart and great
vessels (Fig. 1). Use of these technologies can help in guiding
clinical diagnosis, treatment, and follow-up of cardiac diseases.
Spatio-temporal imaging is a valuable research tool to under-
stand cardiac motion and perfusion, and their relationship with
stages of disease.
Technological advances in cardiac imaging techniques pro-
vide 3-D information with continuously increasing spatial and
temporal resolution. Therefore, a single cardiac examination
can result in a large amount of data (particularly in multiphase
3-D studies). These advances have led to an increasing need for
efficient algorithms to plan 3-D acquisitions, automate the ex- Fig. 1. Diagram of the heart.
traction of clinically relevant parameters, and provide tools for
their visualization.
sional (2-D) images and their accuracy was, therefore, limited
Segmentation of cardiac chambers is an invariable prerequi-
[14].
site for quantitative functional analysis. Although many clinical
The literature on model-driven segmentation of cardiac im-
studies still rely on manual delineation of chamber boundaries,
ages has grown rapidly in the last few years and this trend is
this procedure is time-consuming and prone to intraobserver
likely to continue. To the best of our knowledge no survey is
and interobserver variability. Therefore, many researchers have
available that reviews this work. This paper presents a compre-
addressed the problem of automatic LV and RV segmentation.
hensive and critical review of the state-of-the-art in geometric
Since the shape of the cardiac ventricles is approximately
modeling of the cardiac chambers, notably the LV, and their po-
known, it seems natural to incorporate prior shape knowledge
tential for functional analysis. In order to set reasonable bounds
into the segmentation process. Such model-driven techniques
to the extent of this survey, we have confined ourselves to peer-
have received ample attention in medical image analysis in
reviewed archival publications2 proposing methods for LV (RV)
the last decade [3], [4]. A few advantages over model-free
segmentation, shape representation, and functional and/or mo-
approaches are: 1) the model itself can constrain the segmenta-
tion analysis, that fulfill the following selection criteria:
tion process that is illposed in nature owing to noise and image
artifacts; 2) segmentation, image analysis and shape modeling • the technique is model-based;
are simultaneously addressed in a common framework; 3) • the reconstructed model is 3-D3 ;
models can be coarse or detailed depending on the desired • illustration on cardiac images is provided.
degree of abstraction; 4) in some approaches, most of the This review is organized as follows. Section II gives a brief
chamber’s shape can be explained with a few comprehensible overview of the different acquisition modalities that have been
parameters which can subsequently be used as cardiac indexes used in imaging the heart. Section III overviews and defines
(cf. [5]–[9] among others). the most relevant clinical parameters that provide information
Use of geometric models is not completely new to the anal- on cardiac function. Section IV presents a systematic classifica-
ysis of cardiac images. As a matter of fact, traditional methods tion of cardiac models by type of geometrical representation/pa-
of obtaining parameters such as LVV and mass from echocar- rameterization. Attention is also given to the different types of
diography and angiocardiography were based on (simple) geo- input data and features for model recovery. This section is sum-
metrical models [10]–[13]. However, their use was mainly moti- marized in Table I. Section V discusses cardiac modeling ap-
vated by the need of extracting 3-D parameters from two-dimen-
2A few exceptions were made when the approaches were considered relevant
and journal versions were not available.
1At the time of writing, the authors could not find similar statistics 3Even if the imaging technique is not 3-D like, for instance, in the recon-
for Europe. There is, however, an ongoing European survey on CVDs struction of 3-D models from multiple nonparallel slices or from multiple 2-D
whose results are expected to appear soon [2]. projections.
4 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

TABLE I
OVERVIEW OF CARDIAC MODELING METHODS. SEE KEYS IN FIG. 3

proaches with respect to the functional parameters they provide II. IMAGING TECHNIQUES FOR CARDIAC EXAMINATION
and the degree of evaluation of these methods. This section leads
to Table II that links the clinical target of obtaining functional The physical properties on the basis of which the imaging
information of the heart (Section III) to the various technical ap- device reconstructs an image (e.g., radioactive emission of an
proaches presented in Section IV. Finally, Section VI closes the isotope) are intimately related to some specific functional as-
survey with conclusions and suggestions for future research. pects of the heart (e.g., its perfusion properties). Each imaging
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 5

1) first-pass studies in which the injected bolus dose is moni-

tored during its first passage through the heart and great vessels
and 2) gated equilibrium studies, in which the tracer mixes with
the blood pool before data collection. First pass acquisitions are
typically 2-D, while gated equilibrium studies can be 2-D or 3-D
(SPECT). Isotope imaging can be used to assess parameters like
EF [30] and regional wall motion analysis [30]–[32]. It is also
used to study myocardial perfusion [33] in cases of ischemia or
myocardial infarction, and to assess myocardial viability. The
Fig. 3. Keys to Table I. overwhelming majority of radionuclide studies performed for
perfusion assessment are SPECT.
modality presents advantages and limitations that influence the
achievable modeling accuracy. This section briefly reviews the D. Cardiac CT
techniques most frequently used for 3-D clinical investigation Conventional CT [19] had virtually no place in cardiovas-
of the heart. More extensive reviews and complementary read- cular examinations. Nowadays spiral CT [34], [35] is becoming
ings can be found in [15]–[22]. increasingly popular for cardiac imaging, with image quality
rivaling that of magnetic resonance. DSR [18] uses multiple
A. Angiocardiography X-ray tubes and image intensifier chains to produce “real time”
Angiocardiography is the X-ray imaging of the heart fol- multiple cross sections with similar acquisition times to ultra-
lowing the injection of a radio-opaque contrast medium. Al- fast CT but is not commercially available [15]. EBCT [36] or
though 2-D in principle, this technique can provide projections Ultrafast CT is both relatively inexpensive to perform and ca-
from two angles using a biplane system. Selective enhancement pable of providing 3-D information on coronary calcium de-
of the lesion to be demonstrated can be accomplished by po- posits (plaque) and cardiac cavities’ anatomy and function. A
sitioning an intravascular catheter through which the contrast current limitation of this system (relative to DSR) is that the
medium is guided and injected. Angiocardiography is usually spatial resolution in the transaxial direction is much less than in
good at anatomic delineation of lesions but much less satisfac- the in-plane (often transverse) direction.
tory in determining their severity and the degree of hemody-
namic disturbance that they have produced. This technique has E. MRI
been used for a long time to assess EF and volumes [10] based Cardiac MRI [21] is now an established, although still rapidly
on simplified geometric models [11]–[13] of the LV, but most advancing, technique providing information on morphology
radiologists use visual assessment based on experience [15]. and function of the cardiovascular system [37]. Advantages
of cardiac MRI include a wide topographical field of view
B. Cardiac US
with visualization of the heart and its internal morphology and
Two-dimensional US of the heart or “echocardiography” surrounding mediastinal structures, the capability of multiple
[16], [22] allows the anatomy and movements of intracardiac imaging planes, and a high soft-tissue contrast discrimination
structures to be studied noninvasively. The application of pulsed between the flowing blood and myocardium without the
and continuous-wave Doppler principles to 2-D echocardiog- need for contrast medium or invasive techniques. Long- and
raphy (2DE) permits blood flow direction and magnitude to short-axis views of the heart, as used in echocardiography, can
be derived and mapped onto a small region-of-interest of the be obtained routinely since arbitrary imaging planes can be
2DE image. In CFM, the pulsed-wave signal with respect to selected.
blood velocity and direction of flow throughout the imaging Another advantage of MRI is that it can provide both
plane is color coded, and produces a color map over the 2DE anatomical and functional information about the heart. Several
image. One of the limiting factors of 2DE is the US window researchers have used MRI to assess global and regional,
(presence of attenuating tissues in the interface between the US RV and LV function as represented by SV, EF and LV mass
transducer and the organ of interest). To overcome this problem [38]–[43], wall-thickening [44], myocardial motion [45], and
transesophageal echocardiography can be used, which allows circumferential shortening of myocardial fibers [46]. Data from
for high-quality color flow images at the expense of being MRI is more accurate than that derived from LV angiocardio-
invasive. graphy, where the calculation is based on the assumption that
Three-dimensional echocardiography (3DE) [17] is a rela- the LV is ellipsoidal in shape. Volume measurements by MRI
tively new development in US that allows 3-D quantitation of are independent of cavity shape, with the area from contiguous
organ geometry since the complete organ structure can be im- slices integrated over the chamber of interest.
aged. This technique has been used to compute LV volume and In contrast with other techniques, including 2DE and angio-
mass [23]–[28] and to perform wall motion analysis [29]. cardiography, anatomic information is easily defined on MRI.
The advantages of MRI over 2DE are a wider topographical
C. Isotope Imaging window and a superior contrast resolution.
Isotopes have been used to study LV function and myocar- A decade ago, MR tagging was introduced independently by
dial perfusion. Radionuclide techniques for monitoring global Zerhouni [47] and Axel [48]. This technique is able to create and
and regional ventricular function fall into two major categories: track material points (points attached to a fixed location of the
6 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

TABLE II
OVERVIEW OF CARDIAC MODELING METHODS: REPORTED CLASSICAL FUNCTIONAL PARAMETERS AND THEIR VALIDATION. SEE KEYS AND NOTES IN FIG. 4

myocardium) over time. Myocardial deformations can, there- tion, usually orthogonal to the imaging plane. Tissue deforma-
fore, be studied in a noninvasive manner. SPAMM creates two tion will be indicated by the displacement of black (saturated)
orthogonal sets of parallel planes (sheets) of magnetic satura- bands in the image that correspond to the intersections of the
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 7

be exhaustive, Appendix A summarizes a number of nonclas-

sical shape and motion descriptors that demonstrate the extra
possibilities provided by some advanced methods.

A. Global Functional Analysis

Weber and Hawthorne [49] proposed a classification of car-
diac indexes according to their intrinsic dimensionality: linear,
surface and volumetric descriptors. Linear parameters have been
used intensively in the past since they can easily be derived
from 2-D imaging techniques like 2DE and X-ray angiocardio-
graphy.4 However, they assume an “idealized” geometry of the
LV and strongly depend on external or internal reference and co-
ordinate systems. Besides total ventricular wall area, other sur-
face indexes based on curvature and derived parameters have
been investigated from 2-D studies [50]–[53]. More recently,
many image processing approaches to LV modeling have sug-
gested true 3-D global and local shape indexes based on surface
properties.
In practice, assessment of cardiac function still relies on
simple global volumetric measures like LVV and mass, and
EF. These and other basic parameters will be presented in the
following paragraphs.
Left Ventricular Volume (LVV): is a basic parameter required
to derive other LV indexes like, e.g., EF. Angiocardiography
and echocardiography have been traditionally used to assess this
quantity. In the latter case, three approaches have been applied:
represent the LV volume 1) as the volume of a single shape (e.g.,
truncated ellipse); 2) as the sum of multiple smaller volumes of
similar configuration (e.g., Simpson’s method), and 3) as a com-
bination of different figures [14, p. 585]. The achieved accuracy
in the assessment of LVV with echocardiography varies largely
Fig. 4. Keys and notes to Table II. with the model used to represent the LV. The best results have
been obtained using Simpson’s rule where in vitro studies have
imaging plane (tag grid). This grid only provides the in-plane revealed a relative error ranging from 5.9% to 26.6% depending
motion component (2-D motion). To reconstruct the 3-D motion on the particular implementation and the number of short-axis
of the material points, a number of 2-D tagged image sections slices used in the computation [14, p. 588]. It has been shown
must be obtained in at least two orientations. Further postpro- that echocardiography consistently underestimates ventricular
cessing is then required to interpolate the displacement field and cavity, while angiocardiography consistently overestimates true
to eventually perform strain analysis. volumes [14]. In a recent study by Lorenz et al. [42] with a
canine model and autopsy validation, it has been shown that
III. CLASSICAL DESCRIPTORS OF CARDIAC FUNCTION cine MRI is a suitable and accurate method to estimate RVV
and LVV. In this study, MR-based and autopsy volumes agreed
Development of models of the cardiac chambers has within 6 ml, yielding no statistically significant differences.
emerged from different disciplines and with various goals. Left Ventricular Volume (LVM): LV hypertrophy, as defined
Cardiac models have been used for deriving functional infor- by echocardiography, is a predictor of cardiovascular risk and
mation, for visualization and animation, for simulation and higher mortality [14, p. 599 and references therein]. Anatomi-
planning of surgical interventions, and for mesh generation for cally, LV hypertrophy is characterized by an increase in muscle
FE analysis. mass or weight.
This survey will be confined to the application of modeling LVM is mainly determined by two factors: chamber volume,
techniques for obtaining classical functional analysis. Classical and wall thickness. There are two main assumptions in the com-
functional analysis can be divided into global functional putation of LVM: 1) the interventricular septum is assumed to
analysis (Section III-A), and motion/deformation analysis be part of the LV and 2) the volume, , of the myocardium is
(Section III-B), from which the most clinically relevant param- equal to the total volume contained within the epicardial borders
eters can be obtained. of the ventricle, , minus the chamber volume, ;
Model-based methods also allow one to derive new descrip-
4Such parameters are, for instance, left ventricular internal dimension
tors of cardiac shape and motion. Such advanced descriptors
(LVID), relative wall thickness (RWT), and estimates of fractional shortening
have been mainly presented in the technical literature and their ( 1D)
of the cardiac fibers % and their velocity (V ) . For a detailed analysis
clinical relevance has still to be assessed. Without pretending to of these parameters the reader is referred to Vuille and Weyman [14].
8 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

LVM is obtained by multiplying by the density of the muscle B. Motion and Deformation Analysis
tissue (1.05 g/cm ) Motion analysis5 : A number of techniques have been used
in order to describe and quantify the motion of the heart. They
(1) can be divided into three main categories [56]: i) detecting en-
(2) docardial motion by observing image intensity changes, ii) de-
termining the boundary wall of the ventricle, and subsequently
LVM is usually normalized to total body surface area or weight tracking it, and iii) attempting to track anatomical [57]–[60],
in order to facilitate interpatient comparisons. Normal values of implanted [61]–[67] or induced [47], [48], [68]–[74] myocar-
LVM normalized to body weight are g/kg [42]. dial landmarks. There are a few problems involved with each of
Stroke Volume (SV): is defined as the volume ejected be- these techniques. Assumptions must be made about the motion
tween the end of diastole and the end of systole. (motion model) in the first two groups in order to obtain a unique
point-wise correspondence between frames. To this end, optic
flow methods [75]–[80]6 and phase contrast MR [82]–[85] have
end-diastolic volume been applied for i), and curvature-based matching [86]–[90]
end-systolic volume (3) has been used to find point correspondences in ii). Landmark-
based methods [47], [48], [57]–[74] provide information on ma-
Alternatively, SV can be computed from velocity-encoded terial point correspondence. However, this information is mostly
MR images of the aortic arch by integrating the flow over a com- sparse and, again, assumptions on the type of motion have to be
plete cardiac cycle [54]. Similar to LVM and LVV, SV can be made in order to regularize the problem of finding a dense dis-
normalized to total body surface. This corrected SV is known as placement field. The use of implanted markers adds the extra
SVI. Healthy subjects have a normal SVI of ml/m [42]. complication of being invasive, which precludes routine use of
Ejection Fraction (EF): is a global index of LV fiber short- this technique in humans. Although implanted markers are usu-
ening and is generally considered as one of the most meaningful ally regarded as the gold standard, there are some concerns in
measures of the LV pump function. It is defined as the ratio of the literature about their influence on both image quality and
the SV to the EDV modification of the motion patterns.
Wall Thickening (WT): Azhari et al. [91] have compared WT
% % (4) and wall motion in the detection of dysfunctional myocardium.
From their study, it was concluded that WT is a more sensi-
tive indicator of dysfunctional contraction [91]. This finding has
Lorenz et al. measured normal values of EF with MR [42]. They
triggered several researchers to define methods to quantify wall
found values of % (57%–78%) for the LV, and %
thickness. Azhari et al. [91], and Taratorin and Sideman [92]
(47%–76%) for the RV. Similar values were obtained with ultra-
carried out a regional analysis of wall thickness by dividing the
fast CT, echocardiography, and X-ray angiocardiography [14],
myocardium into small cuboid elements. The local wall thick-
[42].
ness is then defined as the ratio between the volume of the par-
Cardiac Output (CO): The role of the heart is to deliver an
ticular element and the average area of its endocardial and epi-
adequate quantity of oxygenated blood to the body. This blood
cardial surfaces [44].
flow is known as the cardiac output and is expressed in liters
The most widely employed method for WT computation,
per minute. Since the magnitude of CO is proportional to body
however, is the centerline method [93] and several improve-
surface, one person may be compared to another by means of
ments thereof [41], [94]–[96]. Starting with the endocardial
the CI, that is, the CO adjusted for body surface area. Lorenz et
and epicardial contours at each slice, the centerline method,
al. [42] reported normal CI values of l/min/m and a
in its original formulation, measures WT in chords drawn
range of 1.74–4.03 l/min/m .
perpendicular to a line that is equidistant to both contours (the
CO was originally assessed using Fick’s method or the indi-
centerline). Although more accurate than methods relying on
cator dilution technique [55]. It is also possible to estimate this
a fixed coordinate system, this method still assumes that the
parameter as the product of the volume of blood ejected within
contours are perpendicular to the long axis of the LV. If this is
each heart beat (the SV) and the HR.
not the case, the myocardial wall thickness is overestimated
which invariably occurs, for instance, in slices that are close
(5) to the apex. Buller and co-workers [41], [94] introduced an
improvement on this method by estimating at each location
In patients with mitral or aortic regurgitation, a portion of the the angle between the wall and the imaging plane. Recently,
blood ejected from the LV regurgitates into the left atrium or
ventricle and does not enter the systemic circulation. In these
patients, the CO computed with angiocardiography exceeds the 5At this point it is worth mentioning an excellent on-line bibliographic data-
forward output. In patients with extensive wall motion abnor- base maintained by the Special Interest Group on Cardiac Motion Analysis
malities or misshapen ventricles, the determination of SV from (SigCMA) that can be accessed at http://www-creatis.insa-lyon.fr/sigcma. It
also provides general bibliographic information on model-based cardiac image
angiocardiographic views can be erroneous. Three-dimensional analysis.
imaging techniques provide a potential solution to this problem 6For a survey of optic flow methods in computer vision, see Beauchemin and
since they allow accurate estimation of the irregular LV shape. Barron [81].
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 9

