Phương pháp chẩn đoán hình ảnh (Phần 3)

2089_book.fm Page 51 Tuesday, May 10, 2005 3:38 PM

2

Medical-Image
Processing and Analysis
for CAD Systems
Athanassios N. Papadopoulos, Marina E. Plissiti,
and Dimitrios I. Fotiadis

CONTENTS
2.1
2.2

Introduction
Basics of a CAD System
2.2.1 Computer-Aided Methodologies in Mammography
2.2.2 Historical Overview
2.2.3 CAD Architecture
2.2.4 Preprocessing
2.2.5 Segmentation
2.2.6 Feature Analysis (Extraction, Selection, and Validation)
2.2.7 Classification System (Reduction of False Positives or
Characterization of Lesions)
2.2.7.1 Conventional Classifiers
2.2.7.2 Artificial Neural Networks (ANNs)
2.2.7.3 Fuzzy-Logic Systems
2.2.7.4 Support-Vector Machines
2.2.8 Evaluation Methodologies
2.2.9 Integrated CAD Systems
2.3 Computer-Aided Methodologies for Three-Dimensional Reconstruction

of an Artery
2.3.1 IVUS Image Interpretation
2.3.2 Automated Methods for IVUS ROI Detection
2.3.2.1 IVUS Image Preprocessing
2.3.2.2 IVUS Image Segmentation
2.3.3 Limitations in Quantitative IVUS Image Analysis
2.3.4 Plaque Characterization in IVUS Images
2.3.5 Three-Dimensional Reconstruction
2.4 Conclusions
References

Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 52 Tuesday, May 10, 2005 3:38 PM

52

Medical Image Analysis

2.1 INTRODUCTION
Over the last 15 years, several research groups have focused on the development of
computerized systems that can analyze different types of medical images and extract
useful information for the medical professional. Most of the proposed methods use
images acquired during a diagnostic procedure. Such images are acquired using a
variety of techniques and devices, including conventional radiography, computerized
tomography, magnetic resonance imaging, ultrasound, and nuclear medicine. Computerized schemes have been widely used in the analysis of one-dimensional medical
signals such as Electrocardiogram (ECG), Electromyogram (EMG), Electroencephalogram (EEG), etc. However, the majority of medical signals are two-dimensional
representations. Computerized systems designed for the automated detection and
characterization of abnormalities in these images can provide medical experts with

useful information. Such systems are commonly referred to as computer-aided
detection/diagnosis systems (CAD).
A computer-aided detection procedure does not provide a medical diagnosis.
Rather, the computerized system is developed to detect signs of pathology in medical
images by extracting features that are highly correlated with the type and the
characteristics of the abnormality or the disease under investigation. If a specific
area in a radiological image meets the requirements, the computerized scheme
identifies it, and the radiologist can review it to improve the accuracy of the detection
procedure. On the other hand, computer-aided diagnosis schemes, based on the same
or additional features, characterize the identified region according to its pathology.
A CAD system is defined as a combination of image-processing techniques and
intelligent methods that can be used to enhance the medical interpretation process,
resulting in the development of more efficient diagnosis. The computer outcome
assists radiologists in image analysis and diagnostic decision making. In addition,
a CAD system could direct the radiologist’s attention to regions where the probability
of an indication of disease is greatest. A CAD system provides reproducible and
quite realistic outcomes.
In this chapter, we review two of the most common procedures in CAD systems.
The first is related to microcalcification detection and classification in mammograms.
In this procedure, features of microcalcifications are extracted, and intelligent methods are then used to classify these features. The second procedure is based on the
fusion of intravascular ultrasound and biplane angiographies aiming at the threedimensional (3-D) reconstruction of an artery.

2.2 BASICS OF A CAD SYSTEM
Most of the automated CAD approaches include feature-extraction procedures. However, several studies of semi-automated approaches have been reported wherein
radiologists manually perform feature-mining procedures by employing various
feature-extraction modules [1, 2]. CAD systems can be classified in two categories
according to their objectives: (a) those that are used to detect regions of pathology
and (b) those that are used to classify the findings based on their features, which
indicate their histological nature.
Copyright 2005 by Taylor & Francis Group, LLC

2089_book.fm Page 53 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

53

The role of these computerized systems is to improve the sensitivity of the
diagnostic process and not to make decisions about the health status of the patient.
However, the “D” in CAD should stand for “diagnosis” [3], although several reports
in literature utilize the word “detection” [4], which is undoubtedly an essential part
of the diagnostic procedure.
For the design and development of an automated CAD system, several issues
must be considered, including the quality of the digitized images, the sequence of
the processing steps, and the evaluation methodology. Most of the studies use filmscreen images that are digitized using high-performance film digitizers. Recent
studies employ high-quality medical images obtained directly in digital format using
advanced imaging systems (filmless technology). The specific characteristics of the
film digitizer significantly influence the quality of the image. In the case of filmscreen technology, the maximum optical density of the film is a critical parameter
in the quality of the final digitized image. In cases where the upper limit of the
optical density is low, an estimation of noise is possible during the digitization
procedure, especially on the background area (air) of the image. Utilization of filmscreen systems with higher optical densities might lead to the reduction of such
noise due to digitization.

2.2.1 COMPUTER-AIDED METHODOLOGIES

IN

MAMMOGRAPHY

Mammography is one of the radiological fields where CAD systems have been
widely applied because the demand for accurate and efficient diagnosis is so high.
The presence of abnormalities of specific appearance could indicate cancerous circumstances, and their early detection improves the prognosis of the disease, thus
contributing to mortality reduction [5]. However, diagnostic process is complicated
by the superimposed anatomical structures, the multiple tissue background, the low
signal-to-noise ratio, and variations in the patterns of pathology. Thus, the analysis
of medical images is a complicated procedure, and it is not unusual for indications
of pathology, such as small or low-contrast microcalcifications, to be missed or
misinterpreted by radiologists. On the other hand, clinical applications require realtime processing and accuracy in diagnosis. Based on these high standards in diagnostic interpretation, numerous intelligent systems have been developed to provide
reliable automated CAD systems that can be very helpful, providing a valuable
''second opinion'' to the radiologist [6, 7].

2.2.2 HISTORICAL OVERVIEW
Computerized analysis of radiological images first appeared in the early 1960s [8,
9]. One of the first studies employing computers in the area of mammography was
published by Winsberg et al. in 1967 [10]. In this approach, the right- and left-breast
shapes were compared to detect symmetry differences. Computation of local image
characteristics from corresponding locations with high variations indicated the presence of a disease. Ackerman et al. [11] defined four computer-extracted features for
the categorization of mammographic lesions as benign or malignant. Another study
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 54 Tuesday, May 10, 2005 3:38 PM

54

Medical Image Analysis

by the same research group [12] proposed a computational procedure for the processing of a feature set with 30 characteristics that are obtained by radiologists for
the classification of lesions according to their malignancy. At the same time, several

other works targeting detection and characterization of microcalcification clusters
appeared in the literature. Wee et al. [13] classified microcalcification clusters as
benign or malignant using the approximate horizontal length, the average internal
gray level, and the contrast of individual microcalcifications. The cluster pattern
together with features such as size, density, and morphological characteristics of the
cluster were also used for microcalcification characterization [14]. In the late 1970s,
Spiesberger [15] was the first to propose an automated system for the detection of
microcalcifications.
At the end of the 1980s, the literature was enriched by studies reporting several
image-processing algorithms and computational processes that provided satisfactory
descriptions and efficient procedures for the detection of microcalcifications [16–18].
In 1990, Chan et al. reported that under controlled circumstances, a CAD system
can significantly improve radiologists' accuracy in detecting clustered microcalcifications [19].