Bolson and Sheehan [95], [96] have introduced the center- IV. OVERVIEW OF MODELING TECHNIQUES
surface method (true 3-D extension of the centerline method)
which makes use of a reference medial surface to compute the A large effort has been devoted to the analysis and segmen-
chords and subsequent wall thickness. tation of cardiac images by methods guided by prior geometric
Strain Analysis (SA): is a method to describe the internal de- knowledge. When focusing on the way models are geometri-
formation of a continuum body. It is an appealing tool to study cally represented, three main categories can be distinguished:
and quantify myocardial deformation. Here, we shall briefly in- 1) surface models, 2) volumetric models, and 3) deformation
troduce some of the concepts related to SA. A comprehensive models. In all cases both discrete and continuous models have
exposition of this theory can be found in Fung [97]. been proposed as well as implicitly defined surface models
To describe the deformation of a body the position of any (Fig. 2).
point in the body needs to be known with respect to an initial Alternatively, one may classify model-based approaches by
configuration; this is called the reference state. Moreover, to considering the information that is used as input for model re-
describe position a reference frame is needed. In the following a covery. This categorization is highly determined by the imaging
Cartesian reference frame will be assumed. It is also common to modality for which the method has been developed. There are a
use curvilinear coordinates for which some of the expressions variety of inputs for model recovery: 1) multiple 2-D projection
simplify. images, 2) multiple oriented 2-D slices, 3) fully 3-D grey-level
A myocardial point, , has coordinates and a neigh- images, 4) 3-D point sets, 5) phase-contrast velocity fields, and
boring point, , has coordinates . Let be moved 6) MR tagging information.
to the coordinates , and its neighbor to . The de- In this survey we will compare the different methods with
formation of the body is known completely if we know the re- respect to type of model representation, and types of input data
lationship and features that the model is recovered from. Table I, in which
the different approaches are grouped according to the type of
(6) model representation, summarizes this section.

or its inverse
A. Surface Models
(7)
Many approaches to cardiac modeling focus on the endocar-
For every point in the body we can write dial (and/or epicardial) wall. Three subcategories are proposed:
1) continuous models with either global, local, or hybrid pa-
(8) rameterizations, 2) discrete models, and 3) implicitly defined
deformable models.
where is called the displacement of the particle . In order 1) Continuous Models: In the early studies of cardiac im-
to characterize the deformation of a neighborhood, the first par- ages by 2DE and angiocardiography, cardiologists used simpli-
tial derivatives of (6)–(8) are computed. These derivatives can fied models of the LV in order to compute functional parame-
be arranged in matrix form to define the deformation gradient ters like ventricular volume and mass from 2-D images. Most
tensor: . The deformation gra- of the times, simple ellipsoidal models were considered. See,
dient tensor enables to estimate the change in length between e.g., Vuille and Weyman [14] and Dulce et al. [40] for a com-
the neighboring points and , when they are de- prehensive review of such models and a comparison of their
formed into and . Let and be these lengths accuracy. In the last decades, however, approaches have ap-
before and after deformation. Then peared that make use of 3-D acquisitions to reconstruct models
varying from global parameterizations of the LV surface [5],
[27], [103], [110]–[112], [114], [153] to hierarchically parame-
(9)
terized models [9], [105], [107], [123], [137].
Global approaches: In this category, we will discuss sur-
where is the Green strain tensor [97] face representations that are based on simple geometric models.
In general they can provide, with a limited number of global
parameters, a rough shape approximation. We also include in
this category surface representations obtained as series of basis
functions with global support.
Cauvin et al. [103] model the LV as a truncated bullet, a
(10)
combination of an ellipsoid and a cylinder, that is fitted to the
morphological skeleton of the LV. Metaxas and Terzopoulos
where is the Kronecker tensor. From the strain tensor it is [154] have proposed superquadrics [155] to model simple ob-
possible to decompose the strains into two groups: axial and jects with a small number of parameters. Since the introduction
shear strains. The former correspond to the diagonal elements of superquadrics, several extensions have appeared in the liter-
and represent changes in length aligned with the axes of the ref- ature. Chen et al. [109] apply superquadrics with tapering and
erence frame while the latter correspond to off-diagonal terms bending deformations to model the LV in an integrated approach
or deformations where two axes are coupled. for image segmentation and shape analysis. The method iterates
10 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

points are detected in all frames using a Canny-Deriche edge

detector [158] in spherical coordinates [159]. Selected points
in subsequent frames are matched to the current frame using
a modification of the ICP algorithm [159]–[161]. Based on
corresponding point pairs, the parameters of a planispheric
transformation are retrieved by least-squares approximation.
This transformation allows to describe motion with just a few
parameters and to relate them to a canonical decomposition
(radial motion, twisting motion around the apico-basal axis,
and long-axis shortening).
Hierarchical approaches: Some authors have addressed
the problem of building hierarchical representations where a
Fig. 2. Proposed classification of cardiac modeling approaches. model described with few parameters is complemented with
extra deformations that capture finer details. Gustavsson et al.
between a region-based clusterization step [156], using statis- [105], for instance, employ a truncated ellipsoid to obtain a
tics of image intensity and gradient, and a shape-based step that coarse positioning of the LV cavity from contours drawn in two
checks the consistency between the current segmentation and a short-axis and three apical echocardiographic views. Further
superquadric model. Park, Metaxas and Axel [6] have extended model refinement is achieved using cubic B-spline curves
the flexibility of superquadrics by introducing parameter func- approximating manually segmented contours in multiple views.
tions: radial and longitudinal contraction, twisting and long-axis Chen et al. [107] and Bardinet et al. [9] use superquadrics [155]
deformation. These allow for a more detailed representation to coarsely describe the LV. Their approaches fundamentally
of the LV while keeping the intrinsic geometrical meaning of differ in the representation of the additional deformation field.
the superquadric parameters. LV mid-wall motion is recovered Chen et al. use spherical harmonics in order to approximate
using preprocessed MR tagging data obtained by sampling the the residual error between the superquadric estimate of the
LV mid-wall surface from the 3-D FE model of Young and Axel endocardial LV wall and the true wall location. Spherical
[68]. harmonics have the advantage that fine-tuning can be improved
Staib and Duncan [112] use sinusoidal basis functions to ad infinitum with increasing number of harmonics. However,
decompose the endocardial surface of the LV. The overall adding a new coefficient influences the shape of the model ev-
smoothness of the surface is controlled by decreasing the number erywhere (nonlocal basis functions). Bardinet et al. [9] extend
of harmonics in the Fourier expansion. Model recovery is cast the basic superquadric deformations (tapering and bending)
into a Bayesian framework in which prior statistics of the through the use of FFDs, a technique introduced in computer
Fourier coefficients are used to further limit the flexibility of graphics by Sederberg and Parry [162]. The superquadric is
the model. Matheny and Goldgof [111] compare different 3-D attached to a flexible, box-like frame, inducing a nonrigid
and four-dimensional (4-D) surface harmonic descriptions for deformation on the superquadric. Bardinet et al.use trivariate
shape recovery. Time can be incorporated in two ways in B-splines to parameterize this deformation field. In a later
the model: 1) hyper spherical harmonics, where an event in work, Bardinet et al. [113] apply their method to estimate LV
space-time is converted from Cartesian coordinates to hyper wall motion. This is accomplished by deforming the full model
spherical coordinates and 2) “time-normal” coordinates which (superquadric FFD) in the first frame, and modifying only
are formed by including a temporal dependency to each spatial the FFD in the subsequent frames. By tracking points with the
coordinate. Experiments carried out with a 3-D CT data set of a same parametric coordinates along the cardiac cycle, a number
canine heart have indicated that hyper spherical harmonics can of dynamic parameters like WT and twisting motion are com-
represent the beating LV with higher accuracy than direct normal puted. Germano et al. [122], [123] have developed a system for
extensions of spherical, prolate spheroidal and oblate spheroidal automatic quantification of LV function from gated-perfusion
harmonics. Coppini et al. [27] reconstruct a 3-D model of the SPECT images. An iterative algorithm fits an ellipsoidal model
LV based on apical views in US images. LV boundaries are to a semiautomatically obtained segmentation. This iterative
obtained by grouping edges with a feed-forward NN integrating algorithm incrementally adapts the ellipsoid’s parameters and
information about several edge features (position, orientation, center of mass so that accurate registration of the model is
strength, length, and acquisition angle). This allows discarding obtained even in the presence of large perfusion defects. The
many edge points that are not plausible LV boundary points. The ellipsoid defines a coordinate system that is used to refine
3-D LV geometry is modeled as a spherical elastic surface under the model. A Gaussian model of the count profiles is used
the action of radial springs (attracting the model to the edge to compute radial offsets corresponding to the endocardial
points); a Hopfield [157] NN is used to solve the minimization and epicardial walls. Although simple in its formulation, this
problem involved in the reconstruction of this surface. Declerck method has proven very useful in determining most of the
et al.[114] have introduced a spatio-temporal model to segment classical cardiac functional parameters [33] from SPECT
the LV and to analyze motion from gated-SPECT sequences. images and has been extensively validated in humans [122],
The model relies on a planispheric transformation that maps [123], [163].
endocardial points in one time frame to the corresponding Local approaches: A number of methods have been
material points in any other frame. First, endocardial edge reported to provide surface reconstruction using piecewise
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 11

polynomial surfaces, e.g., B-splines or bicubic Hermite surface applications can be found in the book by Metaxas [165] and in
patches. These techniques have appeared mainly in the context the survey by McInerney and Terzopoulos [3].
of surface reconstruction from multiple cross sections [29], McInerney and Terzopoulos [124] have applied this theory
[143] or projections [98]–[101], [115]. Given the ill-posed na- to the segmentation and tracking of the LV in DSR image se-
ture of this problem, most of these techniques require extensive quences. A FE balloon [166] deformable model is discretized
user interaction. Usually, a set of landmarks or fiducial points using triangular elements, and deformed according to a first-
are determined from each cross section/projection and, using order approximation of the Lagrange equations of motion. User-
high-level knowledge about the viewpoint and the geometry of defined point constraints can be interactively inserted to guide
the LV, a local surface approximation using surface patches is the deformation of the model and to avoid local minima of the
performed. potential energy in which the model is embedded. In the La-
A rather different approach is the one by Pentland and grangian formulation, 3-D image sequences can easily be han-
Horowitz [102] who applied modal analysis and FE to recon- dled by making the potential energy a function of time. Mon-
struct a 3-D model of the LV from X-ray transmission data. tagnat, Delingette and Malandain [131] apply simplex meshes
Modal analysis offers a principled physically based strategy [167] to reconstruct the LV from multiple views of a rotating US
for reducing the number of DOFs of the model and to obtain probe. Images are acquired in cylindrical coordinates coaxial
an over-constrained problem for shape recovery. Optic flow is with the apico-basal axis. Accordingly, images are filtered in
used to derive the deformation of the 3-D model from the 2-D cylindrical coordinates. Boundary points are detected based on
views, and a Kalman filter for tracking the structures over time. a combination of image gradient and intensity profiles normal
Instead of working with multiple cross sections or projec- to the surface. Finally, detected edge points are cast into point
tion images, Goshtasby and Turner [110] segment LV and RV attraction-forces deforming the model according to Newton’s
endocardial surfaces from 3-D flow-enhanced MR images. In law of motion. Ranganath [125] reconstructs 3-D models of the
this case, the endocardial surface is modeled as a deformable LV from MRI images using multiple 2-D snakes [168] and de-
cylinder using rational Gaussian surfaces [164]. The model is vising efficient mechanisms for interslice and interframe con-
deformed to fit the zero-crossings of the image Laplacian. To tour propagation. Biedenstein et al. [132] have recently pub-
avoid attraction by spurious edges, prior to fitting, the feature lished an elastic surface model and applied it to SPECT studies.
map is masked by a rough LV region-of-interest obtained by in- The elastic surface is deformed according to a second-order par-
tensity thresholding. tial differential equation. The external (image) forces are de-
RV models: Some efforts have also been directed toward rived from the radioactive distribution function and push the
geometric modeling of the RV. This chamber has a more com- elastic surface toward the center surface of the LV wall. Wall
plex shape than the LV. Spinale et al. [101] fit semiellipses to thickness can be then computed as the distance between the
model the crescentic shape of the RV from biplane ventriculo- elastic surface and the mass points of the radioactivity distri-
grams. Czegledy and Katz [104] model the RV using a cres- bution gradient. Huang and Goldgof [120] have presented an
centic cross-sectional model composed of two intersecting cir- adaptive-size mesh model within a physics-based framework for
cles of different radii. This 3-D model is parameterized by only shape recovery and motion tracking. The optimum mesh size is
a few linear dimensions that can be measured directly from CT, inferred from image data, growing new nodes as the surface un-
MR, or US images. From these dimensions, the RVV is ap- dergoes stretching or bending, or destroying old nodes as the
proximated using analytical expressions. Denslow [108] model surface contracts or becomes less curved. The method is em-
the RV as the difference of two ellipsoids (an ellipsoidal shell ployed to analyze LV motion from a DSR dataset. To establish
model). The parameters from this shell are estimated from MR point correspondences, an adaptive-size mesh is generated for
images (a long-axis and a four-chamber view) and from those, the first frame to be analyzed; subsequent frames further deform
volume estimates can be derived. Sacks et al. [106] model the this mesh while keeping its configuration fixed.
endocardial and epicardial walls of the RV by biquadric surface Physics-based modeling frequently makes an assumption that
patches (contours were manually traced from MR images), and can be problematic: internal constraints are usually represented
have studied surface curvature and wall thickness changes along in the form of controlled-continuity stabilizers [169]. It is known
the cardiac cycle using this representation. that, in the absence of image forces, deformable models tend to
2) Discrete Models: An alternative to continuous surface shrink. To avoid this, Rueckert and Burger [128] simultaneously
representations is the use of discrete surface models. Several model the two cardiac chambers (RV and LV) using a GDT. The
methods have been reported in the literature and they can be standard stabilizers on the deformed model are replaced by a
grouped in the following way. stabilizer on the deformation field between a rest model and a
Physics-based models: Physics-based modeling has deformed model. A GDT consists of three parts: 1) a set of ver-
attracted the attention of many computer vision researchers. In tices that defines the rest state (the template), 2) a set of vertices
this framework, surface recovery is cast into the deformation of that defines a deformed state (an instance of the template), and
a virtual body (the geometric model plus its material properties) 3) a penalty function that measures the amount of deformation
under virtual external forces derived from image/point features, of the template with respect to its equilibrium shape (the stabi-
or user-defined constraints. In the final (deformed) state, this lizer). Another solution to the above mentioned problem, was
virtual body reaches an equilibrium between the external forces proposed by Nastar and Ayache [127] who model a surface as a
and internal (regularization) constraints. A good overview quadrilateral or triangular mesh of virtual masses. Each mass is
of the theory of physics-based deformable models and its attached to its neighbors by perfect identical springs with pre-
12 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

defined stiffness and natural length. The system deforms under static. Legget et al. [26], [174] use piecewise smooth subdivi-
the laws of dynamics. In addition to elastic and image forces, an sion surfaces [175] to reconstruct the LV geometry from man-
“equilibrium force” determines the configuration of the mesh in ually traced contours in 3-D US images. Some elements of the
the absence of external forces. mesh can be labeled so that they allow for sharp edges (e.g., at
Spatio-temporal models: Several researchers have devel- the mitral annulus and apex) and to define regional surface de-
oped models that explicitly incorporate spatial and temporal scriptors. Also, from 3-D US images, Gopal et al. [25] apply tri-
variations of LV shape. Faber et al. [118] use a discrete 4-D angulated surfaces to reconstruct the geometry of latex balloons
model to segment the LV fom SPECT and MR images through phantoms mimicking the LV. Three-dimensional reconstruction
a relaxation labeling scheme [170]. Endocardial and epicardial is directly obtained by triangulating the points of manually de-
surfaces are modeled as a discrete template defined in a mixed lineated contours from a stack of quasiparallel slices.
spherical/cylindrical coordinate system co-axial with the LV 3) Implicitly Defined Deformable Models: Either in contin-
long-axis. Each point in the template represents a radius uous or discrete form, the models in the two previous para-
connected to this axis. The model is spatio-temporal since the graphs were characterized by having an explicit surface param-
compatibility functions computed in the relaxation labeling eterization. A surface model can also be defined by means of
scheme involve neighboring points both in space and time. an implicit function. For instance, in the level-set approach,
In this way, surface smoothness and temporal coherence of a model is obtained as the zero level set of a higher-dimen-
motion are taken into account. Tu et al. [126] have proposed sional embedding function. This technique, sometimes referred
a 4-D model-based LV boundary detector for 3-D CT cardiac as geodesic deformable models have been introduced indepen-
sequences. The method first applies a spatio-temporal gradient dently by Caselles et al. [176] and Malladi et al. [177] based
operator in spherical coordinates with a manually selected on the work by Osher and Sethian [178]. Geodesic deformable
origin close to the center of the LV. This operator is only models have been applied by Yezzi et al. [133], [134] to the seg-
sensitive to moving edges, and less sensitive to noise compared mentation of MR cardiac images. Recently, Niessen et al. [136]
to a static edge detector. An iterative model-based algorithm have extended the method to treat multiple-objects and have ap-
refines the boundaries by discarding edge points that are far plied it to the segmentation of 3-D cardiac CT and MR images.
away from the global model. The model is parameterized Although geodesic models have the ability to handle changes in
by spherical harmonics including higher order terms as the topology, unwanted and uncontrollable topological changes can
refinement proceeds. occur in images of low-contrast edges or with boundary gaps
Polyhedral models: LV polyhedral representations have since this is a purely data driven approach.
been applied by several authors [25], [26], [116], [119], [121], There are other types of implicit models not related to level-
[129]–[131] in the literature. The approaches differ either in sets. Tseng et al. [179], for instance, use a NN to define a CDT to
the type of polygonal primitive (e.g., triangular or quadrilateral the LV boundary. A feed-forward NN is trained to learn the dis-
meshes) or the details of the shape recovery algorithm (imaging tance function to the endocardial and epicardial contours using
modality, input data or recovery features). Shi et al. [129], [130] a few hand-segmented image slices. The surface of the LV is
use a Delaunay triangulation [171] to build a surface descrip- then represented as the zeroes of the distance function. The NN
tion from a stack of 2-D contours obtained with a combined can generalize the boundaries of the LV in the slices not in-
gradient- and region-based algorithm [172]. This representation cluded in the training set, thus serving as an aid to segment a
is subsequently used for motion analysis based on point cor- 3-D image for which the user has to provide the segmentation
respondences. Bending energy under a local thin-plate model of a few slices only. Under an affine deformation model, the dis-
is used as a measure of match between models of consecutive tance transform is used to match different temporal frames and
frames. Friboulet, Magnin and Revel [119] have developed a to derive motion parameters. Wall thickness is computed by the
polyhedral model to analyze the motion of the LV from 3-D centerline method [93] using two CDT NNs for describing the
MR image sequences. LV contours are manually outlined using endocardial and epicardial surfaces.
a track-ball. After applying morphological and linear filtering to A third approach to implicit modeling is the use of surface
diminish quantization noise, the contours are radially resampled primitives which are defined in implicit form. Lelieveldt et
with constant angular step. Finally, the stack of resampled con- al. [137] segment thoracic 3-D MR images using hierarchical
tours is fed into a triangulation procedure [173] which generates blending of hyperquadrics [180] and concepts of CSG [181].
a polyhedral surface with approximately equal-sized triangles. The method provides an automatic, coarse segmentation of
Faber et al. [121] use a combination of cylindrical and spher- a multiple-object scene with little sensitivity to its initial
ical coordinate systems to build a discrete model of the LV in placement. The most representative organs in the thorax (lungs,
SPECT perfusion images. A radius function defined in a discrete heart, liver, spleen, and cardiac ventricles) are incorporated in
(orientation) space of longitudinal and circumferential coordi- the model which can be hierarchically registered to the scanner
nates describes the LV. For each orientation, the radius is deter- coordinate system using only a few coronal, sagittal, and
mined by finding the position of maximal perfusion (which is transversal survey slices. Owing to the contextual information
argued to occur in the center of the myocardium). After low-pass present in the model, this sparse information has successfully
filtering to remove outlier radii, the radius function is mapped been used to estimate the orientation of the long-axis of the
back to Cartesian space where the surface is represented using LV. This allows an observer-independent planning of 3-D
triangular or quadrilateral meshes. This approach shares some long-axis acquisitions in patients [182]. This technique was not
features of the work described in Faber et al. [118] but is purely designed to estimate accurate cardiac functional parameters but
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 13

can be used to generate a first initialization for more accurate is that material properties can be set based on experimental
algorithms. knowledge about myocardial mechanical properties, and not
on a virtual mechanical analog which usually leads to ad hoc
B. Volumetric Models parameter settings.
As opposed to the plethora of surface representations, the use
of volumetric models in the analysis and segmentation of car- C. Deformation Models
diac images has received little attention. Hitherto, we have focussed on representing either the endo-
O’Donnell et al. [7], [8] were the first to suggest a volu- cardial (or epicardial) surface, or the volume comprised within
metric model to recover myocardial motion from MR tagging. the myocardial muscle. Tissue deformation, however, can be
The model, coined hybrid volumetric ventriculoid, can be de- modeled without necessarily modeling the ventricular bound-
composed into three parts: 1) a thick-walled superquadric, 2) a aries. To this end, material point correspondences in different
local offset either in nonparametric [7] or parametric [8] form, temporal frames are required. These correspondences can be ob-
and 3) a local deformation in the form of a polyhedrization. tained by matching certain geometric properties over time (gen-
The thick-walled superquadric represents a high-level abstrac- eral techniques). If images are acquired using MR tagging tech-
tion model of the myocardium that is further refined by the local nology, several other approaches can be applied that exploit the
offsets. Altogether, these two parts constitute the rest model of explicit correspondences inferible from tag displacements (MR
the myocardium that is rigidly scaled to the dimensions of a new tagged-based techniques).
dataset. The local deformation field is responsible of capturing 1) General Techniques: Several techniques have been pro-
the detailed shape variability of different datasets. Park et al. [5] posed in the literature for deformation recovery based on shape
have extended their LV surface model [6] to a super-ellipsoid properties only. These methods are attractive because of their
model with parameter functions. The model is fitted to tagged generality. On the other hand, one must reckon with the va-
MR images providing a compact and comprehensive descrip- lidity of the underlying assumptions and/or motion models be-
tion of motion. Radial and longitudinal contraction, twisting, fore they are applied to analyze image sequences corresponding
long-axis deformation, and global translation and rotation are to normal and pathological myocardial motion patterns.
readily available from the parameter functions. Alternatively, a) Continuous models: Amini and Duncan [86] have de-
standard SA can be carried out. It is also possible to estimate veloped a surface model based on the assumption of conformal
other volumetric parameters like SV, CO, LVV, and LVM. In motion, where angles between curves are preserved but not
order to fit the model, a set of boundary points is manually delin- distances between points. The LV surface is divided into locally
eated and a set of tags are semiautomatically tracked along the quadric patches from which differential properties can be com-
cardiac cycle using the algorithm of Young et al. [72]. There- puted. Interframe patch correspondences are obtained using a
fore, the accuracy of all volumetric measurements depends on metric that is minimal for conformal motion. An assumption
the manual outlining. of this model is that the subdivision into surface patches and
Haber, Metaxas and Axel [140] have developed a model of the number of neighboring patches visited during the matching
biventricular geometry using FEs in a physics-based modeling process are sufficient to accommodate for the largest stretching
context. The 3-D motion of the RV is analyzed by defining that can occur between frames. Bartels et al. [144], [184] model
external forces derived from SPAMM MR tagging data [141]. material deformations with multidimensional splines. The
Creswell et al. [138] and Pirolo et al. [139] describe a mathemat- method shares properties of optical flow techniques to estimate
ical (biventricular) model of the heart built from 3-D MR scans motion fields. However, those approaches do not return an
of a canine specimen. Manual contour delineation of the epi- explicit model of the deformations (only displacements at
cardial, and LV and RV endocardial boundaries provides a set discrete positions are provided). The main assumption of this
of points that is approximated with cubic nonuniform rational technique is that, for a given material point, luminance is a
B-splines (NURBS [183]). From this representation, a hexahe- conserved quantity. As in optic flow techniques, with only
dral FE model is built in order to generate a realistic geometric this assumption the solution remains under-constrained and,
model for biomechanical analysis. therefore, a regularization term must be added. Illustrations of
Recently, Shi et al. [142] have introduced an integrated the method on 2-D cardiac X-ray sequences are provided and
framework for volumetric motion analysis. This work extends the formulation readily extends to 3-D sequences. However, it
the surface model of Shi et al. [129] by combining surface is questionable whether luminance conservation can provide a
motion, extracted from MR magnitude images, and motion cues reliable cue for deformation recovery in regions with homo-
derived from MR phase contrast (velocity) images. The latter geneous intensity, or in the presence of imaging artifacts and
provide motion information inside the myocardial wall but noise. For MR tagging, in particular, the approach must be
are known to be less accurate at the boundaries [85]. The two adapted since luminance is not conserved due to the physics of
sources of motion evidence (boundary and mid-wall motion) the imaging process [75].
are fused by solving the discretized material constitutive law of b) Discrete models: Benayoun and Ayache [90] propose
the myocardium assuming a linear isotropic elastic material. In an adaptive mesh model to estimate nonrigid motion in 3-D
this framework, the measured boundary and mid-wall motion image sequences. The size of the mesh is locally adapted to
estimates at two consecutive frames are used as boundary and the magnitude of the gradient, where the most relevant infor-
initial conditions of a FE element formulation. An advantage mation is supposed to appear (e.g., cardiac walls). Mesh adap-
of this method with respect to physically-based techniques tation is carried out at the first frame only; subsequent frames
14 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

only deform the mesh to recover motion. The underlying hy- A similar approach is followed by O’Dell et al. [145]. One-di-
pothesis is that the deformation is small. Meshes at two time mensional (1-D) displacements are obtained by three indepen-
instants are registered through an energy-minimizing approach dent sets of tag lines: one in the cardiac long-axis, and two or-
matching differential image properties (curvature and gradient). thogonal sets in the short-axis view. Reconstruction of the de-
Recently, Papademetris et al. [150]–[152] have proposed a de- formation field is performed in two interpolation steps. The first
formation model inspired by continuum mechanics. The method step assumes a global affine transformation between two time
recovers a dense deformation field using point correspondences frames. This is done to eliminate global bulk motion, and linear
obtained with the point-tracking algorithm of Shi et al. [129]. stretches and shear. In a second step, the residual deformation
Regularization is accomplished by measuring the internal en- is interpolated using a prolate spheroidal decomposition to de-
ergy of the myocardial tissue assuming a linear elastic body scribe the curvilinear deformations expected in the heart.
model. This is equivalent to a regularization term on the strain Both Young et al. [68], [72] and O’Dell et al. [145] assume
tensor space and not on the displacement field.7 Anisotropy of that the reference frame, to which the SA is related, is the un-
the fibrous structure of the LV is accounted for in the internal deformed state. This is normally the first frame in the sequence
energy by making the model stiffer in the fiber direction [185]. (planar tag surfaces). Although this simplifies the problem by al-
2) MR Tagging-Based Techniques: The introduction of MR lowing to decouple the motion component normal to the tagging
tagging has stimulated researchers to develop models of car- plane, these methods cannot be used to compute strains between
diac tissue deformation. Compared to motion recovery based two arbitrary frames. The latter can be useful in order to retro-
on point correspondences or optic flow, MR tagging has the ad- spectively select the reference frame to coincide precisely with
vantage that, in principle, material point correspondences can the diastole or systole, or to compute strains over a subset of
be estimated from tag information. In this section, different ap- the cardiac cycle. To circumvent this limitation, Moulton et al.
proaches for modeling the deformation fields are reviewed. Ac- [146] have proposed a Lagrangian approach that explicitly com-
curate tag localization is a prerequisite for subsequent deforma- putes the intersection of the tag surfaces in two arbitrary frames.
tion recovery and, therefore, it is a closely related topic to de- Tag surfaces are obtained by interpolating the tag curves that are
formation models. A brief overview of tag tracking techniques stacked in different imaging planes. Surface intersections define
is given in Appendix B. a set of material lines for each time frame. These points were
a) Continuous models: Several approaches have been used to perform strain calculations employing a -version of FE
proposed in which the parameterization of the deformation basis functions.
field is a continuous function. The availability of continuous Radeva et al. [147] use two coupled volumetric models: a
deformation maps allows the computation of local strains. tissue deformation field and a model describing the LV geom-
Young et al., for instance, developed a model-based approach etry. The first model is represented by a cubic trivariate B-spline
for tracking tag intersections [68] and tag stripes [72]8 that (coined B-solid by the authors); the second model is represented
has been validated using silicone gel phantoms [188]. A by two coupled surfaces (endocardium and epicardium) fitted
deformation field that maps the first (undeformed) frame to a to boundary points. It is assumed that the boundaries are ei-
subsequent (deformed) frame is modeled through a piecewise ther manually delineated or (semi)automatically detected from
polynomial function. Two fitting steps are involved in this the tagged images. The B-solid is deformed under thin-plate in-
method. First, the material points (tag intersections or stripes) in ternal constraints, and under two external forces. The first corre-
each deformed frame, , are reconstructed in the coordinate sponds to tagging information: the iso-parametric curves of the
system of the undeformed state, (reconstruction fit). In model are deformed to align with the tag strips. Simultaneously,
the latter frame, tag surfaces are arranged in true planes since the B-solid is attracted toward the LV boundaries by integrating
no motion has occurred yet. In the second step, the material a distance function to edge points on the epicardial and endocar-
points for , expressed in the reference frame , dial surfaces. Therefore, in this method, boundary and tag infor-
are used to reconstruct a displacement field relative to mation are incorporated in a unified approach. Since this method
(deformation fit).9 has been applied in combination with short-axis tagged images
only, it yields in-plane 2-D displacements. In a recent paper,
7Related regularization schemes are the global and body smoothing terms
Huang et al. [149] have extended the method to analyze true
described in Young and Axel [68] which act on the deformation gradient tensor. 3-D deformations using a spatio-temporal model. The method
However, they are not directly interpretable as an internal deformation energy.
8Amini [186] have compared landmark-based (tag intersections) against differs from the one of Radeva et al.in that no boundary infor-
curved-based tag (stripes) tracking based on the simulator of Waks [187]. mation is now incorporated. On the other hand, a spatio-tem-
It was concluded that as the number of stripes/landmarks increases, the two poral B-solid is constructed through a 4-D tensor product spline
methods give similar performances. Under large deformations, the degradation
of the curve-based techniques is more graceful compared to landmark-based
(3-D ). The fitting process to SPAMM data is governed by
methods. a normal constraint which enforces the attraction produced by
9Both fitting steps handle sparse data and, therefore, regularization is needed. each tag plane to be in its normal direction. Since multiple, or-
Regularization, however, is known to introduce artifactual strains. The effect of thogonal tag planes are available, this allows a full 3-D recon-
three regularization terms has been studied in [68]: 1) a thin-plate spline sta-
bilizer, 2) a global smoothing regularizer minimizing the deformation gradient struction of the deformation field.
tensor, F , and 3) a local body regularizer minimizing the deformation gradient Kerwin and Prince [74] have developed an alternating pro-
tensor expressed in some natural local coordinate system (e.g., aligned in cir- jection technique to accurately estimate the 3-D location of the
cumferential, longitudinal and radial directions). Based on simulations of an
axis-symmetric deformation of a thick walled incompressible cylinder, it was intersection points of the tag grid. The deformation field be-
shown that all three constraints yield similar results in the SA. tween two frames is recovered using thin-plate-spline interpo-
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 15