2.2.3 CAD ARCHITECTURE
CAD systems proposed in the literature are based on techniques from the field of
computer vision, image processing, and artificial intelligence. The main stages of a
typical CAD scheme are: preprocessing, segmentation, feature analysis (extraction,
selection, and validation), and classification utilized either to reduce false positives
(FPs) or to characterize abnormalities (Figure 2.1). A description of the methods
employed in each stage is given in the following sections.

2.2.4 PREPROCESSING
In this stage, the subtle features of interest are enhanced and the unwanted characteristics of the image are de-emphasized. The enhancement procedure results in a
better description of the objects of interest, thus improving the sensitivity of the
detection system and leading to better characterization in the case of diagnosis. The
enhancement of the contrast of the regions of interest, the sharpening of the abnormalities’ boundaries, and the suppression of noise is performed in this stage. Several
methodologies have been reported in the literature based on conventional imageprocessing techniques, region-based algorithms, and enhancement through the transformation of original image into another feature space. Global processing can be
performed, or local adjusting enhancement parameters can be used to accommodate
the particularity of different image areas.

Morphological, edge-detection, and band-pass filters have been utilized. An
enhanced representation can be obtained using subtraction procedures on the processed image [18]. One of the earliest contrast-enhancement methodologies was the
modification of image histogram [20] and its equalization [21]. The resulting image
contains equally distributed brightness levels over the gray-level scale. Because the
mammogram contains areas of different intensity, a global modification is poor.
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 55 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

55

Digital mammogram

Preprocessing

Computer-aided
detection scheme

Segmentation

Feature extraction - selection

Computer-aided diagnosis

Classification module - reduction
of FP findings

Feature extraction - selection

Computer-aided
characterization scheme

Classification module - likelihood
of malignancy

FIGURE 2.1 CAD architecture.

Performance can be improved utilizing local adjustments of the processing parameters (adaptive histogram equalization) [22]. Another technique restricts the methodology to certain contrast values to increase the effective range of contrast in the
specific areas (contrast-limited adaptive histogram equalization) [23].
Unsharp masking is a routinely used procedure to enhance the fine-detail structures. A high-spatial-frequency component multiplied by a weight factor is added
on the original image. In the case of linear unsharp filtering, the above parameters
are constant throughout the entire image. In nonlinear methodologies, the weighting
factor depends on the intensity of the examined region (background/foreground), or it
can be applied differently in different resolution levels in multiscale approaches [24].
Contrast stretch is a rescaling of image gray levels based on linear or nonlinear
transformations. In linear transformations, the difference between the background
and foreground areas is increased to improve the contrast of both areas. Introducing
a nonlinear transformation, the contrast of the different parts of the image is modified,
selectively enhancing the desired gray levels. In most medical images, objects of
interest have nonstandard intensities, thus the selection of a proper “intensity window” is not sufficient for contrast enhancement.
The adaptive neighborhood contrast-enhancement method improves the contrast
of objects or structures by modifying the gray levels of the neighborhood (contextual
region) of each pixel from which the object is composed. After the identification of
homogeneous areas (using, for example, a growing technique) several conditions
are imposed to downgrade unconventional high-contrast areas or low-level noise and
to enhance regions surrounded by variable background [25]. Techniques that enhance
regions of interest by estimating their difference from their background areas are

Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 56 Tuesday, May 10, 2005 3:38 PM

56

Medical Image Analysis

called region-based enhancement techniques. Typical region growing techniques,
which employ contrast and statistical conditions, result in the definition of the extent
and the shape of the objects [26].
Multiresolution methods, based mainly on wavelet analysis, are used to enhance
the features of mammographic images [27]. A multiscale analysis of the original
mammogram to several subband images provides the advantage of studying each
subband independently using scale characteristics. Each subband provides information based on different scales resulting in the representation of high- or low-frequency elements on separate images. Thus, noise or similar type components of the
image can be described in high resolution (small scale), while subtle objects with
defined extent or large masses are described in medium-resolution and low-resolution
levels (medium and coarse scales), respectively. Hence, the significant image features
can be selectively enhanced or degraded in different resolution levels [28]. Furthermore, adaptive approaches in wavelet enhancement techniques that ensure the avoidance of the utilization of global parameters have been reported [29].
Fuzzy-logic techniques are also used for contrast enhancement of microcalcifications [30]. Global information (brightness) is employed to transform an image to
a fuzzified version using a function, while local information (geometrical statistics)
is employed to compute the nonuniformity.
Methods that are based on deterministic fractal geometry have been used to
enhance mammograms [31–33]. A fractal-image model was developed to describe
mammographic parenchymal and ductal patterns using a set of parameters of affine
transformations. Microcalcification areas were enhanced by taking the difference
between the original image and the modeled image.

2.2.5 SEGMENTATION

In this stage, the original mammographic image is segregated into separate parts,
each of which has similar properties. The image background, the tissue area, and
the muscle or other areas can be separated because they are characterized using
generic features. Moreover, apart from the generic classification of image regions,
a CAD segmentation procedure can identify regions containing small bright spots
that appeared in groups and that correspond to probable microcalcifications and their
clusters. The complexity of a segmentation procedure depends on the nature of the
original image and the characteristics of the objects that have to be identified. A
mammographic image contains several regions having different attenuation coefficients and optical densities, resulting in intensity variations. In addition, because a
mammogram is a two-dimensional (2-D) representation of a 3-D object, the overlying areas develop a complex mosaic composed of bright regions that may or may
not be a real object. Thus, the implementation of a global single threshold or a set
of fixed thresholds that defines intensity ranges is not an efficient segmentation
procedure. Moreover, the employment of a global intensity threshold usually
increases the number or the size of the selected regions introducing noise, which
makes the procedure inefficient because noise removal requires further treatment.
In any case, after the first partitioning has been achieved, region-growing techniques,
following specific homogeneity and differentiation criteria, can be utilized to define
the real extent and the exact borders of the segmented region.
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 57 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

57

To overcome the limitations of a global thresholding methodology, local thresholding criteria must be utilized from the beginning. The definition of the parameters
that satisfy the demands of the segmentation algorithm increase the efficiency of the
technique. The corresponding measures were calculated for a specific window size.