lation. Myocardial points are distinguished from those in static A. Validation

tissues by checking whether they pass across the imaging plane Three main groups of papers can be distinguished: 1) with no
over time. In points that do not fulfill the previous criterion, a evaluation or only qualitative illustrations, 2) with quantitative
test is performed to check their inclusion within the outlined evaluation on nonhuman datasets, and 3) with quantitative eval-
myocardial borders prior to rejection from the analysis. Such a uation on human datasets. This classification has been used in
rejection scheme is important for proper visualization and anal- constructing Table II.
ysis of myocardial motion. Although there are always exceptions confirming the rule,
Recently, Young [148] has introduced the concept of model Table II indicates several trends. Most papers in the first cate-
tags that represent the material surfaces within the heart tissue gory correspond to articles presenting technical or methodolog-
which are tagged with magnetic saturation. Model tags are “at- ical aspects of advanced modeling techniques. The result sec-
tached” to the heart and deform with it. They are embedded tions in these papers are restricted to either technical aspects or
within a 3-D FE model describing the geometry of the LV; this proof-of-concept illustration on realistic images hypothesizing
model is linear in the transmural direction and employs bicubic the potential of the technique. Only a few of them have seen
Hermite interpolation in the circumferential and longitudinal di- follow-up articles confirming those hypotheses in large studies.
rections. Instead of finding the 3-D location of the tag plane in- Further evaluation of these techniques is required in order to de-
tersections, this approach finds the intersections of the model termine their usefulness in clinical tasks.
tags with the imaging planes (MTIs). The FE model is subse- Approaches in the second category are numerous. Method-
quently deformed so that the MTI match the tag stripes in each ologies in this category have been evaluated on simulated
image plane. Matching is carried out by a local search algorithm images or in phantom experiments. These have the advantage
guided by an orientation filter. Additionally mechanisms are in- of providing ground truth to assess the accuracy and repro-
corporated to allow efficient user interaction and to correct for ducibility of the techniques. Owing to the use of idealized
erroneous MTI matches. geometries and measurement conditions, extrapolation of the
b) Discrete models: Moore et al. [69] use MR tagging to results to in vivo human studies remains to be demonstrated.
reconstruct the location of material points through the cardiac Some papers in this second category have evaluated their tech-
cycle by interpolating the positions of the tags from short- niques on ex vivo or in vivo animal models. Several researchers
and long-axis image planes using an iterative point-tracking have reported experiments with dogs [90], [101], [106], [108],
algorithm. Discrete tag locations are arranged in cuboid volume [111]–[113], [124], [127], [129], [136], [139], swines [101],
elements which are identified in the deformed and reference [108], [112], [146] or calfs [104], [192], [193].10 Only a few
frames. For each element, a 3-D strain tensor is calculated studies have compared measurements, obtained from ex vivo
using the generalized inverse method [189]. Since the SA is [192], [193] or in vivo [101], [129], [150] animal studies,
performed on a coarse discrete grid, only average strains can be against other standard-of-reference techniques.
retrieved. The tag tracking procedure of this method compen- MR tagging techniques for reconstruction of myocardial mo-
sates for through plane motion. An important conclusion from tion or tissue deformation deserve separate attention. Most in
this work is that SA can be largely influenced by through plane vivo animal and human studies have reported on Monte Carlo
motion if this is not corrected for. analysis of sensitivity to errors in tag localization and tracking,
Denney and Prince [73] employ a multidimensional sto- and on the ability to recover the location of tags in different
chastic approach to obtain a dense discrete model of the frames [69], [74], [145], [146], [191].11 Several models have
displacement field from a sparse set of noisy measurements been used in the literature to benchmark the accuracy of motion
(tag displacements). The displacement field is constrained and deformation recovery. These evaluations were based, for in-
to be smooth and incompressible (isochoric deformation). stance, on spherical and cylindrical models of cardiac motion
This formulation leads to a partial stochastic model of the [60], [69], [73], [145], FE solutions with realistic geometries
deformation field that can be solved using Fisher’s estimation [146], artificially generated motion trajectories [113] or syn-
framework [190]. thetic images using the cardiac motion simulator [74], [149],
[197] developed by Waks et al. [187] that builds upon the kine-
V. DISCUSSION matic model of Arts et al. [198]. Recently, a study was car-
ried out by Declerck et al. [199] that thoroughly compared four
Comparison of the performance of different techniques is a techniques [73], [145], [200], [201] for motion tracking from
difficult task due to the diversity of approaches, the different or tagged MR. This paper provides results on normal and patho-
complementary information obtained from them, the different logical subjects. Although the general trends of motion were
imaging modalities and image acquisition protocols, and, last captured correctly by all methods, this study shows that there
but not least, the lack of a standard way of reporting on perfor-
mance. In order to draw some comparative conclusions, we have 10Remarkably, a large amount of evaluations involving canine models have
classified the existing methodologies according to the degree of been acquired with the DSR. However, the reduced clinical availability of this
their validation (Section V-A). At the same time, we introduce technique and its specific image properties makes it difficult to extrapolate the
a number of performance criteria (Section V-B). In this com- results of the evaluation to other clinical imaging techniques.
11Validation MR tagging itself for describing tissue deformation has been
parison we have focussed on techniques leading to traditional
addressed by Young [196] using a silicone gel phantom. Strains derived
cardiac indexes, viz global (Section III-A) and motion parame- from MR tagging were compared to the analytic equilibrium strains under a
ters (Section III-B). Table II summarizes this discussion. Mooney–Rivlin material law.
16 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

are notable differences in the displacement and strain computa- are highly versatile and can accommodate detailed shape
tions provided by each technique. variations. Most of the quantitative evaluation studies have
Finally, the third category includes studies that reported appli- been reported on local flexible models, most of which are able
cation on human volunteers and patients, including quantitative to cope even with complex topologies. On the other hand,
results in terms of cardiac functional parameters. The size of the restricting the space of possible shapes is usually difficult
populations in most of these studies was small. With only three or requires substantial manual intervention or guidance [98],
exceptions, all studies were conducted on less than a dozen of [99], [101], [143]. Hierarchical or top-down approaches aim
subjects. at a reduction in computational time and at improving robust-
ness by incrementally unconstraining the space of allowed
B. Performance Criteria shape variation [7]–[9], [33], [105], [107]. One weak point
In the following subsections we elaborate on the criteria that in hierarchical approaches is the need for ad hoc scheduling
we have used to compare the different methods. mechanisms to determine when one level in the representation
1) Model Complexity or Flexibility: The complexity or hierarchy should be fixed and a new level added, and up to
flexibility of a technique has been categorized in four groups which level the model should be refined. Furthermore, opti-
according to the number of DOFs12 or parameters involved. 1) mization procedures involved in the recovery of hierarchical
Compact models with only a few parameters (on the order of a models have to be designed with particular care. It is unclear
dozen). Prototypical examples are superquadrics. 2) Flexible how it can be ensured that a succession of optimizations at
models with large number of DOFs and parameterized with different modeling levels actually leads to the optimum global
global-support basis functions. Representative examples are deformation. Also, the question arises how to link different
harmonic parameterizations of several types. 3) Flexible models levels of model detail with the resolution of the underlying
with large number of DOFs and parameterized with local-sup- image data, and how to interact with the models if, after all,
port basis functions. Members of this family are B-spline and manual editing is required. Still, hierarchical model repre-
polyhedral models. 4) Flexible hierarchical models encom- sentations are an active and challenging field in 3-D medical
passing a reduced set of DOFs coarsely describing shape, plus image segmentation research where several investigators have
an extended set of DOFs giving extra flexibility to the model. presented encouraging results in cardiac [7]–[9], [107], [113],
Representative of this family are superquadrics with FFDs. [122], [123] and thorax modeling [137], [202].
Complexity is, to some extent, related to the computational 2) Robustness and Effective Automation: Processing prior
demand of an algorithm. Highly flexible algorithms are usually to model recovery, automation of the recovery algorithm itself,
related to higher computation time for deforming them to a and the presence of ad hoc parameters are factors that determine
given image dataset.13 On the other hand, it is also a measure the robustness of a technique and its effective automation. By
of the ability of a modeling technique to accommodate for fine effective automation we refer to the automation of the overall
shape details. approach, from raw images until the presentation of the func-
Although idealized models of ventricular geometry (mainly tional parameters.
ellipsoids or ellipsoidal shells) are appealing for their parsimony Before a given model can be fitted or deformed to a dataset,
and for historical reasons, Table II shows that no study has quan- almost every technique requires some type of preprocessing to
titatively demonstrated their accuracy in computing simple mea- convert the raw grey-level images into a representation suitable
surements as LVV and EF. Compact models have developed in for shape recovery. Section IV has suggested a classification
two different directions. On one hand, in particular for the RV, of types of input data. For the sake of simplicity, Table II
some researchers have evaluated combinations of simple models only indicates the degree of manual involvement to obtain the
that roughly derive RVV from a small number of linear measure- corresponding input data. Four categories were considered:
ments [104], [108]. The models, however, remain highly con- (N) no preprocessing required, (I) manual initialization of
strained and have been tested on ex vivo casts experiments only. landmarks/models, (A) (semi) automated initialization of
A second direction has been to trade off the compactness of the landmarks/models integrated into the technique, and (M) fully
superquadric models and their flexibility without the need of hi- manual segmentation of landmarks/contours. Although vari-
erarchical decompositions [5], [6]. In this manner, flexibility is ability inherent to the preprocessing can have a marked effect
added in an elegant way by which each parameter function has on the overall performance of a technique, this factor is usually
an interpretation in terms of local and global shape changes. disregarded in the evaluation of algorithms. A remarkable
Most approaches that reached the stage of quantitative exception is the evaluation of MR tag tracking algorithms
evaluation are based on flexible or hierarchical representations. using Monte Carlo analysis to assess the influence of erroneous
Both present challenges and advantages. Flexible representa- tag localization in the recovery of tissue deformation [69],
tions (e.g., polyhedral meshes or harmonic decompositions) [74], [145], [146], [191]. Model initialization is also related to
the issue of preprocessing. Although a few techniques make
12Here, we disregard the obvious rigid transformation parameters to instan- explicit mention of the procedure required to initialize the
tiate the model in world coordinates. model (cf., e.g., [5], [6], [72], [132], [136], and [137]), model
13Actually, it is the conjunction of model parameterization and the recovery initialization in a 3-D environment can be nontrivial or require
strategy which determines the computational load of an approach. It would have expert guidance.
been very interesting to report computation time with each technique. Unfortu-
nately, variability in hardware architecture over time and techniques renders any Another factor undermining robustness and reliability of a
quantitative comparison unrealistic. technique, is the presence of ad hoc parameters that have to
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 17

be set by the user. This can be particularly problematic when of further research. So far the main flow of efforts has been fo-
such parameters are highly dependent on a given dataset. This cussed on adopting generic geometrical representations to build
is a known problem, for instance, of many physics-based de- cardiac shape models (e.g., superquadrics, B-splines, polyhe-
formable models for which several weights must be tuned to dral meshes, Fourier descriptors, etc.). As a consequence, in
balance the smoothing constraints to the external energy terms. generating a realistic LV shape, the representations are either
However, in the literature, analysis of sensitivity of the result too restrictive or require a considerable amount of parameters.
to the weighting parameters is mostly missing. In Table II, we The question arises of how to infer a compact representation
have classified the different techniques into two categories ac- giving rise to realistic shapes, possibly learned from examples.14
cording to the presence of user-defined ad hoc parameters: Modeling approaches that go from shape examples to a specific
no parameters or parameters with corresponding analysis of sen- shape representation can reduce computational demands and
sitivity, and parameters for which no sensitivity analysis improve their robustness. A small number of efficiently selected
was performed. The fact that several methods do not present ad model parameters reduces the dimensionality of the model re-
hoc parameters does not have to be confounded with overall covery problem, and naturally constrains its results owing to
robustness. Even within the approaches with quantitative evalu- model specificity.
ation, many papers in the category either require substantial Further investigation of suitable image features will be
preprocessing [9], [25], [26], [68], [145], [146], [150], [151] or needed to improve shape recovery. In particular, incorporation
human guidance [98], [99], [104], [108], [116], [117], [179]. of domain knowledge about the type of image modality (and
Both factors influence the robustness and reproducibility of the acquisition protocols) can play an important role in increasing
derived functional information. the accuracy of shape recovery techniques.
Finally, Table II also indicates the degree of user guidance Most of the modeling techniques presented in this review
(automation) of the fitting procedures for given input data (pre- were either purely geometric or inspired in a virtual physical
processing) and set of ad hoc parameters. Three degrees of au- analog (physics-based approaches). Recently, a few papers
tomation were used to classify the approaches: relying on have introduced known biomechanical properties of the heart
substantial human guidance, manual interaction can be nec- in the formulation of models that analyze cardiac images [142],
essary for guiding/correcting the deformation, and fully au- [150]–[152]. Further development of such approaches, and
tomated. In general terms, the larger the need for human inter- their application to segmentation tasks, can be a natural way of
vention during the fitting procedure, the less robust a technique extending the ideas of physics-based methods and of relating
will become, and the more prone it will be to interobserver/in- some of the ad hoc parameters with experimental evidence
traobserver variability of the final results. provided by biomechanics.
2) Research on Interactive Model-Based Segmenta-
tion: Table II supports the idea that model-based cardiac
VI. CONCLUSIONS AND SUGGESTIONS FOR FUTURE RESEARCH segmentation has not reached the status of being effectively
In this paper, we have reviewed techniques for 3-D geometric automated since current techniques either require substantial
modeling and analysis of cardiac images. In particular, we have expert guidance, ad hoc parameter fine-tuning or nontrivial
focussed on those techniques leading to traditional indexes of preprocessing. Although full automation is a desirable end
cardiac function. We have proposed a systematic classification goal, its difficulty has been acknowledged many times in the
of the approaches based on the type of representation of the geo- literature. There is a growing consensus that user interaction is,
metric model, and the type of input data required for model re- to some extent, unavoidable, and that it has to be considered as
covery (Table I). Furthermore, we have given a critical assess- an integrated part of the segmentation procedure. Therefore,
ment of these approaches according to the type of functional development of efficient tools for 3-D interaction will play an
parameters that they provide, their degree of evaluation, and important role in the near future. “Efficient” entails that with
the performance achieved in terms of modeling flexibility, com- minimal and intuitive user interaction the operator keeps control
plexity, and effective automation (Table II). over the segmentation process to correct or overrule its results
From the surveyed literature, four main lines of future efforts where it has failed, and to guide the algorithm in abnormal
can be distinguished. situations (e.g., in front of a pathological case). Of course, the
1) Research on Modeling and Model deformation Tech- issue of reproducibility in case of human intervention needs
niques: The last two decades have witnessed an enormous attention. Where well-defined repetitive tasks are recognized,
amount of efforts in 3-D models of LV and RV. This holds true or where a local user interaction can be extrapolated to a
for all imaging modalities (cf. Table I). In spite of the large broader area, the process should be automated, thus improving
number of attempts, no approach has simultaneously achieved segmentation throughput and repeatability. How to devise such
robustness, automation, model flexibility and computational efficient and intuitive mechanisms for 3-D manipulation of
speed. Manual outlining and analysis of cardiac images is still models and volumetric data, and how to integrate them into the
the most popular technique in clinical environments. deformation of the models remain topics of future research.
Several issues will require more attention in order to inte-
grate the advances of modeling techniques into clinical prac-
14An interesting approach is to extract statistical models from sample shapes
tice. Accurate 3-D modeling techniques are, in general, com-
[203] and to capture the most representative DOFs via principal component
putationally intensive. Exploration of flexible modeling tech- analysis. Although interesting results have been obtained in 2-D applications,
niques that make an efficient use of their DOFs will be worthy more research is needed to solve practical problems in their 3-D extensions.
18 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

3) Research on Functional Cardiac Descriptors: There are APPENDIX A

many shape and motion parameters other than traditional in- NONTRADITIONAL SHAPE AND MOTION DESCRIPTORS
dexes (cf. Appendix A). Unfortunately, although these new in- Three-dimensional model-based analysis of LV shape and
dexes seem to provide richer information and/or a more detailed motion has the potential of providing rich morphological
analysis of cardiac function, their clinical evaluation has been and functional information. Current clinical assessment of
very limited. As a consequence, it is difficult to determine their cardiac function is based mainly on global parameters as LVV
clinical relevance and the extra information provided with re- and EF. However, several researchers have demonstrated in
spect to traditional indexes like LVV, EF, etc. The lack of clin- the past the importance of local functional indexes as WT
ical evaluations may be related to the fact that advanced 3-D and segmental motion analysis [93], [204]–[206], and local
modeling techniques, from which these parameters can be de- curvature and shape [50]–[53] as potential CIs. Unfortunately,
rived, are computationally expensive and require considerable most of these studies were based on 2-D imaging techniques.
user intervention. The need of considerable preprocessing and Although they can indicate major trends about cardiac shape,
postprocessing procedures, ad hoc parameter settings and tech- a 3-D analysis would be beneficial to better account for the
nical understanding of the modeling technique itself may ex- true cardiac geometry. In this section, we briefly summarize
plain why most of the described approaches are not available as several new indexes proposed in the literature to describe shape
stand-alone prototypes on which clinical studies can be carried and/or motion. Some of these indexes have been presented as
out routinely. a by-product of a specific modeling technique while others are
There is certainly place for development of novel shape and easily computable from any model representation. Therefore,
motion descriptors. However, there is even a larger need for this distinction seems a natural classification.
evaluation of already existing indexes on reference data sets
and/or large scale clinical studies. It is remarkable that this lack A. Generic Descriptors
of large evaluation studies is present even in many techniques
Mean and Gaussian Curvature: The principal curvatures
aiming at the extraction of traditional functional parameters
( and , respectively) measure the maximum and minimum
(Table II).
bending of a regular surface. Rather than using principal
It is unrealistic to expect that every new technique proposed in
curvatures, it is more common to use two derived quantities
the future will go through the process of a thorough clinical eval-
known as Gaussian and mean
uation study. Unfortunately, many research institutes working
curvatures. By analyzing the signs of the pair it is
on geometric modeling and shape analysis are not located in a possible to locally distinguish between eight surface types
clinical environment. Access to state-of-the-art image material [207].
and derived parameters for testing and benchmarking purposes Friboulet et al. [88] have studied the distribution of the
is, therefore, difficult. In this respect, a public, common data- Gaussian curvature in the LV at different phases of the cardiac
base of a representative set of images from different modalities cycle. From this study it was concluded that this distribution
would be highly beneficial. This database should establish a few remains structurally stable over time. Whereas the LV free wall
standard data sets (both synthetic and clinical study cases) with provides rich and dense curvature information, the curvature
as much independent measurements as possible of mass, SV, etc. at the septal wall is less suitable to establish point correspon-
With the current speed of development in the imaging modali- dences. Similar findings were made by Sacks et al. [106] with
ties, such a database should be updated regularly to be represen- respect to the RV free wall: the RV free wall has relatively
tative of the state-of-the-art imaging technology. uniform distribution of principal curvatures, and the surface
4) Multidisciplinary Approaches: When imaging and mod- geometry of the RV free wall does not change significantly
eling techniques get more complex, the interplay of clinicians, from end diastole to end systole.
medical physicists, and technologists in a common environment Shape Index and Shape Spectrum: Although mean and
becomes increasingly important. Several issues have to be ad- Gaussian curvatures are related to the concept of curvedness,
dressed in a cooperative fashion: the interrelationship between there still remains scale information in these shape descriptors.
image acquisition and cardiac modeling, the development of ef- To overcome this problem, Clarysse et al. [208] have used the
fective visualization techniques of 4-D datasets, realization of shape index and the curvedness , two parameters that
intuitive interfaces to interact with geometric models at the var- were introduced by Koenderink and van Doorn [209] and are
ious stages of initialization, deformation and eventual correction defined as follows:
of results, and concise transferal of clinical information from
images/models to the cardiologists. (11)
It is to be expected that approval by clinicians of a model-
based technique that provides functional parameters will depend
on close collaboration between technicians involved in image (12)
acquisition, computer scientists devoted to the development of
efficient modeling and model recovery techniques, and cardi- While is inversely proportional to the object size, defines
ologists providing feedback about the desired information and a continuous distribution of surface types ranging from cup-like
display methods, the validity of the assumptions and the design umbilic to peak-like umbilic points. It can be
of evaluation studies. shown that while the shape index is invariant by homothecy, the
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 19

curvedness is not. In this way, shape information and size can the cardiac cycle follow a similar evolution to the temporal vari-
be easily decoupled. ation of the principal strains obtained by Young et al. [100] using
The shape spectrum [210], , is a global shape index SA techniques.
defined as the fractional area of the LV with shape index value
, at time B. Model-Specific Shape Descriptors
Geometric Cardiogram (GCG): Azhari et al. [212] describe
(13) a method for classification of normal and abnormal LV geome-
tries by defining a “geometrical cardiogram” (GCG), a helical
sampling of the LV geometry from apex to base [213]. The GCG
where at end systole and at end diastole are subsequently analyzed via
total area of the surface ; a KLT to compress their information. A truncated set of the KLT
small region around the point ; basis vectors is used to project the GCG of individual patients
1-D Dirac delta function. into a lower dimensional space, and the mean square error be-
Cardiac deformation can be analyzed by tracking the shape tween the projected and original GCG is used to discriminate
index and curvedness of SSPs over time. SSPs are connected between normal and abnormal LV [214]. From this vectorial
surface patches whose points have similar shape indexes, i.e., representation LVV and EF [213], and WT [91] can also be com-
the shape index falls within a given range . Clarysse et puted.
al.have shown the potential applicability of these indexes by Deformable Superquadric and Related Models: One of the
analyzing phantoms of normal and diseased LVs. A LV model first 3-D primitives used to model the LV was the superquadric.
of dilated cardiomyopathy, and a model of an ischemic LV (both It is a natural extension of the simplified geometric models origi-
akinetic and hypokinetic in the left anterior coronary territory) nally used in 2DE [14] and angiocardiography [10]–[13]. Along
were generated using 4-D spherical harmonics. The curvedness with three main axes indicating principal dimensions, the su-
spectrum was significantly altered by both pathologies, even perquadric models can be endowed with additional parametric
when they were localized (ischemic models). Reduction of the deformations as linear tapering and bending [9], [109], FFDs
global function in the dilated myocardium had no significant [113], displacement fields [7], [8] or parametric functions pro-
repercussion on the shape index spectra. This could be an viding information about radial and longitudinal contraction,
indicator that this pathology mostly affects the magnitude of twisting motion, and deformation of the LV long-axis [5], [6]
motion only. An alternative to global analysis is to track the and wall thickness [6]. In particular, Park et al. [5], [6] suggest to
curvature parameters in predetermined regions. Clarysse et al. decompose deformation and motion into a few parametric func-
have tracked three reference points over time: the apex, a point tions that can be presented to the clinician in the form of simple
in the anterior wall, and a point in the cup of the pillar anchor. plots. All these functions are either independent of the total LV
Using the local temporal variation of the curvedness and shape volume (e.g., twisting) or can be normalized with respect to the
index, it was possible to distinguish between the normal and dimensions of the LV (e.g., radial and longitudinal contraction).
diseased model. A potential problem of this techniques is the This allows interpatient comparisons of contraction and shape
reliable tracking of SSPs. If local deformations are too large change.
the trace of points might be lost. Global motion analysis based on departure from an affine
Local stretching: Mishra et al. [87] have presented a model: Friboulet et al. [119] modeled the LV using a polyhe-
computational scheme to derive local epicardial stretching dral mesh at each frame of the cardiac cycle. The state of the
under conformal motion. In conformal motion, it is assumed LV was characterized by the center of gravity and the moments
that motion can be described by a spatially-variant but lo- of inertia of the polyhedral mesh. The deformation between two
cally isotropic stretching factor. In particular, for any two frames was hypothesized to follow an affine model. By defining
corresponding patches before and after motion, and , the a metric to compare two different polyhedral representations,
local stretching factor, , can be computed from the change the authors were able to quantify the difference between the
in Gaussian curvature and a polynomial stretching model by actual interframe deformation and the corresponding deforma-
means of the relationship tion derived from an affine motion model. Several parameters
of global motion are then derived: the temporal variation of the
longitudinal and transversal moments of inertia, and the propor-
(14) tion of total motion explained by the affine model. By means of
case studies it was demonstrated that these global indexes are
where able to discriminate between normal and highly
polynomial stretching model (linear or diseased LVs. On the other hand, the global nature
quadratic in [87]); of these indexes precludes the quantification of localized, inho-
, and coefficients of the first fundamental form mogeneous dysfunction of the LV.
[211]; Motion decomposition through planispheric transfor-
coordinates of a local parameterization of the mation: Declerck et al. [114] have proposed a canonical
surface patch. decomposition of cardiac motion into three components: radial
Mishra et al. [87] present a method to solve for in (14) and motion, twisting motion around the apico-basal axis, and
show that the local epicardial stretching factors computed over long-axis shortening. This decomposition is achieved through
20 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

a transformation of the Cartesian coordinates of the LV wall intersections is modeled as a mass-spring mesh of triangles.
to a planispheric space. In this space, a 4-D transformation Tag intersections are tracked by means of a correlation-based
is defined that regularly and smoothly parameterizes the external energy and, eventually, adding interactive constraints.
spatio-temporal variation of the LV wall. Since the canonical Finally, this technique allows to compute average strains on the
decomposition of motion can be directly obtained in the triangular patches. Another method for automatic tracking of
planispheric space, these descriptors also vary smoothly along the SPAMM grid has been presented by Kumar and Goldgof
the cardiac cycle. Finally, by tracking the position of material [70]. In the first frame, template matching is applied to provide
points over time in the planispheric space and subsequent an initial position of the tag grid. In this frame, the tag grid has
mapping to Cartesian coordinates, it is possible to reconstruct a high contrast and a regular arrangement. In the subsequent
their 3-D trajectories. frames, each line of the tag grid is independently tracked using
Modal analysis—deformation spectrum: Nastar and Ay- a discrete thick snake with a width of two pixels (the typical
ache have introduced the concept of deformation spectrum [127] tag width). The product of the image intensity in the two
which can be applied within the framework of modal analysis pixels is used as external energy to attract the snakes to the tag
[153]. The deformation spectrum is the graph representing the lines. Although these methods for extracting tag intersections
value of the modal amplitudes as a function of mode index. can be useful for 3-D deformation analysis, in the original
The deformation spectrum corresponding to the deformation be- formulations, the methods proposed in [72], [188], and [70]
tween two image frames describes which modes are excited in have all been applied to 2-D SA.
order to deform one object into another. It also gives an indi- There exist other approaches not based on snakes. Zhang
cation of the strain energy [127] of the deformation. As a con- et al. [219] decouple horizontal and vertical tag tracking via
sequence, a pure rigid deformation has zero strain energy. Two Fourier decomposition and spectral masking. In order to com-
deformations are said to be similar when the corresponding de- pensate for spectral cross modulation from perpendicular lines,
formation fields are equivalent up to a rigid transformation. In local histogram equalization is needed prior to spectral anal-
order measure the dissimilarity of two deformation fields, the ysis. Detection of tag lines is simplified in the preprocessed
lower-order modes related to rigid transformation are discarded. images and a simple local search can then be used to track
The difference of the deformation spectra so computed, can be local intensity minima (tag lines) over time. Kerwin and Prince
used to define a metric between shapes (e.g., the LV in two [197] have developed a method to simultaneously detect and
phases of the cardiac cycle) that can be applied to classify them track tag surfaces without the need for prior 2-D tag tracking.
into specific classes (e.g., normal/abnormal motion patterns). Tag surfaces are modeled using a kriging update model [220],
Finally, the amplitude of the different modes can be tracked over [221]. This model parameterizes tag surfaces using a global
time. Using Fourier spectral analysis, Nastar and Ayache have quadratic surface plus a local stochastic displacement. A re-
shown that these modes concentrate in a few low-frequency co- cursive spatio-temporal scheme is developed that updates the
efficients. kriging model. Measurements to update the model are obtained
through a local search for tag lines. In this search, a matched
filter is employed modeling the intensity profile across a tag line.
APPENDIX B Recently, Osman et al. [222], [223] have introduced and eval-
MR TAG LOCALIZATION TECHNIQUES uated a method for cardiac motion tracking based on the con-
Early attempts to model myocardial tissue deformation cept of HARP. The method uses isolated spectral peaks in the
tracked tag grid intersections manually over time [68]. Other Fourier domain of MR tagged images as a cue for tag tracking.
researchers [69], [73], [74], [141], [145], [191] have used The inverse Fourier transform of a spectral peak is a complex
semiautomatic tools [215]–[217], based on snakes, to locate image whose computed angle is called HARP image. In Osman
and track tag intersections and to define myocardial contours. et al. [222], [224] it is shown how this angle can be treated as a
Although they still require user interaction, these tools can material property that can be related to myocardial strain. This
speed up the manual procedure while reducing interobserver technique has the advantage that is fast, fully automatic and pro-
variability [218]. vides dense material properties. So far the method has been ap-
Young et al. [72] propose an interactive scheme for tag plied to 2-D images and, thus, only provides information about
tracking. The 2-D tag grid is modeled as a whole (active “apparent motion.” In Osman and Prince et al. [225], the au-
carpet). Separate manual segmentation of the LV boundaries thors present several visualization techniques that can be used
is required to compute myocardial strains only. Tag tracking to display the information provided by HARP images.
is performed using a modified snake [168] algorithm. Since
tags show up in these images as dark lines (intensity valleys),
the image intensity is used as external energy. Additionally REFERENCES
interactive guidance is supported by introducing user-defined [1] American Heart Association, 1999 Heart and Stroke Statistical Update
constrains. Only points in the myocardium mask are tracked (1998). [Online]. Available: http://www.americanheart.org
[2] M. L. Simoons, “The european heart survey: Inventory of the care of
in each frame while carpet points outside the myocardium cardiovascular diseases,” presented at the XXIst Eur. Soc. Cardiology
(inactive points) provide a weak form of continuity. Kraitchman Congress, Barcelona, Catalonia, 1999. Presentation of the ESC President
et al. [188] have introduced an interactive method for tracking Elect.
[3] T. McInerney and D. Terzopoulos, “Deformable models in medical
tag intersections. The method shares some features of the image analysis: A survey,” Med. Image Anal., vol. 1, no. 2, pp. 91–108,
active carpet model of Young et al. [72]. The carpet of tag June 1996.
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 21