Some of the local thresholding criteria are:
The mean intensity values plus/minus a number of standard deviation (SD)
values of intensity [16]
The difference of the intensity value of a seed pixel from the maximum and
minimum intensity values of pixels that belong to a specific neighborhood
around a seed pixel [34]
A contrast measure equal to the difference of intensity between object and
background region [35]
An object is selected only if the feature value belongs to the highest 2% of the values
obtained.
In a similar but more flexible way, adaptive filtering methodologies have been
proposed, defining parameters or measures adjusted to a specific area. A feature
called prediction error (PE) is the difference between the actual pixel value and the
weighted sum of the eight nearest-neighbor pixels [36]. If PE follows a Gaussian
distribution, calcifications are not present. Functions using first, second, and third
moments of the PE are used to generate a threshold value that reveals the presence of
the microcalcifications. In another study [37], given a local maximum pixel value x0,y0,
an edge pixel is given by the value of x,y that maximizes the difference in pixel values
between pixels at x,y and x0,y0, divided by the distance between the two pixels.
Mathematical morphology filtration has been used to segment the microcalcifications. Classical erosion and dilation transformations, as well as their combinations
such as open, close, and top-hat transformations, are employed [38].
In statistical approaches, several histogram-based analysis and Markov random
field models are used [39, 40]. Markov random fields have been used to classify
pixels to background, calcification, line/edge, and film-emulsion errors [41]. Multiscale analysis based on several wavelet transformations has been used to enable the
segmentation process to be performed using the different scales-levels [42, 43].
Furthermore, as in the preprocessing module, techniques have been applied exploiting fractal [44] and fuzzy-logic methodologies [45].

2.2.6 FEATURE ANALYSIS (EXTRACTION, SELECTION,

AND

VALIDATION)

In this stage, several features from the probable microcalcification candidates are
extracted to reduce false positives. In any segmentation approach, a considerable
number of normal objects are recognized as pathological, which results in reduced
efficiency of the detection system. To improve the performance of the scheme, several
image features are calculated in an effort to describe the specific properties or
characteristics of each object. The most descriptive of these features are processed
by a classification system to make an initial characterization of the segmented
samples.
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 58 Tuesday, May 10, 2005 3:38 PM

58

Medical Image Analysis

TABLE 2.1
Features for the Detection and Characterization
of Microcalcifications and Their Clusters
Microcalcification (MC) Cluster
Classification Features

Radiologists’
Characterization Features

Number of MCs in cluster

Cluster area
Mean MC area
SD of MC area
Mean MC compactness
Mean MC elongation
SD of MC elongation
SD of MC intensity
Mean MC background intensity
Mean contrast
Cluster eccentricity
Mean distance from cluster centroid
Neighboring with a larger cluster
Cluster entropy
Spreading of MCs in cluster
Cluster elongation
Mean local MC background
Mean MC intensity
SD of MC compactness
SD of distances from cluster centroid
Area of the cluster convex hull
Length of the cluster convex hull

Cluster elements (separable/countable)
Cluster size
MC size
Shape of elements within cluster
Shape of elements within cluster
Shape of elements within cluster
Shape of elements within cluster
Density of calcifications

Density of calcifications
Contrast of calcifications
Shape of cluster
Calcification distribution
Cluster distribution
Calcification distribution
Calcification distribution
Cluster shape
Density of calcifications
Density of calcifications
Shape of elements within cluster
Calcification distribution
Shape of cluster
Shape of cluster

Although the number of calculated features derived from different feature spaces
is quite large, it is difficult to identify the specific discriminative power of each one.
Thus, a primary problem is the selection of an effective feature set that has high
ability to provide a satisfactory description of the segmented regions. Early studies
utilized features that were similar to the features that radiologists employ during
their diagnosis. However, as mentioned previously, additional features not employed
by the doctors also have high discrimination power. Table 2.1 provides a list of
typical morphological features of individual microcalcification and their clusters.
Specific features could be extracted, such as the surround region dependence
matrix (SRDM), gray-level run length (GLRL), and gray-level difference (GLD)
[46]. Laplacian or Gaussian filtration can be used in the validation of features [47].
Using wavelet analysis, features such as energy, entropy, and norms of differences
among local orientations can be extracted [48].
The use of a large number of features does not improve the classification
performance. Indeed, the use of features without discriminative power increases the

complexity of the characterization process. In addition, the probability of misclassification increases with the number of features. Moreover, the prediction variability
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 59 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

59

is larger, and the classifier is sensitive to outliers. Finally, the more features included
in a given classifier, the greater is the dimension of a training set needed for the
same degree of reliability [49]. The selection of the optimal feature subset is a
laborious problem. Only an exhaustive search over all subsets of features can provide
the system with a reliable subset. Usually, the criterion of selecting an efficient
subset of features is the minimization of misclassification probability (classification
error). However, for the testing of a subset, a classifier must be chosen, and it is
important to consider that different classifiers and different methods for the estimation of error rate could lead to the selection of a different feature subset.
One of the most important issues of a mammographic CAD system is the
selection of a standard feature set and the classification method that is used to extract
regions of pathological interest while minimizing false-positive findings. The selection of the appropriate features can be based on “weighting factors” proposed by
radiologists [50–53] or on algorithmic procedures that identify the most discriminant
features.
The feature space can be a transformed space that has lower dimension than the
original, although its discriminating power could be higher. To achieve this, PCA
(principal component analysis), which is based on the elimination of features that
contribute less, can be used [54, 55].
Alternatively, the most discriminative features can be selected, reducing in this
way the size of the feature set. Several methods have been proposed, such as:
Stepwise discriminant analysis [56]

Sequential Forward Selection (SFS) and Sequential Backward Selection
(SBS) [57]
Genetic algorithms [58]
Stepwise discriminant analysis is based on the sequential trial of different feature
subsets. The one that results in the smallest error rate is chosen as the most convenient
[59–61]. Sequential forward selection is a bottom-up search procedure where one
feature at a time is added to the feature set. At each stage, the feature to be included
in the feature set is selected from among the remaining features [57, 62, 63]. Genetic
algorithms have been used to select features that could enhance the performance of
a classifier (for distinguishing malignant and benign masses). In the same way,
genetic algorithms have been used to optimize the feature set for the characterization
of microcalcifications [64, 65].

2.2.7 CLASSIFICATION SYSTEM (REDUCTION OF FALSE POSITIVES OR
CHARACTERIZATION OF LESIONS)
Diagnosis is an integrated medical procedure that is defined as the art or act of
recognizing the presence of a disease from its signs or symptoms. During the entire
process, especially in the case of differential diagnosis, it is obvious that there are
several dilemmas for the rejection or acceptance of probable diseases. Thus, a
classification system is an essential part of a CAD system. Classification schemes
range from techniques that classify lesions according to their different types (stellate,
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 60 Tuesday, May 10, 2005 3:38 PM

60

Medical Image Analysis

circumscribed masses, or calcifications) [66] to techniques that produce binary
diagnosis, characterizing the findings as malignant or benign.
The classifiers that are utilized in the area of the detection of mammographic
microcalcification are those employed in most of the medical image-analysis procedures. They could be categorized in the following classes:
Conventional classifiers
Artificial neural networks
Fuzzy-logic systems
Support-vector machines
2.2.7.1 Conventional Classifiers
2.2.7.1.1 Rule-Based Systems (Decision Trees)
The decision tree is one of the most widely used techniques for the extraction of
inductive inference. As a learning method, it aims at the definition of an approximating discrete-value target function in which the acquired knowledge is represented
as a decision tree. The architecture of the classifier includes a set of “if-then” rules.
A decision-tree scheme includes a main root node, from where the classification
procedure starts, and several leaf nodes where the classification of the instance is
given. Each node in the tree specifies a check of an attribute of the instance, and
each branch descending from that node corresponds to one of the possible values
for this specific attribute. An instance is categorized beginning from the root node
and, by checking the attribute specified by this node, moving down to the specific
tree branch that is responsible for the value of this attribute. A similar procedure is
replicated if a new tree is rooted at the new node.
From the early studies of microcalcification detection and characterization in
mammography, rule-based systems provide a remarkable assistance in the simulation
of the diagnosis process carried out by a radiologist [67, 68]. Although, the conversion of medical rules to “if-then” rules is a feasible task, the development of a highperformance system has not been achieved. This is due to the absence of attributevalue pair representations in medical data and the lack of disjunctive descriptions
or large data sets for system training that include all the specific disease cases.
2.2.7.1.2 Bayesian Quadratic and Linear Classifiers (Statistical)
A Bayesian classifier is based on the approximation of the class-conditional probabilistic density functions (PDFs). Each PDF expresses the frequency of occurrence
of each sample in the feature space. Typically, an unknown sample is classified to
a class with the highest value of its PDF. The problem is that the precise approximation of the PDFs has to be defined [62].
Quadratic and linear classifiers are statistical (parametric) methods that utilize