[4] A. Singh, D. Goldgof, and D. Terzopoulos, Eds., Deformable Models [27] G. Coppini, R. Poli, and G. Valli, “Recovery of the 3-D shape of the left
in Medical Image Analysis. Los Alamitos, CA: IEEE Comput. Soc. ventricle from echocardiographic images,” IEEE Trans. Med. Imag., vol.
Press, 1998. 14, pp. 301–317, June 1995.
[5] J. Park, D. N. Metaxas, and L. Axel, “Analysis of left ventricular wall [28] G. D. Stetten and S. M. Pizer, “Medial-node models to identify and
motion based on volumetric deformable models and MRI-SPAMM,” measure objects in real-time 3-D echocardiography,” IEEE Trans. Med.
Med. Image Anal., vol. 1, no. 1, pp. 53–72, Mar. 1996. Imag., vol. 18, pp. 1025–1034, Oct. 1999.
[6] J. Park, D. N. Metaxas, A. A. Young, and L. Axel, “Deformable models [29] J. Maehle, K. Bjoernstad, S. Aakhus, H. G. Torp, and B. A. Angelsen,
with parameter functions for cardiac motion analysis from tagged MRI “Three-dimensional echocardiography for quantitative left ventricular
data,” IEEE Trans. Med. Imag., vol. 15, pp. 278–289, June 1996. wall motion analysis: A method for reconstruction of endocardial sur-
[7] T. O’Donnell, A. Gupta, and T. Boult, “The hybrid volumetric ventricu- face and evaluation of regional dysfunction,” Echocardiography, vol.
loid: New model for MR-SPAMM 3-D analysis,” in IEEE Comput. in 11, no. 4, pp. 397–408, 1994.
Cardiol.. Vienna, Austria, Sept. 1995, pp. 5–8. [30] G. Germano, J. Erel, H. Kiat, P. B. Kavanagh, and D. S. Berman, “Quan-
[8] T. O’Donnell, T. Boult, and A. Gupta, “Global models with parametric titative LVEF and qualitative regional function from gated 201-T1 per-
offsets as applied to cardiac motion recovery,” in Comput. Vis. Pattern fusion SPECT,” J. Nucl. Med., vol. 38, no. 5, pp. 749–754, May 1997.
Recogn. San Francisco, CA, June 1996, pp. 293–299. [31] T. L. Faber, E. M. Stokely, G. H. Templeton, M. S. Akers, R. W. Parkey,
[9] E. Bardinet, L. D. Cohen, and N. Ayache, “A parametric deformable and J. R. Corbett, “Quantification of three-dimensional left ventricular
model to fit unstructured 3D data,” Comput. Vis. Image Underst., vol. segmental wall motion and volumes from gated tomographic radionu-
71, no. 1, pp. 39–54, July 1998. clide ventriculograms,” J. Nucl. Med., vol. 30, no. 5, pp. 638–649, May
[10] J. W. Kennedy, W. A. Baxley, M. M. Figley, H. T. Dodge, and J. R. 1989.
Blackmon, “Quantitative angiocardiography: I. The normal left ventricle [32] T. L. Faber, M. S. Akers, R. M. Peshock, and J. R. Corbett, “Three-di-
in man,” Circulation, vol. 34, pp. 272–278, Aug. 1966. mensional motion and perfusion quantification in gated single-photon
[11] H. J. Hermann and S. H. Bartle, “Left ventricular volumes by angiocar- emission computed tomograms,” J. Nucl. Med., vol. 32, no. 12, pp.
diography: Comparison of methods and simplification of techniques,” 2311–2317, Dec. 1991.
Cardiovasc. Res., vol. 2, no. 4, pp. 404–414, Oct. 1968. [33] G. Germano, P. B. Kavanagh, and D. S. Berman, “An automatic ap-
[12] J. C. Davila and M. E. Sanmarco, “An analysis of the fit of mathematical proach to the analysis, quantitation and review of perfusion and function
models applicable to the measurement of left ventricular volume,” Amer. from myocardial perfusion SPECT images,” Int. J. Card. Imag., vol. 13,
J. Cardiol., vol. 18, pp. 31–42, July 1966. no. 4, pp. 337–346, Aug. 1997.
[13] H. T. Dodge, H. Sandler, W. A. Baxley, and R. R. Hawley, “Usefulnes [34] B. Ohnesorge, T. Flohr, C. Becker, A. Knez, A. F. Kopp, K. Fukuda, and
and limitations of radiographic methods for determining left ventricular M. F. Reiser, “Cardiac imaging with rapid, retrospective ECG synchro-
volume,” Amer. J. Cardiol., vol. 18, pp. 10–24, July 1966. nized multilevel spiral CT” (in German), Der Radiologe, vol. 40, no. 2,
[14] C. Vuille and A. E. Weyman, “Left ventricle I: General considerations, pp. 111–117, Feb. 2000.
assessment of chamber size and function,” in Principles and Practice of [35] K. Klingenbeck-Regn, S. Schaller, T. Flohr, B. Ohnesorge, A. F. Kopp,
Echocardiography, 2nd ed, A. E. Weyman, Ed. Philadelphia, PA: Lea and U. Baum, “Subsecond multislice computed tomography: Basics and
and Febiger, 1994. applications,” Eur. J. Radiol., vol. 31, no. 2, pp. 110–124, Aug. 1999.
[15] M. J. Raphael and R. M. Donaldson, “The normal heart: Methods of [36] D. P. Boyd and C. Haugland, “Recent progress in electron beam tomog-
examination,” in Textbook of Radiology and Imaging, 6th ed, D. Sutton, raphy,” Med. Imag. Tech., vol. 11, pp. 578–585, 1993.
Ed. New York: Churchill Livingstone, 1998, pp. 541–565. [37] J. Magn. Reson. Imag. (Special Issue: Cardiovascular MRI), vol. 10, no.
[16] H. A. McCann, J. C. Sharp, T. M. Kinter, C. N. McEwan, C. Barillot, and 5, Nov. 1999.
J. F. Greenleaf, “Multidimensional ultrasonic imaging for cardiology,” [38] A. M. Keller, R. M. Peshock, C. R. Malloy, L. M. Buja, R. Nunnally,
Proc. IEEE, vol. 76, pp. 1063–1071, Sept. 1988. R. W. Parkey, and J. T. Willerson, “In vivo measurement of myocardial
[17] T. R. Nelson, D. B. Downey, D. H. Pretorius, and A. Fenster, Three-Di- mass using nuclear magnetic resonance imaging,” J. Amer. Coll. Car-
mensional Ultrasound. Philadelphia, PA: Lippincott Williams & diol., vol. 8, no. 1, pp. 113–137, July 1986.
Wilkins, 1999. [39] G. B. Cranney, C. S. Lotan, L. Dean, W. Baxley, A. Bouchard, and G.
[18] R. A. Robb, E. A. Hoffman, L. J. Sinak, L. D. Harris, and E. L. Ritman, M. Pohost, “Left ventricular volume measurement using cardiac axis
“High-speed three-dimensional X-ray computed tomography: The dy- nuclear magnetic resonance imaging,” Circulation, vol. 82, no. 1, pp.
namic spatial reconstructor,” Proc. IEEE, vol. 71, pp. 308–319, Mar. 154–163, July 1990.
1983. [40] M. C. Dulce, G. H. Mostbeck, K. K. Friese, G. R. Caputo, and C. B.
[19] D. P. Boyd and M. J. Lipton, “Cardiac computed tomography,” Proc. Higgings, “Quantification of the left ventricular volumes and function
IEEE, vol. 71, pp. 298–307, Mar. 1983. with cine MR imaging: Comparison of geometric models with three-
[20] C. B. Higgins, “Overview of MR of the heart—1986,” Amer. J. dimensional data,” Radiology, vol. 188, no. 2, pp. 371–376, Aug. 1993.
Roentgenol., vol. 146, no. 5, pp. 907–918, May 1986. [41] R. J. van der Geest, A. de Roos, E. E. van der Wall, and J. H. C. Reiber,
[21] G. A. Wright, “Magnetic resonance imaging,” IEEE Signal Processing “Quantitative analysis of cardiovascular MR images,” Int. J. Card.
Mag., pp. 56–66, Jan. 1997. Imag., vol. 13, no. 3, pp. 247–258, June 1997.
[22] J. U. Quistgaard, “Signal acquisition and processing in medical diag- [42] C. H. Lorenz, E. S. Walker, V. L. Morgan, S. S. Klein, and T. P. Graham
nostic ultrasound,” IEEE Signal Processing Mag., pp. 67–74, Jan. 1997. Jr., “Normal human right and left ventricular mass, systolic function, and
[23] M. Ariet, E. A. Geiser, S. M. Lupkiewicz, D. A. Conetta, and C. R. Conti, gender differences by cine magnetic resonance imaging,” J. Cardiovasc.
“Evaluation of a three-dimensional reconstruction to compute left ven- Magn. Res., vol. 1, no. 1, pp. 7–21, 1999.
tricular volume and mass,” Amer. J. Cardiol., vol. 54, no. 3, pp. 415–420, [43] J. T. Marcus, M. J. W. Götte, L. K. de Waal, M. R. Stam, R. J. van
Aug. 1984. der Geest, R. M. Heethaar, and A. C. van Rossum, “The influence of
[24] G. Treece, R. Prager, A. Gee, and L. Berman, “Volume measurements through-plane motion on left ventricular volumes measured by magnetic
in sequential freehand 3-D ultrasound,” in Lecture Notes in Computer resonance imaging: Implications for image acquisition and analysis,” J.
Science, A. Kuba, M. Sámal, and A. Todd-Pokropek, Eds. Berlin, Ger- Cardiovasc. Magn. Res., vol. 1, no. 1, pp. 1–6, 1999.
many: Springer-Verlag, 1999, vol. 1613, Inform. Processing Med. Imag., [44] R. Beyar, E. P. Shapiro, W. L. Graves, W. J. Rogers, W. H. Guier, G.
pp. 70–83. A. Carey, R. L. Soulen, E. A. Zerhouni, M. L. Weisfeldt, and J. L.
[25] A. S. Gopal, D. L. King, J. Katz, L. M. Boxt, D. L. King Jr., and M. Weiss, “Quantification and validation of left ventricular wall thickening
Y. Shao, “Three-dimensional echocardiographic volume computation by by a three-dimensional volume element magnetic resonance imaging ap-
polyhedral surface reconstruction: In vitro validation and comparison to proach,” Circulation, vol. 81, no. 1, pp. 297–307, Jan. 1990.
magnetic resonance imaging,” J. Amer. Soc. Echocardiogr., vol. 5, no. [45] E. R. McVeigh, “MRI of myocardial function: Motion tracking tech-
2, pp. 115–124, Mar. 1992. niques,” Magn. Reson. Imag., vol. 14, no. 2, pp. 137–150, 1996.
[26] M. E. Legget, D. F. Leotta, E. L. Bolson, J. A. McDonald, R. W. Martin, [46] N. R. Clark, N. Reichek, P. Bergey, E. A. Hoffman, D. Brownson,
X. N. Li, C. M. Otto, and F. H. Sheehan, “System for quantitative three L. Palmon, and L. Axel, “Circumferential myocardial shortening in
dimensional echocardiography of the left ventricle based on magnetic the normal human left ventricle. Assessment by magnetic resonance
field position and orientation sensing system,” IEEE Trans. Biomed. imaging using spatial modulation of magnetization,” Circulation, vol.
Eng., vol. 45, pp. 494–504, Apr. 1998. 84, no. 1, pp. 67–74, July 1991.
22 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

[47] E. A. Zerhouni, D. M. Parish, W. J. Rogers, A. Yang, and E. P. Shapiro, [67] F. L. Villarreal, L. K. Waldman, and W. Y. W. Lew, “Technique for mea-
“Human heart: Tagging with MR imaging—A method for noninvasive suring regional two-dimensional finite strains in canine left ventricle,”
assessment of myocardial motion,” Radiology, vol. 169, no. 1, pp. Circ. Res., vol. 62, no. 4, pp. 711–721, Apr. 1988.
59–63, Oct. 1988. [68] A. A. Young and L. Axel, “Three-dimensional motion and deformation
[48] L. Axel and L. Dougherty, “Heartwall motion: Improved method of spa- of the heart wall: Estimation with spatial modulation of magneti-
tial modulation of magnetization for MR imaging,” Radiology, vol. 172, zation—A model-based approach,” Radiology, vol. 185, no. 1, pp.
no. 2, pp. 349–350, Aug. 1989. 241–247, Oct. 1992.
[49] K. T. Weber and E. W. Hawthorne, “Descriptors and determinants of [69] C. C. Moore, W. G. O’Dell, E. R. McVeigh, and E. A. Zerhouni, “Calcu-
cardiac shape: An overview,” Fed. Proc., vol. 40, no. 7, pp. 2005–2010, lation of three-dimensional left ventricular strains form biplanar tagged
May 1981. MR images,” J. Magn. Reson. Imag., vol. 2, no. 2, pp. 165–175, Mar.
[50] G. B. Mancini, S. F. De Boe, E. Anselmo, S. B. Simon, M. T. Le Free, and 1992.
R. A. Vogel, “Quantitative regional curvature analysis: An application [70] S. Kumar and D. Goldgof, “Automatic tracking of SPAMM grid and the
of shape determination for the assessment of segmental left ventricular estimation of deformation parameters from cardiac MR images,” IEEE
function in man,” Amer. Heart J., pt. 1, vol. 113, no. 2, pp. 326–334, Trans. Med. Imag., vol. 13, pp. 122–132, Mar. 1994.
Feb. 1987. [71] S. E. Fischer, G. C. McKinnon, M. B. Scheidegger, W. Prins, D. Meier,
[51] G. B. J. Mancini, S. F. DeBoe, M. J. McGillem, and E. R. Bates, “Quan- and P. Boesiger, “True myocardial motion tracking,” Magn. Res. Med.,
titative regional curvature analysis: A prospective evaluation of ventric- vol. 31, no. 4, pp. 401–413, Apr. 1994.
ular shape and wall motion measurements,” Amer. Heart J., pt. 1, vol. [72] A. A. Young, D. L. Kraitchman, L. Dougherty, and L. Axel, “Tracking
116, no. 6, pp. 1611–1621, Dec. 1988. and finite element analysis of stripe deformation in magnetic resonance
[52] D. A. Kass, T. A. Traill, M. Keating, P. I. Altieri, and W. L. Maughan, tagging,” IEEE Trans. Med. Imag., vol. 14, pp. 413–421, Sept. 1995.
“Abnormalities of dynamic ventricular shape change in patients with [73] T. S. Denney Jr. and J. L. Prince, “Reconstruction of 3D left ventricular
aortic and mitral valvular regurgitation: Assessment by Fourier shape motion from planar tagged cardiac MR images: An estimation theoretic
analysis and global geometric indexes,” Circ. Res., vol. 62, no. 1, pp. approach,” IEEE Trans. Med. Imag., vol. 14, pp. 1–11, Dec. 1995.
127–138, Jan. 1988. [74] W. S. Kerwin and J. L. Prince, “Cardiac material markers from
[53] G. Barletta, M. Baroni, R. del Bene, A. Toso, and F. Fantini, “Regional tagged MR images,” Med. Image Anal., vol. 2, no. 4, pp. 339–353,
and temporal nonuniformity of shape and wall movement in the normal Dec. 1998.
left ventricle,” Cardiology, vol. 90, no. 3, pp. 195–201, Dec. 1998. [75] J. L. Prince and E. R. McVeigh, “Motion estimation form tagged MR
[54] C. Kondo, G. R. Caputo, R. Smelka, E. Foster, A. Shimakawa, and C. B. image sequences,” IEEE Trans. Med. Imag., vol. 11, pp. 238–249, June
Higgins, “Right and left ventricular stroke volume measurements with 1992.
velocity-encoded cine MR imaging: In vitro and in vivo validation,” [76] S. C. Amartur and H. J. Vesselle, “A new approach to study cardiac
Amer. J. Roentgenol., vol. 157, no. 1, pp. 9–16, July 1991. motion: The optical flow of cine MR images,” Magn. Res. Med., vol.
[55] L. D. Hillis, B. G. Firth, and M. D. Winniford, “Analysis of factors af- 29, no. 1, pp. 59–67, Jan. 1993.
fecting the variability of Fick versus indicator dilution measurements of [77] M. Tistarelli and G. Marcennaro, “Using optical flow to analyze the
cardiac output,” Amer. J. Cardiol., vol. 56, no. 12, pp. 764–768, Nov. motion of human body organs from bioimages,” in Proc. Math. Meth.
1986. Biomed. Imag. Anal., Seattle, WA, USA, 1994, pp. 100–109.
[56] J. C. McEachen, A. Neohorai, and J. S. Duncan, “A recursive filter for [78] T. S. Denney Jr. and J. L. Prince, “Optimal brightness functions for op-
temporal analysis of cardiac motion,” in Proc. Math. Meth. Biomed. tical flow estimation of deformable motion,” IEEE Trans. Imag. Pro-
Imag. Anal., Seattle, WA, 1994, pp. 124–133. cessing, vol. 3, pp. 178–191, Mar. 1994.
[57] M. J. Potel, J. M. Rubin, S. A. MacKay, A. M. Aisen, J. Al-Sadir, and R. [79] , “A frequency domain performance analysis of Horn and
E. Sayre, “Methods for evaluating cardiac wall motion in three dimen- Schunck’s optical flow algorithm for deformable motion,” IEEE Trans.
sions using bifurcation points of the coronary arterial tree,” Inv. Rad., Imag Processing, vol. 4, pp. 1324–1327, Sept. 1995.
vol. 18, no. 1, pp. 47–57, Jan. 1983. [80] L. Dougherty, J. C. Asmuth, A. S. Blom, L. Axel, and R. Kumar, “Vali-
[58] H. C. Kim, B. G. Min, M. M. Lee, J. D. Seo, Y. W. Lee, and M. C. Han, dation of an optical flow method for tag displacement estimation,” IEEE
“Estimation of local cardiac wall deformation and regional wall stress Trans. Med. Imag., vol. 18, pp. 359–363, Apr. 1999.
from biplane coronary cineangiograms,” IEEE Trans. Biomed. Eng., vol. [81] S. S. Beauchemin and J. L. Barron, “The computation of optical flow,”
32, pp. 503–512, July 1985. ACM Comput. Surveys, vol. 27, no. 3, pp. 444–467, Sept. 1995.
[59] C. W. Chen and T. S. Huang, “Epicardial motion and deformation es- [82] S. M. Song and R. M. Leahy, “Computation of 3D velocity fields from
timation from coronary artery bifurcation points,” in Proc. Int. Conf. 3D cine CT images of a human heart,” IEEE Trans. Med. Imag., vol. 10,
Comput. Vis., Osaka, Japan, Dec. 1990. pp. 295–306, Sept. 1991.
[60] A. A. Young, P. J. Hunter, and B. H. Smaill, “Estimation of epicardial [83] S. M. Song, R. M. Leahy, D. P. Boyd, B. H. Brundage, and S. Napel,
strain using the motions of coronary bifurcations in biplane cineangiog- “Determining cardiac velocity fields and intraventricular pressure dis-
raphy,” IEEE Trans. Biomed. Eng., vol. 39, pp. 526–531, May 1992. tribution from a sequence of ultrafast CT cardiac images,” IEEE Trans.
[61] D. C. Harrison, A. Goldblatt, E. Braunwald, G. Glick, and D. T. Mason, Med. Imag., vol. 14, pp. 386–397, June 1994.
“Studies on cardiac dimensions in intact unanesthetized man. I. Descrip- [84] J.-M. Gorce, D. Friboulet, and I. E. Magnin, “Estimation of three-di-
tion of techniques and their validation,” Circ. Res., vol. 13, pp. 448–455, mensional cardiac velocity fields: Assessment of a differential method
1963. and application to three-dimensional CT data,” Med. Image Anal., vol.
[62] J. S. Rankin, P. A. McHale, C. E. Artentzen, D. Ling, J. C. Green- 1, no. 3, pp. 245–261, Mar. 1996.
field, and R. W. Anderson, “The three-dimensional dynamic geometry [85] F. G. Meyer, R. T. Constable, A. J. Sinusas, and J. S. Duncan, “Tracking
of the left ventricle in the conscious dog,” Circ. Res., vol. 39, no. 3, pp. myocardial deformation using phase contrast MR velocity fields: A sto-
304–313, Sept. 1976. chastic approach,” IEEE Trans. Med. Imag., vol. 15, pp. 453–465, Aug.
[63] R. W. Brower, H. T. ten Katen, and G. J. Meester, “Direct method 1996.
for determining regional myocardial shortening after bypass surgery [86] A. A. Amini and J. S. Duncan, “Bending and stretching models for LV
from radiopaque markers in man,” Amer. J. Cardiol., vol. 41, no. 7, pp. wall motion analysis from curves and surfaces,” Imag. Vis. Comput., vol.
1222–1229, June 1978. 10, no. 6, pp. 418–430, July 1992.
[64] N. B. Ingels Jr., G. T. Daughters, E. B. Stinson, and E. L. Alderman, [87] S. K. Mishra, D. B. Goldgof, T. S. Huang, and C. Kambhamettu, “Cur-
“Evaluation of methods for quantitating left ventricular segmental wall vature-based nonrigid motion analysis from 3D point correspondences,”
motion in man using myocardial markers as a standard,” Circulation, Int. J. Image Syst. Tech., vol. 4, pp. 214–225, 1992.
vol. 61, no. 5, pp. 966–972, May 1980. [88] D. Friboulet, I. E. Magnin, C. Mathieu, A. Pommert, and K. H. Hoehne,
[65] G. D. Meier, A. A. Bove, W. P. Santamore, and P. R. Lynch, “Contractile “Assessment and visualization of the curvature of the left ventricle from
function in canine right ventricle,” Amer. J. Physiol., vol. 239, no. 6, pp. 3D medical images,” Comput. Med. Imag. & Graph., vol. 17, no. 4–5,
H794–804, Dec. 1980. pp. 257–262, July 1993.
[66] T. Arts, W. C. Hunter, A. S. Douglas, A. M. M. Muijtjens, J. W. Corsel, [89] C. Kambhamettu and D. B. Goldgof, “Curvature-based approach to
and R. S. Reneman, “Macroscopic three-dimensional motion patterns of point correspondence recovery in conformal nonrigid motion,” Comput.
the left ventricle,” Adv. Exp. Med. Biol., vol. 346, pp. 383–392, 1993. Vis. Graph. Imag. Processing, vol. 60, no. 1, pp. 26–43, July 1994.
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 23