Gaussian distributions for the PDFs. The mean vector and the covariance matrix are
estimated from the training set of each class. In the case of a Bayesian quadratic
classifier (BQC), the classification boundary forms a quadratic curve. In the case of
a Bayesian linear quadratic (BLQ) classifier, instead of using different covariance
matrices for the individual classes, one unified covariance matrix is used for all
classes, and the classification border is a straight line.
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 61 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

61

2.2.7.1.3 Nonparametric
When the underlying distributions of the samples are quite complex, additional
techniques can be employed to approximate the PDFs. The K-nearest neighbor and
the Parzen estimate belong to this category. In the K-nearest-neighbor technique,
the classification boundary is directly constructed instead of calculating the PDFs
[69]. For an unknown sample, distances to the individual training samples are
calculated, and the major class in the nearest K samples is selected. The Parzen
estimate method is used if the distribution is complex, and its generation is quite
difficult. Numerous kernel functions that describe the individual training samples
are summed up to calculate the complex PDF [70].
2.2.7.2 Artificial Neural Networks (ANNs)
A neural network is a structure that can be adjusted to produce a mapping of
relationships among the data from a given set of features. For a given set of data
{xi , yi }iN=1, the unknown function, y = f(x), is estimated utilizing numerical algorithms.
The main steps in using ANNs are: First, a neural-network structure is chosen in a

way that should be considered suitable for the type of the specific data and the
underlying process to be modeled. The neural network is trained using a training
algorithm and a sufficiently representative set of data (training data set). Finally, the
trained network is evaluated with different data (test data set), from the same or
related sources, to validate that the acquired mapping is of acceptable quality.
Several types of neural networks have been reported, such as feedforward [12,
20, 36, 43, 48, 55, 57], radial basis function [71], Hopfield [72], vector quantization,
and unsupervised types such as self-organizing maps [73]. A review of the role of
the neural networks in image analysis is reported by Egmont et al. [74]. Because
feedforward back-propagation and radial basis function neural networks are the most
common, a brief description of these network architectures can be meaningful.
Typically, a neural network is a structure involving weighted interconnections among
neurons (nodes), which are typically nonlinear scalar transformations. Figure 2.2
shows an example of a two-hidden-layer neural network with three inputs, x = {x1,
x2, x3}, that feed each of the five neurons composing the first hidden layer. The five
outputs from this layer feed each of the three neurons that compose the second
hidden layer, which, in a similar way, are fed into the single-output-layer neuron,
yielding the scalar output, yˆ . The layers of neurons are called hidden because their
outputs are not directly seen in the data. The inputs to the neural network are feature
vectors with dimensions equal to the amount of the most significant features. Several
training algorithms are implemented before selecting the one that is “most suitable”
for the network training. Gradient descent, resilient back-propagation, conjugate
gradient, quasi-Newton, and Levenberg-Marquardt are some of the most common
training methods [75].
2.2.7.3 Fuzzy-Logic Systems
Classification reduces the nonstatistical uncertainty. Statistical uncertainty can be
handled in several ways, so the nonstatistical uncertainty must be decreased to
Copyright 2005 by Taylor & Francis Group, LLC

2089_book.fm Page 62 Tuesday, May 10, 2005 3:38 PM

62

Medical Image Analysis

1st Hidden level

Input level

2nd Hidden level

Output level

FIGURE 2.2 A feedforward neural network with three inputs and two hidden layers composed of five and three neurons, respectively, and one output neuron.

develop more reliable classification approaches. Fuzzy set theory is an approach to
resolve this problem. Initially, fuzzy sets are integrated into rule-based expert systems to improve the performance of decision-support systems. Fuzzy procedures
can also be used to automatically generate and tune the membership functions on
the definition of different classes. Image-processing techniques have been reported
employing different feature sets defined in a fuzzy way. Intelligent methodologies
and pattern-recognition techniques have been used to introduce fuzzy clustering and
fuzzy neural-network approaches [76].
However, fuzzy sets can be utilized in more than one stage of a classifier design.
Fuzzy inputs can also be used, wherein the original input values are converted to a
more “blurry” version. For instance, instead of using the exact values of the feature
vector, a new vector consisting of feature values expressing the degree of membership of the specific value to the fuzzy sets (e.g., small, medium, large) can be used.
Fuzzy reasoning can be utilized in classification processes in which the inferences
are not strictly defined. The categories in a medical classification procedure are
exclusive. Thus, every sample belongs to a specific category. However, in some

cases, an unknown sample belongs to more than one class, but with a different degree
of membership. In such cases, the classification scheme is based on the utilization
of fuzzy classes.
2.2.7.4 Support-Vector Machines
Another category of classification methods that has recently received considerable
attention is the support-vector machine (SVM) [77–80]. SVMs are based on the
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 63 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

63

Nonlinear
Function

Data space

Feature space

FIGURE 2.3 A nonlinear SVM maps the data from the feature space D to the high-dimensional feature space F using a nonlinear function.

definition of an optimal hyperplane that separates the training data to achieve a
minimum expected risk. In contrast to other classification schemes, an SVM aims
to minimize the empirical risk Remp while maximizing the distances (geometric
margin) of the data points from the corresponding linear decision boundary (Figure
2.3). Remp is defined as

Remp ( a ) =

1
2l

l

∑ y − f ( x , a)
i

i

(2.1)

i =1

where
xi ∈ RN, i = 1, …, l, is the training vector belonging to one of two classes
l is the number of training points
yi ∈ {−1, 1} indicates the class of xi
ƒ is the decision function
The training points in the space RN are mapped nonlinearly into a higher dimensional space F by the function (a priori selected) : RN → F. It is in this space (feature
space) where the decision hyperplane is computed. The training algorithm uses only
the dot products ((xi)⋅(xj)) in F. If there exists a “kernel function” K such that K(xi,xj)
= (xi)⋅(xj), then only the knowledge of K is required by the training algorithm. The
decision function is defined as
l

f ( x) =

∑ y a K ( x , x) + b
i i

i

(2.2)

i =1

where ai represents the weighting factors and b denotes the bias. After training, the
condition ai > 0 is valid for only a few examples, while for the others ai = 0. Thus,
the final discriminant function depends only on a small subset of the training vectors,
which are called support vectors. Several types of kernels have been reported in the
literature, such as the polynomial type of degree p
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 64 Tuesday, May 10, 2005 3:38 PM

64

Medical Image Analysis

K ( xi , x ) = ( xi ⋅ x + 1) p

(2.3)

and the Gaussian kernel
K ( xi , x ) = e

2

− xi − x / 2 σ 2

(2.4)

where σ is the kernel width.