[90] S. Benayoun and N. Ayache, “Dense and nonrigid motion estimation [113] E. Bardinet, N. Ayache, and L. D. Cohen, “Tracking and motion analysis
in sequences of medical images using differential constraints,” Int. J. of the left ventricle with deformable superquadrics,” Med. Image Anal.,
Comput. Vis., vol. 26, no. 1, pp. 25–40, Jan. 1998. vol. 1, no. 2, pp. 129–150, June 1996.
[91] H. Azhari, S. Sideman, J. L. Weiss, E. P. Shapiro, M. L. Weisfeldt, W. [114] J. Declerck, J. Feldmar, and N. Ayache, “Definition of a four-dimen-
L. Graves, W. J. Rogers, and R. Beyar, “Three-dimensional mapping sional continuous planisferic transformation for the tracking and the
of acute ischemic regions using MRI: Wall thickening versus motion analysis of the left-ventricle motion,” Med. Image Anal., vol. 2, no. 2,
analysis,” Amer. J. Physiol., pt. 2, vol. 259, no. 5, pp. H1492–1503, Nov. pp. 197–213, June 1998.
1990. [115] Y. Sato, M. Moriyama, M. Hanayama, H. Naito, and S. Tamura, “Ac-
[92] A. M. Taratorin and S. Sideman, “3D functional mapping of left ventric- quiring 3D models of nonrigid moving objects from time and viewpoint
ular dynamics,” Comput. Med. Imag. Graph., vol. 19, no. 1, pp. 113–129, varying image sequences: A step toward left ventricle recovery,” IEEE
Jan. 1995. Trans. Pattern Anal. Machine Intell., vol. 19, pp. 253–259, Mar. 1997.
[93] F. H. Sheehan, E. L. Bolson, H. T. Dodge, D. G. Mathey, J. Schofer, and [116] E. A. Geiser, S. M. Lupkiewicz, L. G. Christie, M. Ariet, D. A. Conetta,
H. K. Woo, “Advantages and applications of the centerline method for and C. R. Conti, “A framework for three-dimensional time-varying re-
characterizing regional ventricular function,” Circulation, vol. 74, no. 2, construction of the human left ventricle: Sources of error and estimation
pp. 293–305, Aug. 1986. of their magnitude,” Comput. Biomed. Res., vol. 13, no. 3, pp. 225–241,
[94] V. G. M. Buller, R. J. van der Geest, M. D. Kool, and J. H. C. Reiber, June 1980.
“Accurate three-dimensional wall thickness measurement from multi- [117] E. A. Geiser, M. Ariet, D. A. Conetta, S. M. Lupkiewicz, L. G. Christie
slice short-axis MR imaging,” in Proc. IEEE Computers in Cardiology. Jr., and C. R. Conti, “Dynamic three-dimensional echocardiographic re-
Los Alamitos, CA, Sept. 1994, pp. 245–248. construction of the intact human left ventricle: Technique and initial ob-
[95] E. L. Bolson and F. H. Sheehan, “Centersurface model for 3D analysis of servations in patients,” Amer. Heart J., vol. 103, no. 6, pp. 1056–1065,
regional left biventricular function,” in Proc. IEEE Computers in Car- June 1982.
diology, London, U.K., Sept. 1993, pp. 735–738. [118] T. L. Faber, E. M. Stokely, R. M. Peshock, and J. R. Corbett, “A model-
[96] E. L. Bolson, F. H. Sheehan, M. E. Legget, H. Jin, J. A. McDonald, based four-dimensional left ventricular surface detector,” IEEE Trans.
P. D. Sampson, R. W. Martin, G. Bashein, and C. M. Otto, “Applying Med. Imag., vol. 10, pp. 321–329, Sept. 1991.
the centersurface model to 3-D reconstructions of the left ventricle for [119] D. Friboulet, I. E. Magnin, and D. Revel, “Assessment of a model for
regional functional analysis,” in Proc. IEEE Computers in Cardiology, overall left ventricular three-dimensional motion from MRI data,” Int.
Vienna, Austria, Sept. 1995, pp. 63–67. J. Card. Imag., vol. 8, no. 3, pp. 175–190, Sept. 1992.
[97] Y. C. Fung, Biomechanics: Motion, Flow, Stress, and Growth. New [120] W.-C. Huang and D. Goldgof, “Adaptive-size meshes for rigid and non-
York: Springer Verlag, 1990. rigid analysis and synthesis,” IEEE Trans. Pattern Anal. Machine Intell.,
[98] A. L. Yettram and C. A. Vinson, “Geometric modeling of the human left vol. 15, pp. 611–616, June 1993.
ventricle,” J. Biomech., vol. 101, pp. 221–223, 1979. [121] T. L. Faber, C. D. Cooke, J. W. Peifer, R. I. Pettigrew, J. P. Vansant, J.
[99] A. L. Yettram, C. A. Vinson, and D. G. Gibson, “Computer modeling R. Leyendecker, and E. V. Garcı´ a, “Three-dimensional displays of left
of the human left ventricle,” J. Biomech., vol. 104, no. 2, pp. 148–152, ventricular epicardial surface from standard cardiac SPECT perfusion
May 1982. quantification techniques,” J. Nucl. Med., vol. 36, no. 4, pp. 697–703,
[100] A. A. Young, P. J. Hunter, and B. H. Smaill, “Epicardial surface estima- Apr. 1995.
tion from coronary angiograms,” Comput. Vis. Graph. Imag. Processing, [122] G. Germano, H. Kiat, P. B. Kavanagh, M. Moriel, M. Mazzanti, H.-T. Su,
vol. 47, no. 1, pp. 111–127, July 1989. K. F. van Train, and D. S. Berman, “Automatic quantification of ejection
[101] F. G. Spinale, B. A. Carabello, and F. A. Crawford Jr., “Right ventric- fraction from gated myocardial perfusion SPECT,” J. Nucl. Med., vol.
ular function and three-dimensional modeling using computer-aided de- 36, no. 11, pp. 2138–2147, Nov. 1995.
sign,” J. Appl. Physiol., vol. 68, no. 4, pp. 1707–1716, Apr. 1990. [123] G. Germano, P. B. Kavanagh, J. Chen, P. Waechter, H.-T. Su, H. Kiat,
[102] A. Pentland and B. Horowitz, “Recovery of nonrigid motion and struc- and D. S. Berman, “Operator-less processing of myocardial perfusion
ture,” IEEE Trans. Pattern Anal. Machine Intell., vol. 13, pp. 730–742, SPECT studies,” J. Nucl. Med., vol. 36, no. 6, pp. 2127–2132, Nov. 1995.
July 1991. [124] T. McInerney and D. Terzopoulos, “A dynamic finite element surface
[103] J. C. Cauvin, J. Y. Boire, M. Zanca, J. M. Bonny, J. Maublant, and model for segmentation and tracking in multidimensional medical
A. Veyre, “3D modeling in myocardial 201TL SPECT,” Comput. Med. images with application to 4D image analysis,” Comput. Med. Imag.
Imag. Graph., vol. 17, no. 4–5, pp. 345–350, July 1993. Graph., vol. 19, no. 1, pp. 69–83, Jan. 1995.
[104] F. P. Czegledy and J. Katz, “A new geometric description of the [125] S. Ranganath, “Contour extraction from cardiac MRI studies using
right ventricle,” J. Biomed. Eng., vol. 15, no. 5, pp. 387–391, Sept. snakes,” IEEE Trans. Med. Imag., vol. 14, pp. 56–64, June 1995.
1993. [126] H. K. Tu, A. Matheny, D. B. Goldgof, and H. Bunke, “Left ventricular
[105] T. Gustavsson, R. Pascher, and K. Caidahl, “Model based dynamic 3D boundary detection from spatio-temporal volumetric computed tomog-
reconstruction and display of the left ventricle from 2D cross-sectional raphy images,” Comput. Med. Imag. Graph., vol. 19, no. 1, pp. 27–46,
echocardiograms,” Comput. Med. Imag. Graph., vol. 17, no. 4–5, pp. Jan. 1995.
273–278, July 1993. [127] C. Nastar and N. Ayache, “Frequency-based nonrigid motion analysis:
[106] M. S. Sacks, C. J. Chuong, G. H. Templeton, and R. Peshock, “In vivo Application to four dimensional medical images,” IEEE Trans. Pattern
3-D reconstruction and geometric characterization of the right ventric- Anal. Machine Intell., vol. 18, pp. 1067–1079, Nov. 1996.
ular free wall,” Ann. Biomed. Eng., vol. 21, no. 3, pp. 263–275, May [128] D. Rueckert and P. Burger, “Geometrically deformable templates for
1993. shape-based segmentation and tracking in cardiac MR image,” in Lec-
[107] C. W. Chen, T. S. Huang, and M. Arrott, “Modeling, analysis, and visu- ture Notes in Computer Science. Venice, Italy: Springer-Verlag, 1997,
alization of left ventricle shape, and motion by hierarchical decomposi- vol. 1223, Energy MinImag. Meth. Comput. Vis. & Pattern Recogn..
tion,” IEEE Trans. Pattern Anal. Machine Intell., vol. 16, pp. 342–356, [129] P. Shi, A. J. Sinusas, R. T. Constable, E. Ritman, and J. S. Duncan,
Apr. 1994. “Point-tracked quantitative analysis of left ventricular surface motion
[108] S. Denslow, “An ellipsoidal shell model for volume estimation of the from 3D image sequences: Algorithms and validation,” IEEE Trans.
right ventricle from magnetic resonance images,” Acad. Radiol., vol. 1, Med. Imag., vol. 19, pp. 36–50, Jan. 2000.
no. 4, pp. 345–351, Dec. 1994. [130] J. S. Duncan, P. Shi, T. Constable, and A. Sinusas, “Physical and geo-
[109] C. W. Chen, J. Luo, K. J. Parker, and T. S. Huang, “CT volumetric metrical modeling for image-based recovery of left ventricular deforma-
data-based left ventricle motion estimation: An integrated approach,” tion,” Prog. Biophys. Mol. Biol., vol. 69, no. 1–2, pp. 333–351, 1998.
Comput. Med. Imag. Graph., vol. 19, no. 1, pp. 85–100, Jan. 1995. [131] J. Montagnat, H. Delingette, and G. Malandain, “Cylindrical echocar-
[110] A. Goshtasby and D. A. Turner, “Segmentation of cardiac cine MR im- diographic image segmentation based on 3D deformable models,” in
ages for extraction of right and left ventricular chambers,” IEEE Trans. Lecture Notes in Computer Science, 1999, vol. 1679, Medical Imaging
Med. Imag., vol. 14, pp. 56–64, Mar. 1995. Computing & Computer-Assisted Intervention, pp. 168–175.
[111] A. Matheny and D. M. Goldgof, “The use of three- and four-dimensional [132] S. Biedenstein, M. Schäfers, L. Stegger, T. Kuwert, and O. Schober,
surface harmonics for rigid and nonrigid shape recovery and representa- “Three-dimensional contour detection of left ventricular myocardium
tion,” IEEE Trans. Pattern Anal. Machine Intell., vol. 17, pp. 967–981, using elastic surfaces,” Eur. J. Nucl. Med., vol. 26, no. 3, pp. 201–207,
Oct. 1995. Mar. 1999.
[112] L. H. Staib and J. S. Duncan, “Model-based deformable surface finding [133] A. Yezzi, A. Tannenbaum, S. Kichenassamy, and P. Olver, “A gradient
for medical images,” IEEE Trans. Med. Imag., vol. 15, pp. 720–731, surface approach to 3-D segmentation,” in Proc. IS&T 49th Annu. Conf.,
May 1996. Minneapolis, MN, May 1996.
24 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 20, NO. 1, JANUARY 2001

[134] A. Yezzi, S. Kichenassamy, A. Kumar, P. Olver, and A. Tannenbaum, [156] C. W. Chen, J. B. Luo, and K. J. Parker, “Image segmentation via adap-
“A geometric snake model for segmentation of medical imagery,” IEEE K
tive -mean clustering and knowledge-based morphological operations
Trans. Med. Imag., vol. 16, pp. 199–209, Apr. 1997. with biomedical applications,” IEEE Trans. Imag. Processing, vol. 7, pp.
[135] Y.-H. Tseng, J.-N. Hwang, and F. H. Sheehan, “Three-dimensional ob- 1673–1683, Dec. 1998.
ject representation and invariant recognition using continuous distance [157] J. J. Hopfield, “Neurons with graded responses have collective compu-
transform neural networks,” IEEE Trans. Neural Networks, vol. 8, pp. tational properties like those of two-state neurons,” in Proc. Nat. Acad.
141–147, Jan. 1997. Sci., vol. 81, May 1984, pp. 3088–3092.
[136] W. J. Niessen, B. M. ter Haar Romeny, and M. A. Viergever, “Geodesic [158] O. Monga, R. Deriche, and J.-M. Rocchisani, “3D edge detection using
deformable models for medical image analysis,” IEEE Trans. Med. recursive filtering: Application to scanner images,” Comput. Vis. Graph.
Imag., vol. 17, pp. 634–641, Aug. 1998. Imag. Processing: Imag. Underst., vol. 53, no. 1, pp. 76–87, Jan. 1991.
[137] B. P. F. Lelieveldt, R. J. van der Geest, M. R. Rezaee, J. G. Bosch, and J. [159] J. Declerck, J. Feldmar, M. L. Goris, and F. Betting, “Automatic reg-
H. C. Reiber, “Anatomical model matching with fuzzy implicit surfaces istration and alignment on a template of cardiac stress and rest SPECT
for segmentation of thoracic volume scan,” IEEE Trans. Med. Imag., vol. images,” IEEE Trans. Med. Imag., vol. 16, pp. 727–737, Dec. 1997.
18, pp. 231–238, Mar. 1999. [160] P. J. Besl and N. D. McKay, “A method for registration of 3D shapes,”
[138] L. L. Creswell, S. G. Wyers, J. S. Pirolo, W. H. Perman, M. W. Vannier, IEEE Trans. Pattern Anal. Machine Intell., vol. 14, pp. 239–255, Feb.
and M. K. Pasque, “Mathematical modeling of the heart using magnetic 1992.
resonance imaging,” IEEE Trans. Med. Imag., vol. 11, pp. 581–589, Dec. [161] J. Feldmar and N. Ayache, “Rigid, affine and locally affine registration
1992. of free-form surfaces,” Int. J. Comput. Vis., vol. 18, no. 2, pp. 99–119,
[139] J. S. Pirolo, S. J. Bresina, L. L. Creswell, K. W. Myers, B. A. Szabó, M. May 1996.
W. Vannier, and M. K. Pasque, “Mathematical three-dimensional solid [162] T. W. Sederberg and S. R. Parry, “Free-form deformation of solid geo-
modeling of biventricular geometry,” Ann. Biomed. Eng., vol. 21, no. 3, metric models,” Proc. SIGGRAPH’86, vol. 20, pp. 151–160.
pp. 199–219, May 1993. [163] G. Germano and D. Berman, Clinical Gated Cardiac
[140] E. Haber, D. N. Metaxas, and L. Axel, “Three-dimensional geometric SPECT. Armonk, NY: Futura, 1999, ch. Quantitative gated perfusion
modeling of cardiac right and left ventricles,” presented at the Biomed. SPECT, pp. 115–146.
Eng. Soc. Annu. Meeting, San Diego, CA, 1997. [164] A. Goshtasby, “Design and recovery of 2-D and 3-D shapes using ra-
[141] , “Motion analysis of the right ventricle from MRI images,” in Com- tional Gaussian curves and surfaces,” Int. J. Comput. Vis., vol. 10, no. 3,
puters in Cardiology, vol. 1496, Medical Imaging Computing & Com- pp. 233–256, 1993.
puter-Assisted Intervention, pp. 177–188. [165] D. N. Metaxas, Physics-Based Deformable Models: Applications to
[142] P. Shi, A. J. Sinusas, R. T. Constable, and J. S. Duncan, “Volumetric Computer Vision, Graphics and Medical Imaging. Norwell, MA:
deformation analysis using mechanics-based data fusion: Application in Kluwer Academic, 1996.
cardiac motion recovery,” Int. J. Comput. Vis., vol. 35, no. 1, pp. 87–107, [166] L. D. Cohen, “On active contour models and balloons,” Comput. Vis.
Nov. 1999. Graph. Imag. Processing: Imag. Underst., vol. 53, no. 2, pp. 211–218,
[143] M. Kuwahara and S. Eiho, “3-D heart image reconstructed from MRI Mar. 1991.
data,” Comput. Med. Imag. Graph., vol. 15, no. 4, pp. 241–246, July [167] H. Delingette, “General object reconstruction based on simplex
1991. meshes,” Int. J. Comput. Vis., vol. 32, no. 2, pp. 111–146, Sept. 1999.
[144] K. A. Bartels, A. C. Bovik, S. J. Aggarwal, and K. R. Diller, “The anal- [168] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour
ysis of biological shape change from multidimensional dynamic im- models,” Int. J. Comput. Vis., vol. 1, no. 4, pp. 321–331, Jan. 1988.
ages,” Comput. Med. Imag. Graph., vol. 17, no. 2, pp. 89–99, Mar. 1993. [169] D. Terzopoulos, “Regularization of inverse visual problems involving
[145] W. G. O’Dell, C. C. Moore, W. C. Hunter, E. A. Zerhouni, and E. R. discontinuties,” IEEE Trans. Pattern Anal. Machine Intell., vol. 8, pp.
McVeigh, “Three-dimensional myocardial deformations: Calculation 413–424, Apr. 1986.
with displacement field fitting to tagged MR images,” Radiology, vol. [170] J. Kittler and J. Illingworth, “Relaxation labeling algorithms—A re-
195, no. 3, pp. 829–835, June 1995. view,” Imag. Vis. Comput., vol. 3, no. 4, pp. 206–216, Nov. 1985.
[146] M. J. Moulton, L. L. Creswell, S. W. Downing, R. L. Actis, B. A. Szabo, n
[171] D. F. Watson, “Computing the -dimensional Delaunay tesselation
M. W. Vannier, and M. K. Pasque, “Spline surface interpolation for with application to Voronoi polytopes,” Comput. J., vol. 24, no. 2, pp.
calculating 3-D ventricular strains from MRI tissue tagging,” Amer. J. 167–172, May 1981.
Physiol., pt. 2, vol. 270, no. 1, pp. H281–297, Jan. 1996. [172] A. Chakraborty, L. H. Staib, and J. S. Duncan, “Deformable boundary
[147] P. Radeva, A. A. Amini, and J. Huang, “Deformable B-solids and im- finding in medical images by integrating gradient and region informa-
plicit snakes for 3D localization and tracking of SPAMM MRI data,” tion,” IEEE Trans. Med. Imag., vol. 15, pp. 859–870, Dec. 1996.
Comput. Vis. Image Underst., vol. 66, no. 2, pp. 163–178, May 1997. [173] E. Ekoulé, F. Peyrin, and C. Odet, “Description d’une procèdure de tri-
[148] A. A. Young, “Model tags: Direct 3D tracking of heart wall motion from angulation entièrament automatique,” in Proc. COGNITIVA 87, Paris,
tagged MR images,” Med. Image Anal., vol. 3, no. 4, pp. 361–372, Dec. France, pp. 88–95.
1999. [174] F. H. Sheehan, E. L. Bolson, R. W. Martin, G. Bashein, and J. McDonald,
[149] J. Huang, D. Abendschein, V. Davila, and A. A. Amini, “Four-dimen- “Quantitative three dimensional echocardiography: Methodology, vali-
sional LV tissue tracking from tagged MRI with a 4D B-spline model,” dation and clinical applications,” in Lecture Notes in Computer Science,
IEEE Trans. Med. Imag., vol. 18, pp. 957–972, Oct. 1999. vol. 1496, Medical Imaging Computing & Computer-Assisted Interven-
[150] X. Papademetris, P. Shi, D. P. Dione, A. J. Sinusas, R. T. Constable, and tion, pp. 102–109.
J. S. Duncan, “Recovery of soft tissue object deformation from 3D image [175] H. Hoppe, “Surface Reconstruction from Unorganized Points,” Ph.D.
sequences using biomechanical models,” in Lecture Notes in Computer dissertation, Univ. Washington, Seattle, WA, 1994.
Science, A. Kuba, M. Sámal, and A. Todd-Pokropek, Eds. Berlin, Ger- [176] V. Caselles, F. Catté, T. Coll, and F. Dibos, “A geometric model for active
many: Springer-Verlag, 1999, vol. 1613, Information Processing and contours in image processing,” Num. Mathematik., vol. 66, pp. 1–31,
Medical Imaging, pp. 352–357. 1993.
[151] , “Recovery of Soft Tissue Object Deformation from 3D Image Se- [177] R. Malladi, J. Sethian, and B. Vemuri, “Shape modeling with front prop-
quences Using Biomechanical Models,” Image Processing and Analysis agation: A level set approach,” IEEE Trans. Pattern Anal. Machine In-
Lab., Dept. Diagnostic Radiology, Yale Univ., New Haven, CT, Tech. tell., vol. 17, pp. 158–174, Feb. 1995.
Rep.1999-01, June 1999. [178] S. Osher and S. Sethian, “Fronts propagating with curvature dependent
[152] X. Papademetris, A. J. Sinusas, D. P. Dione, and J. S. Duncan, “3D car- speed: Algorithms based on the Hamilton-Jacobi formalism,” J. Com-
diac deformation from ultrasound images,” in Lecture Notes in Com- putat. Phys., vol. 79, pp. 12–49, 1988.
puter Science, vol. 1679, Medical Imaging Computing & Computer-As- [179] Y.-H. Tseng, J.-N. Hwang, and F. H. Sheehan, “3-D heart modeling and
sisted Intervention, pp. 421–429. motion estimation based on continuous distance transform neural net-
[153] A. Pentland and S. Sclaroff, “Closed-form solutions for physically based works and affine transform,” J. VLSI Signal Processing Syst. Signal,
shape modeling and recognition,” IEEE Trans. Pattern Anal. Machine Imag., Video Technol., vol. 18, no. 3, pp. 207–218, Apr. 1998.
Intell., vol. 13, pp. 703–714, July 1991. [180] A. J. Hanson, “Hyperquadrics: Smoothly deformable shapes with
[154] D. Metaxas and D. Terzopoulos, “Shape and nonrigid motion estimation convex polyhedral bounds,” Comput. Vis. Graph. Imag. Processing,
through physics-based synthesis,” IEEE Trans. Pattern Anal. Machine vol. 44, no. 2, pp. 191–210, Nov. 1988.
Intell., vol. 15, pp. 580–591, June 1993. [181] A. A. Requicha and H. B. Voelcker, “Solid modeling: A historical sum-
[155] A. H. Barr, “Superquadrics and angle-preserving deformations,” IEEE mary and contemporary assessment,” IEEE Comput. Graph. Applicat.,
Comput. Graph. Applicat., vol. 1, no. 1, pp. 11–23, Jan. 1981. vol. 2, no. 2, pp. 9–24, Mar. 1982.
FRANGI et al.: THREE-DIMENSIONAL MODELING FOR FUNCTIONAL ANALYSIS OF CARDIAC IMAGES: A REVIEW 25