2.2.8 EVALUATION METHODOLOGIES
The evaluation of a classification system is one of the major issues in measuring
the system’s performance. From the early beginning, researchers have utilized several performance indexes to estimate the diagnostic system’s ability to distinguish
accurately the samples in their classes. True-positive (TP) rate and false-positive
(FP) rate are indexes that partially indicate the classification performance of a system.
The TP rate represents the percentage of “diseased” samples that are correctly
classified as “diseased,” and the FP rate represents the percentage of normal samples
that are incorrectly classified as “diseased.” However, in most of the statistical
classification systems, the adjustment of certain algorithmic parameters can modify
their operating points, resulting in the achievement of different pairs of TP and FP
rates. Such behavior introduces questions about the selection of the appropriate
training parameters of the system and results in difficulties in evaluating the system’s
actual performance for different degrees of confidence.
The receiver operating characteristic (ROC) methodology is the most widely
used scheme for evaluating the performance of a CAD system. ROC analysis overcomes the problem of a fixed selection of the classification parameters. A 2-D
graphical representation of all corresponding single points, expressing each pair of
TP and FP rates, gives the overall performance of the system. It is generated by
plotting the true-positive rate (sensitivity) against the false-positive rate (1-specificity) for various threshold values (Figure 2.4). The ROC curve represents the tradeoff between the TP/FP values and changes in the criterion for positivity [81]. The
area under curve (AUC, Az) is a measure of the diagnostic performance of the
classifier. The Az value defines the probability that the classifier will rank a randomly
chosen positive instance higher than a randomly chosen negative instance. It is
possible for a classifier with a lower Az value to have higher classification ability,

in a specific point, than another having higher Az value. Nevertheless, the Az value
is an efficient measure of the classification performance.
Alternative evaluation methodologies are the free ROC (FROC) [82] and the
location-specific ROC (LROC) [83]. In the FROC technique, the detection outcome
of a CAD system, for each image, contains normal or abnormal objects that are
characterized as TP or FP findings if they are in the area of real or fake detections,
respectively. The FROC curve is created by a plot of TP rate vs. the number of false
positive samples per image. In the case of the LROC methodology, only one object
is contained in each image or, in the case of a normal exam, none. The annotation
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 65 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

65

1
0.9
0.8
0.7

Sensitivity

0.6
0.5
0.4
0.3
AZ = 0.83

0.2

AZ = 0.91
0.1

AZ = 0.96

0
0

0.1

0.2

0.3

0.4

0.5
0.6
1-specificity

0.7

0.8

0.9

1

FIGURE 2.4 ROC curves indicating the performance of three different classification systems.

of the database is performed by radiologists, who localize the abnormalities on each
image. A simpler version of the FROC method is the alternative free-ROC (AFROC)
technique [84]. ROC methodologies impose limitations in their application to different medical diagnostic systems such as limited data sets, independence of samples,
the lack of categorical rating in the characterization, and the absence of indexes that
can characterize the detection difficulty of a specific sample [85, 86]. A unified ROC
methodology that can be used efficiently for all CAD systems does not exist.

2.2.9 INTEGRATED CAD SYSTEMS
The research tasks that have been proposed for more than 15 years in the area of
computer-aided detection in mammography have been integrated into efficient clinical devices that can provide useful information to radiologists. To date, three CAD
systems have been approved by the U.S. Food and Drug Administration as clinical
devices valuable in detection of pathological areas/objects in mammography. These
systems are the ImageChecker (R2 Technology) [87], the Second Look Digital/AD
(CADx Medical Systems) [88], and MammoReader (Intelligent Systems Software)
[89].
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 66 Tuesday, May 10, 2005 3:38 PM

66

Medical Image Analysis

However, other systems have also been developed and are being clinically
evaluated. Some of these systems are: Mammex TR (Scanis Inc.) [90], the Promam
(DBA Systems Inc.) [91], and the MedDetect (LMA & RBDC) [92]. The performances of the clinically approved systems have been evaluated by several research

groups or organizations [93–96].

2.3 COMPUTER-AIDED METHODOLOGIES FOR THREEDIMENSIONAL RECONSTRUCTION OF AN ARTERY
The modules of a CAD system for the detection and characterization of abnormalities
in mammography have been described in Section 2.2. Those systems take advantage
of the specific appearance of the breast tissue depicted utilizing X-rays. However,
similar image-analysis and artificial-intelligence techniques can be applied in medical images obtained by different imaging modalities. One such case is intravascular
ultrasound (IVUS) images, which are acquired using ultrasonic signals to depict the
inner structure of arteries. Detection of the actual borders of the lumen and plaque
in vessels is crucial in defining the severity of arterial disease. Diagnostic ultrasound
has become the most common imaging modality, and the number of clinical applications for ultrasound continues to grow.
Coronary artery disease is the most common type of heart disease and the leading
cause of death both in men and women in Europe and the U.S. The main cause of
coronary artery disease is atherosclerosis, which results in hardening and thickening
of the inner lining of arteries. Deposits of fatty substances, cholesterol, cellular waste
products, calcium, and other substances build up in the arterial wall, resulting in the
development of atheromatic plaque. As a consequence, partial or total obstruction
of blood flow in the artery can occur, which can lead to heart attack.
Early diagnosis and accurate assessment of plaque position and volume are
essential for the selection of the appropriate treatment. Biplane coronary angiography
has been used as the “gold standard” for the diagnosis of coronary narrowings and
guiding coronary interventions. On the other hand, intravascular ultrasound (IVUS)
is an interventional technique that produces tomographic images of the arterial
segments. These techniques are considered to be complementary because the first
provides information about the lumen width and the vessel topology, while the
second permits direct visualization of the arterial wall morphology.
Today, IVUS is used extensively as a routine clinical examination that assists
in selecting and evaluating therapeutic intervention such as angioplasty, atherectomy,
and stent placement. The aim of IVUS and angiographical image processing is the
extraction of valuable diagnostic information about the nature of alternations of

lining of arteries and the three-dimensional vessel morphology. Quantitative estimations of plaque thickness, volume, and position in the arterial wall are obtained from
the processing of the acquired images. Sophisticated modeling techniques combining
images from both modalities allow the three-dimensional (3-D) reconstruction of
the arterial segment and provide useful geometrical and positional information about
the shape of the lumen in 3-D space. The following sections describe several automated methods for quantitative analysis of IVUS images and techniques for the
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 67 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

67

adventitia

lumen

intima

media

lumen /intima border

catheter
distal
shadowing

Lipid lake
calcium

media /adventitia border

FIGURE 2.5 (a) Cross-sectional pattern appearance of IVUS images; (b) borders of interest
in IVUS image.

extraction of three-dimensional vessel models with fusion of IVUS and angiographical data.