[182] B. P. F. Lelieveldt, “Anatomical Models in Cardiovascular Image Anal- [204] R. A. Greenbaum and D. G. Gibson, “Regional nonuniformity of left
ysis,” Ph.D. dissertation, Leiden Univ. Med. Ctr., Leiden, The Nether- ventricular wall movement,” Br. Heart J., vol. 45, no. 1, pp. 29–34, Mar.
lands, 1999. 1981.
[183] L. Piegl and W. Tiller, The NURBS Book, 2nd ed, Berlin, Germany: [205] E. Shapiro, D. L. Marier, M. G. St. J. Sutton, and D. G. Gibson, “Re-
Springer-Verlag, 1996. Monographs in Visual Communication. gional nonuniformity of wall dynamics in normal left ventricle,” Br.
[184] K. A. Bartels, A. C. Bovik, and C. E. Griffin, “Spatio-temporal tracking Heart J., vol. 45, no. 3, pp. 264–270, Mar. 1981.
of material shape change via multidimensional splines,” in Math. Meth. [206] N. G. Pandian, D. J. Skorton, S. M. Collins, H. L. Falsetti, E. R. Burke,
Biomed. Imag. Anal., Seattle, WA, 1994, pp. 110–116. and R. E Kerber, “Heterogeneity of left ventricular segmental wall
[185] J. M. Guccione and A. D. McCulloch, “Finite element modeling of ven- thickening and excursion in 2-dimensional echocardiograms of normal
tricular mechanics,” in Theory of Heart, P. J. Hunter, A. D. McCul- human subjects,” Amer. J. Cardiol., vol. 51, no. 10, pp. 1667–1673,
loch, and P. Nielsen, Eds, Berlin, Germany: Springer-Verlag, 1991, pp. June 1983.
122–144. [207] P. J. Besl and R. C. Jain, “Invariant surface characteristics for 3-D object
[186] A. A. Amini, Y. Chen, and D. Abendschein, “Comparison of land-mark- recognition in range images,” Comput. Vis. Graph. Image Processing,
based and curved-based thin-plate warps for analysis of left-ventricular vol. 33, no. 1, pp. 33–80, Jan. 1986.
motion from tagged MRI,” in Lecture Notes in Computer Science, vol. [208] P. Clarysse, D. Friboulet, and I. E. Magnin, “Tracking geometrical de-
1679, Medical Imaging Computing & Computer-Assisted Intervention, scriptors on 3-D deformable surfaces: Application to the left-ventricular
Berlin, Germany, pp. 498–507. surface of the heart,” IEEE Trans. Med. Imag., vol. 16, pp. 392–404,
[187] E. Waks, J. L. Prince, and A. Douglas, “Cardiac motion simulator 1997.
for tagged MRI,” in Math. Meth . Biomed. Imag . Anal., 1996, pp. [209] J. J. Koenderink and A. J. van Doorn, “Surface shape and curvature
182–191. scales,” Imag. Vis. Comput., vol. 10, no. 8, pp. 557–565, Oct. 1992.
[188] D. L. Kraitchman, A. A. Young, C.-N. Chang, and L. Axel, “Semiau- [210] C. Dorai and A. K. Jain, “Shape spectrum based view grouping and
tomatic tracking of myocardial motion in MR tagged images,” IEEE matching of 3D free-form objects,” IEEE Trans. Pattern Anal. Machine
Trans. Med. Imag., vol. 14, pp. 422–433, Sept. 1995. Intell., vol. 19, pp. 1139–1146, Oct. 1997.
[189] A. S. Douglas, W. C. Hunter, and M. D. Wiseman, “Inhomogeneous [211] M. P. Do Carmo, Differential Geometry of Curves and Sur-
deformation as a source of error in strain measurements derived from faces. Englewood Cliffs, NJ: Prentice-Hall, 1976.
implanted markers in the canine lef ventricle,” J. Biomech., vol. 23, no. [212] H. Azhari, E. Grenadier, U. Dinnar, R. Beyar, D. R. Adam, M. L.
4, pp. 331–341, 1990. Marcus, and S. Sideman, “Quantitative characterization and sorting of
[190] F. C. Schweppe, Uncertain Dynamic Systems. Englewood Cliffs, NJ: three-dimensional geometries: Application to left ventricles in vivo,”
Prentice-Hall, 1973. IEEE Trans. Med. Imag., vol. 36, pp. 322–332, Mar. 1989.
[191] T. S. Denney Jr. and E. R. McVeigh, “Model-free reconstruction of [213] H. Azhari, S. Sideman, R. Beyar, E. Grenadier, and U. Dinnar, “An
three-dimensional myocardial strain from planar tagged MR analytical descriptor of three-dimensional geometry: Application to
images,” J. Magn. Reson. Imag., vol. 7, no. 5, pp. 799–810, Sept. the analysis of the left ventricle shape and contraction,” IEEE Trans.
1997. Biomed. Eng., vol. 34, pp. 345–355, May 1987.
[192] D. F. Leotta, B. Munt, E. L. Bolson, C. Kraft, C. M. Martin, R. W. amd [214] H. Azhari, I. Gath, R. Beyar, M. L. Marcus, and S. Sideman, “Discrim-
Otto, and F. H. Sheehan, “Quantitative three-dimensional echocardiog- ination between healthy and diseased hearts by spectral decomposition
raphy by rapid imaging from multiple transthoracic windows: In vitro of their left ventricular 3D geometry,” IEEE Trans. Med. Imag., vol. 10,
validation and initial in vivo studies,” J. Amer. Soc. Echocardiogr., vol. pp. 207–215, June 1991.
10, no. 8, pp. 830–839, Oct. 1997. [215] M. A. Guttman, J. L. Prince, and E. R. McVeigh, “Tag and contour de-
[193] B. I. Munt, D. F. Leotta, E. L. Bolson, K. Coady, R. W. Martin, C. M. tection in tagged MR images of the left ventricle,” IEEE Trans. Med.
Otto, and F. H. Sheehan, “Left ventricular shape analysis from three- Imag., vol. 13, pp. 74–88, Jan. 1994.
dimensional echocardiograms,” J. Amer. Soc. Echocardiogr., vol. 11, no. [216] M. A. Guttman, E. A. Zerhouni, and E. R. McVeigh, “Analysis and visu-
8, pp. 761–769, Aug. 1998. alization of cardiac function from MR images,” IEEE Comput. Graph.
[194] S. Sasayama, H. Nonogi, M. Fujita, T. Sakurai, A. Wakabayashi, C. Applicat., vol. 17, no. 1, pp. 55–63, Jan. 1997.
Kawai, S. Eiho, and M. Kuwahara, “Three-dimensional analysis of [217] L. Axel, D. Bloomgarden, C. N. Chang, D. Kraitchman, and A. A.
regional myocardial function in response to nitroglycerin in patient Young, “SPAMMVU: A program for the analysis of dynamic tagged
with coronary artery disease,” J. Amer. Coll. Cardiol., vol. 3, no. 5, pp. MRI,” in Proc. ISMRM, 1993, p. 724.
1187–1196, May 1984. [218] A. Bazille, M. A. Guttman, E. R. McVeigh, and E. A. Zerhouni, “Im-
[195] G. Germano, J. Erel, H. Lewin, P. B. Kavanagh, and D. S. Berman, pact of semiautomated versus manual image segmentation errors on my-
“Automatic quantitation of regional myocardial wall motion and ocardial strain calculation by MR tagging,” Inv. Rad., vol. 29, no. 4, pp.
thickening from gated technetium-99n sestamibi myocardial perfu- 427–433, Apr. 1994.
sion single-photom emission computed tomography,” J. Amer. Coll. [219] S. Zhang, M. A. Douglas, L. Yaroslavsky, R. M. Summers, V. Dilsizian,
Cardiol., vol. 30, no. 5, pp. 1360–1367, Nov. 1997. L. Fananapazir, and S. L. Bacharach, “A Fourier based algorithm for
[196] A. A. Young, L. Axel, L. Dougherty, D. K. Bogen, and C. S. Parenteau, tracking SPAMMM tags in gated magnetic resonance cardiac images,”
“Validation of tagging with MR imaging to estimate material deforma- Med. Phys., vol. 23, no. 8, pp. 1359–1369, Aug. 1996.
tion,” Radiology, vol. 188, no. 1, pp. 101–108, July 1993. [220] W. S. Kerwin and J. L. Prince, “The Kriging update model and recursive
[197] W. S. Kerwin and J. L. Prince, “Tracking MR tag surfaces using space-time function estimation,” IEEE Trans. Signal Processing, vol.
a spatiotemporal filter and interpolator,” Int. J. Image Syst. 47, pp. 2942–2952, Nov. 1999.
Technol., vol. 10, pp. 128–142, 1999. [221] Handbook of Applied Advanced Geostatistical Ore Reserve Estimation,
[198] T. Arts, W. C. Hunter, A. Douglas, M. M. Muijtjens, and R. S. Ren- Elsevier Sci., Amsterdam, The Netherlands, 1988.
eman, “Description of the deformation of the left ventricle by a kine- [222] N. F. Osman, W. S. Kerwin, E. R. McVeigh, and J. L. Prince, “Cardiac
matic model,” J. Biomech., vol. 25, no. 10, pp. 1119–1127, Oct. 1992. motion tracking using CINE harmonic phase (HARP) magnetic reso-
[199] J. Declerk, T. S Denney, C. Öztürk, W. O’Dell, and E. R. McVeigh, nance imaging,” Magn. Res. Med., vol. 42, no. 6, pp. 1048–1060, Dec.
“Left ventricular motion reconstruction from planar tagged MR images: 1999.
A comparison,” Phys. Med. Biol., vol. 45, no. 6, pp. 1611–1632, June [223] J. Garot, D. A. Bluemke, N. F. Osman, C. E. Rochitte, E. R. McVeigh,
2000. E. A. Zerhouni, and J. L. Prince, “Fast determination of regional my
[200] J. Declerck, N. Ayache, and E. R. McVeigh, “Use of a 4D Planispheric ocardial strain fields from tagged cardiac images using harmonic phase
Transformation for the Tracking and the Analysis of LV Motion with MRI,” Circulation, vol. 101, no. 9, pp. 981–988, Mar. 2000.
Tagged MR Images,” INRIA, Sophia-Antipolis, France, Tech. Rep. [224] N. F. Osman and J. L. Prince, “Motion estimation from tagged MR
3535, Oct. 1998. images using angle images,” Int. Conf. Image Processing, vol. 1, pp.
[201] C. Öztürk and E. R. McVeigh, “Four-dimensional B-spline based motion 704–708, 1998.
analysis of tagged cardiac MR images,” in Proc. SPIE, vol. 3660, 1999, [225] , “Visualizing myocardial function using HARP MRI,” Phys. Med.
pp. 46–56. Biol., vol. 45, no. 6, pp. 1665–1682, June 2000.
[202] J. Lötjönen, P.-J. Reissman, I. E. Magnin, and T. Katila, “Model extrac- [226] C. Taylor and A. Colchester, Eds., “Medical Imaging Computing
tion from magnetic resonance volume data using a deformable pyramid,” & Computer-Assisted Intervention,” in Lecture Notes in Computer
Med. Image Anal., vol. 3, no. 4, pp. 387–406, Dec. 1999. Science, Cambridge, U.K.: Springer-Verlag, Oct. 1999, vol. 1679.
[203] T. F. Cootes, A. Hill, C. J. Taylor, and J. Haslam, “Use of active shape [227] A. Colchester, W. M. Wells, and S. Delp, Eds., “Medical Imaging
models for locating structures in medical images,” Imag. Vis. Comput., Computing & Computer-Assisted Intervention,” in Lect. Notes Comput.
vol. 12, no. 6, pp. 355–365, July 1994. Sci.. Boston, MA: Springer-Verlag, Sept. 1998, vol. 1496.