2.3.1 IVUS IMAGE INTERPRETATION
IVUS is an invasive catheter-based imaging technique that provides 360° radial
images in a plane orthogonal to the long axis of the catheter. IVUS image sequences
consist of cross-sectional images of the arterial segment and are acquired with the
insertion of a catheter in the vessel. The reflection of the ultrasound beam as it passes
through the different layers and the scattering of the material give rise to a typical
image pattern that can then be used to identify different regions in IVUS images.
Figure 2.5 shows a schematic diagram of the cross-sectional anatomy of an artery
as well as an original depiction in IVUS images.
There are two key landmarks in IVUS images that assist in the correct interpretation of arterial structure: the lumen/intima border and the media/adventitia border.
Each one is recognized in IVUS images by its location and its characteristic appearance. As seen in Figure 2.5(b), the first bright interface beyond the catheter itself is
the lumen/intima border. Moreover, the media is usually a discrete thin layer that is
generally darker than intima and adventitia. The appearance of intima, media, and
adventitia follows a double-echo pattern showing a circumferentially oriented parallel bright-dark-bright echo pattern that is referred to as the “typical” three-layered
appearance. In IVUS images of normal arteries, the three-layered appearance may
not be visible because the intima may be too thin or there may be sufficient collagen
and elastin in the media of some arterial segments for it to blend with the surrounding
layers. In addition, in highly diseased vessels, the media may be very thin to register
as a separate layer on ultrasound images. It is more likely that the media is clearly
defined over only a part of the vessel circumference. In such cases or in noisy images,
the identification of the media/adventitia border is obtained by the transition in
Copyright 2005 by Taylor & Francis Group, LLC

2089_book.fm Page 68 Tuesday, May 10, 2005 3:38 PM

68

Medical Image Analysis

“texture” of regions corresponding to plaque and adventitia. In sequential IVUS
frames, plaque can be distinguished from blood flowing in the lumen, because plaque
echoes exhibit a constant pattern, while blood has a highly speckled and changing
pattern over time.
Besides the information about the amount and distribution of the plaque, IVUS
images provide a detailed description of plaque composition. The ultrasonic appearance of atherosclerotic plaque depends on its composition, and several components
of plaque can be identified in IVUS images.
During clinical imaging, several practical methods are used to enhance the
appearance of the different parts of the vessel. Saline injections help real-time
visualization of luminal border [97]. Injection of echo-contrast is another useful
technique for the detection of vessel borders [98]. Although these injections assist
in the better visualization of the arterial segment, they can also interrupt continuous
recording or even increase intracoronary pressure, which will result in erroneous
geometric measurements of the vessel components.

2.3.2 AUTOMATED METHODS

FOR

IVUS ROI DETECTION

The vast amount of data obtained by a single IVUS sequence renders manual

processing a tedious and time-consuming procedure. Furthermore, manually derived
data are difficult to reproduce because interobserver and intraobserver variability
can reach up to 20% [99]. Accurate automated methods for the detection of the
regions of interest in IVUS images improve the reproducibility and the reliability
of quantitative measures of coronary artery disease. Those methodologies usually
take advantage of the characteristic appearance of the arterial anatomy in twodimensional IVUS images and the connectivity of frames in the entire IVUS
sequence.
2.3.2.1 IVUS Image Preprocessing
IVUS frames contain noise, and the actual boundaries of regions of interest (ROIs)
are difficult to identify in many cases. A preprocessing step is essential in removing
speckles and artifacts that can interfere with the detection of desired boundaries.
Usually, in IVUS images, calibration marks are included for quantitative measurements because they provide useful information about the real dimensions of the vessel.
To remove all of the bright pixels constituting the calibration markers, substitution of
their gray-level value by the average or the median value evaluated in the neighborhood
of each pixel must be carried out [100, 101]. This operation may be preceded by
automated identification of the mark location based on the expected position and
isolation of the corresponding pixels using thresholding techniques [100].
Furthermore, the detection of regions of interest in IVUS images is restricted
by the existence of weak edges, and image enhancement is required. To enhance
image features, common image-processing techniques are used: median filtering [99,
101–103], Gaussian smoothing [101, 102], and nonlinear diffusion filtering based
on Euclidean shortening [102]. Repeated application of these filtering techniques is
acceptable for noise reduction. For contrast enhancement, a local columnwise histogram stretching can also be used [99].
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 69 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

69

A technique for blood noise reduction (BNR) in the imaged vessel lumen has
also been proposed [104]. This technique results in the edge enhancement of highfrequency IVUS images using a combination of spatial and temporal filtering, before
an automated algorithm for border detection is applied. The ratio between the highand low-frequency components is calculated using a fast Fourier transform, and
pixels are assigned as blood speckle or as tissue by thresholding this ratio. Different
filtering techniques are applied to blood and tissue. A limitation of the BNR algorithm arises from the hypothesis that tissue tends to be more consistent over time
and space than blood noise. However, pulsating motion of the arterial wall during
the cardiac cycle may disguise temporal or spatial fluctuations in the signals from
the vessels and thus affect the performance of the method.
Many techniques include a coordinate transformation [99, 100, 104, 108] to
restore the original polar format of the image data from the Cartesian values. This
results in the “straightening” of the borders of the regions of interest in IVUS images.
The coordinate transformation allows rectangular kernel sizes and linear convolution
(kernel motion) paths, and assists in the construction of searching graphs for the
extraction of the desired region borders.
2.3.2.2 IVUS Image Segmentation
Segmentation of IVUS images is a difficult task because of their complexity. The
efficiency of segmentation methods, which include a combination of thresholding
techniques, region growing, or dynamic contour models, has been examined in
several studies [99, 102]. In addition, more sophisticated techniques that exploit the
expected pattern of the regions of interest in IVUS data have been proposed [100,
101, 103, 104, 106, 108, 109, 111, 129].
Some of the earlier work on segmentation of IVUS images was based on heuristic
graph-searching algorithms to identify an optimal path in a two-dimensional graph
corresponding to the desired border [100, 104, 105]. For the accurate identification
of borders using graph searching, an appropriate cost function associated with the
graph is necessary.
Sonka et al. [100] have developed a method for detecting the internal and external
elastic lamina and plaque–lumen interface. First, the searching space in the image,

which includes the vessel except for the inner area of the lumen, is determined. After
the application of two different edge operators, the resulting edge subimages are
resampled in directions perpendicular to the outer or inner boundary of the ROI.
Those images are used to construct the laminae-border detection graph and the
plaque-border detection graph. Different cost functions are used for the detection of
each border. A compromise between the edge information of the image and the a
priori knowledge obtained from the shape of the ROI is achieved in the cost function.
After assigning the appropriate cost in all nodes of each graph, the optimal path
forming a closed boundary is defined as the path with the minimum sum of costs
of all nodes of the path.
The previously described BNR algorithm was combined with a graph-searching
method for the detection of external elastic membrane (EEM) and lumen borders
[104]. Gray images are converted into edge ones in rectangular format. A searching
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 70 Tuesday, May 10, 2005 3:38 PM

70

Medical Image Analysis

graph is constructed, with costs associated to the larger dynamic change of gray
level, the direction of change, and the likelihood of intensity in a given ROI. A
different searching strategy is performed for the detection of each border, and the
path with the minimum accumulative cost is generated, considering the continuity
in connecting nodes. Finally, the searched paths are mapped back to the original
image to form the desired border.
A texture-based approach has also been proposed [111] for the segmentation of
IVUS images. Textural operators were used to separate different tissue regions, and

morphological processing was used to refine extracted contours. The first step of
the method is the extraction of texture features and the association of the feature
vector to every image point. A neighborhood of 15 × 15 pixels was used for the
extraction of the fraction of image in runs (FOIIR) measure and the mean gray-level
measure (MGL). A histogram for each measure was constructed, and a threshold t
for both histograms was automatically selected that maximizes the interclass variance between regions separated by threshold t. Thus, since the lumen area is characterized by the absence of textural properties, all pixels with measure FOIIR(x,y)
below the threshold tFOIIR are classified into the lumen region. Accordingly, all pixels
with texture measures MGL(x,y) above the threshold tMGL are grouped into the
adventitia region. Afterward, a contour refinement was performed to remove errors
due to noise or distortions. A priori knowledge about the size and shape of the blood
vessel is used for the removal of inadequate-shaped objects and the selection of
appropriate structuring elements for the morphological processing that follows,
which results in improvement of the detected contours.
Methods that are based on the expected similarity of the regions of interest in
adjacent IVUS frames and that take into account the fact that the sequence of frames
constitutes a three-dimensional object have also been proposed [106–110]. Li et al.
[106] used a combination of transversal and longitudinal contour-detection techniques on the entire IVUS image sequence. The first step in this technique is the
reconstruction of longitudinal views of the vessel, using two perpendicular planes,
parallel to the longitudinal axis of vessel. In these planes, the contours corresponding
to vessel and lumen borders are automatically detected using a minimum-cost algorithm. The longitudinal contours intersecting the planes of the transverse images are
represented as edge points, guiding the final automated contour detection in the
cross-sectional IVUS images by defining the positions that the border line should
pass. A cost matrix is constructed for each IVUS image, with very low values
corresponding to the predefined four points. With the application of the minimumcost algorithm on the cost matrix, an optimal closed curve passing through these
points is obtained, which forms the border of the region of interest. The same strategy
is adopted for several studies on IVUS images [107, 109, 110].
A similar method that also includes stent detection in transversal images has
been proposed [108]. The stent-contour detection is performed only in the transversal
images because the appearance of the stent struts is much less regular in longitudinal
planes. First, the image is polar-transformed using the catheter as the center. A cost

matrix is used, whose element values depend on the intensity and the distance of
the corresponding pixel toward the catheter, and weight factors are also determined.
An initial model is created by applying the minimum-cost algorithm on the matrix,
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 71 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

71

and a second detection of the strut location is performed, resulting in a more refined
stent-contour detection. Longitudinal vessel detection is also performed, and the
vessel contours are detected simultaneously. In particular, both sides of the vessel
are searched for the selection of a strong transition at one side and a transition on
the other side that best match with the morphologic continuity and the geometric
characteristics of the vessel. Stent restrictions are also used, forcing the vessel
contour to lie outside the already detected stent. In this way, limitations on contour
detection that arise from the presence of calcified plaque or side branches in images
are overcome. The lumen contours are detected in the same longitudinal images
using information about the location of the catheter, the previously defined vessel
contour, and the gradient of the image. The contour detection in transversal images
is guided by the attraction points extracted from longitudinal contours.
Segmentation methods based on active contour models have also been proposed
for the processing of IVUS images [101, 103, 112]. The main advantage of active
contour models (“snakes”), compared with traditional edge detection approaches, is
that they incorporate spatial and image information for the extraction of smooth
borders of the regions of interest. An initial estimation of the wanted border must
be given as well as the curve deformations to obtain the final optimum shape. Thus,

isolated artifacts are ignored when they interfere with the smoothness of the curve.
A snake deforms under the influence of internal and external forces [113]. The
position of the snake can be represented by the curve v (s ) =  x (s ), y (s )  , where s ∈
[0, 1] is the arc length, and x, y are the Cartesian coordinates of each point of the
curve. The energy of the snake is given as

∫ ( E ( v ( s )) + E ( v (s ))) ds
1

Esnake =

int

image

(2.5)

0

where Eint represents the internal energy of the snake due to bending, and Eimage is
derived from image data.
The use of active-contour principles is suitable for border detection in IVUS
images because the desired borders are overall piecewise smooth with a low-variance
curvature. Algorithms that are based on active-contour models have to overcome
one major limitation arising from the classical snake properties. In particular, they
must ensure that the initial contour is placed close enough to the desired solution
to avoid unwanted convergence into a wrong (local) minimal solution.
A method based on active-contour models is described in the literature [103].
The initial estimation of the ROI border is given by the observer at the first frame
of the IVUS sequence, near the desired boundaries. The image force is appropriately

modified to force the snake to rest at points that separate large homogeneous regions
(placed on the boundary of lumen/media and media/adventitia). The minimization
of the energy function is performed by a Hopfield neural network [114]. The method
is further modified to detect the outer vessel boundary when calcium is present [129].
Under the perspective that the sequence of IVUS frames constitutes a threedimensional object, active-contour principles in 3-D space can be used to extract
Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 72 Tuesday, May 10, 2005 3:38 PM

72

Medical Image Analysis

the desired lumen and media/adventitia borders. An algorithm based on activecontour models in 2-D and its extension in 3-D is described in the literature [112].
The initial contour is placed around the IVUS catheter, and it can be represented
by r ≡ r θ , θ ∈ 0, 2π . The contour evolves under the influence of three forces:
the internal force, the image force, and the balloon force. Thus

()

Ftotal ( r ) = Fint ( r ) + Fimage + Fbal ( r )

(2.6)

The “balloon” force is added in the energy of the active-contour model and
causes the contour to inflate until the desired borders are detected. The application
of the 2-D algorithm results in a set of contours, which are then combined to form
a 3-D surface and used as the initial guess for the 3-D algorithm, in which appropriate
modifications to the forces and the representation of the contour are introduced.

A three-dimensional segmentation technique has been developed [101] for the
detection of luminal and adventitial borders in IVUS sequences. The method is based
on the deformation of a template by the features present in the 3-D image. This
algorithm is a 3-D extension of the digital dynamic contour (DDC) model reported
by Lobregt and Viergever [115]. The model comprises vertices (which are associated
with net force, acceleration, and velocity) connected with edges. While the vertices
of the model move, the initial contour deforms under the influence of internal and
external forces and a third dumping force that helps to bring the model to rest. The
contour obtains its final shape when the velocity and the acceleration of the vertices
become zero. Expanding the DDC algorithm in three dimensions, a cylindrical shape
is adopted as the initial surface model and it is allowed to deform under the influence
of the same three forces. The model is composed of vertices, determined in individual
contours, and connections between them are then defined. The internal force applied
to this model depends on transverse and longitudinal curvature vectors. Its components are given by:
fin,i, j ,trans = fin,i, j ,trans rˆi, j

(2.7)

fin,i, j ,long = fin,i, j ,long rˆi, j

(2.8)

and

where rˆi, j is a unit radial vector at vertex Vi,j. The magnitudes of transverse and
longitudinal internal forces are properly defined. The external force is the gradient
of a 3-D potential field that results from the preprocessing of IVUS images, and it
can be decomposed into two tangential and one radial component. The damping
force is a decelerating force acting at vertex Vi,j and is proportional to and directed
opposite to vertex velocity vi,j.

Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 73 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems

2.3.3 LIMITATIONS

IN

73

QUANTITATIVE IVUS IMAGE ANALYSIS

Many restrictions in automated segmentation of IVUS images derive from the quality
of the image, such as the lack of homogeneity of regions of interest and the shadowed
regions that are produced by the presence of calcium. The complicated structure of
human vessels and the different components in each part result in an image with
high-intensity variations, even in regions corresponding to the same tissue. In addition, calcified, hard-plaque regions are typically identified by high-amplitude echo
signals with complete distal shadowing. Consequently, it is not possible to identify
the morphology of the outer layers of the arterial segment, and in the absence of
contextual information from image frames adjacent in space and time, single-frame
IVUS images are difficult to analyze, even for the most experienced observers.
It must be reported that systolic–diastolic image artifacts frequently limit the
clinical applicability of automated analysis systems. A method of limiting cyclic
artifacts in IVUS images is based on electrocardiogram-gated (ECG-gated) image
acquisition, which is extensively used to overcome the problem of vessel distensibility and cardiac movement. The principle of ECG-gated image acquisition is
described by von Birgelen et al. [109]. A workstation is used for the reception of a
video input from the IVUS machine and the ECG signal from the patient. Upper

and lower limits for acceptable RR intervals, i.e., the time duration between two
consecutive QRS complexes, are defined (mean value ±50 msec) before image
acquisition begins. Images are acquired 40 msec after the peak of the R wave,
digitized, and stored in the computer. If an RR interval is too long or too short,
images are rejected, and the transducer does not move until the image can be acquired
during a heart cycle with the appropriate RR interval length. After an image is
acquired, the IVUS transducer is withdrawn in axial 0.2-mm increments through the
stationary imaging sheath to acquire the next image at that site. In general, ECGgated image acquisition, when combined with an automated boundary detection
method results in much smoother vessel boundaries.
In many cases, images of IVUS sequence are excluded from further analysis
because of the problems they exhibit. Common problems in IVUS sequences are
poor image quality, side branch attachments in the vessel under examination, extensive nonuniform rotational distortion, extensive calcification of the vessel wall, and
excessive shadows caused by stent struts.
The accuracy of the proposed segmentation algorithms would ideally be determined by the comparison of borders extracted automatically with the real borders
of the regions of interest. However, it is difficult to assess the accuracy and the
reliability of the suggested methods because the precise size and shape of the arterial
segment is unknown in vivo. For that reason, the manual tracing is used as the “gold
standard,” and the information that is often used is the location of these borders as
given by experienced observers, who generally have different opinions.

2.3.4 PLAQUE CHARACTERIZATION

IN

IVUS IMAGES

Plaque composition was shown to correlate with clinical variables in atherosclerotic
coronary artery disease [116, 117]. The composition of the plaque can be identified
in IVUS images, as demonstrated in several studies in comparison with histology
Copyright 2005 by Taylor & Francis Group, LLC

2089_book.fm Page 74 Tuesday, May 10, 2005 3:38 PM

74

Medical Image Analysis

[118, 119]. The classification of plaque in regions of soft (cellular), hard (fibrocalcific), and calcified plaque is based on the characteristic appearance of each one in
IVUS images. The components of soft plaque (highly cellular areas of intimal
hyperplasia, cholesterol, thrombus, and loose connective tissue types) in IVUS
images are regions of low contrast and homogeneous texture. On the other hand,
regions of hard plaque, which may also contain calcium, are characterized by bright
echoes (similar to adventitia), heterogeneous texture, and they are often trailed by
shadowed areas.
An automated method for assessing plaque composition in IVUS images has
been proposed by Zhang et al. [105]. The method proposed by Sonka et al. [100]
was used to detect the borders of the lumen and media/adventitia in the entire IVUS
sequence. To assess plaque composition, narrow wedges, called elementary regions,
were defined in plaque regions, and a classification label was assigned to them,
describing a soft or hard plaque. To classify elementary regions, several texturefeature measurements were computed. Gray-level-based texture descriptors — such
as histogram contrast, skewness, kurtosis, dispersion, variance, and the radial profile
property — are calculated for each elementary region. Co-occurrence matrices were
used, and such features as energy, entropy, maximum probability, contrast, and
inverse difference moment were computed. Two run-length features, such as short
primitives emphasis and long primitives emphasis as well as Brownian fractal dimension were also calculated. After having calculated these features, correlated ones
were removed, and among all features, radial profile, long run emphasis, and the
fractal dimension were identified as providing the best features for classifying soft
and hard plaques in IVUS images. These features were used for the training of a
classifier with piecewise linear discrimination functions. Afterward, each elementary

region was classified as containing soft or hard plaque. For the hard-plaque regions,
a further classification of hard plaque and shadow subregions was performed. When
the classification had been applied on the entire IVUS sequence, the plaque type of
each pixel was determined as the majority type among the pixels of the same spatial
location in a total of seven consecutive frames.
In the study of Vince et al. [120], the efficacy of texture-analysis methods in
identifying plaque components was assessed in vitro. IVUS images were captured,
and regions of interest were identified by microscopic examination of the histological
sections. Three plaque classes were considered: calcified, fibrous (dense collagenous
tissue), and necrotic core (lipidic pool with evident necrosis). Texture-analysis
procedures were applied in the region of interest, and the following statistical
techniques were evaluated: first-order statistics, Haralick’s method, Laws’s texture
energies, neighborhood gray-tone difference matrices (NGTDM), and the texturespectrum method. The selection of these methods was based on their ability to
differentiate soft tissue and textural patterns in two-dimensional gray-scale images.
After the implementation of these approaches, classification of texture features was
performed. The clustering ability of each of the examined texture-analysis techniques
was assessed. Haralick’s method demonstrated tight clustering of calcified, fibrous,
and necrotic regions with no overlap.

Copyright 2005 by Taylor & Francis Group, LLC


2089_book.fm Page 75 Tuesday, May 10, 2005 3:38 PM

Medical-Image Processing and Analysis for CAD Systems
z

z

y

x

75

x

(a)

(b)

y

FIGURE 2.6 (a) Estimation of the three-dimensional trajectory path from the biplane angiographical data; (b) mapping of IVUS frames along the pullback path in three-dimensional
space.

2.3.5 THREE-DIMENSIONAL RECONSTRUCTION
Three-dimensional reconstruction of the vessel based on IVUS yields more information than two-dimensional IVUS imaging alone in the visualization and assessment of coronary artery disease and the choice of intervention. To produce threedimensional renderings of vessel geometry, approaches that rely exclusively on IVUS
data perform a straight stacking of adjacent frames [107, 121, 122]. However, these
approaches do not account for the real spatial geometry of the coronary artery,
completely neglecting the influence of the vessel curvature, which induces an error
in quantitative measurements of the vessel [123].
In general, the determination of the position in 3-D space of an object, whose
shape and size are unknown, requires more than one view. For that reason, techniques
have recently been developed [124–127] to reconstruct the true spatial geometry by
combining IVUS and biplane angiography. These two modalities are well complementary and suitable for fusion, since biplane angiography provides longitudinal
projections of the vessel lumen, while IVUS provides transversal cross-sections of
the lumen and the wall.
The main concept of these approaches is illustrated in Figure 2.6. From the
angiographical data, a reconstruction of the catheter path during its pullback in 3D space (i.e., the pullback path) is obtained, and IVUS images are placed appropriately along this path. The steps of this procedure are depicted in Figure 2.7. Several

sources of errors can affect the accuracy of the 3-D vessel model. Apart from the
problems that each modality is associated with, problems that are closely related to
the fusion between both image modalities — such as the determination of the
pullback path, the estimation of the catheter twist, and the absolute orientation of
IVUS frame sequence — need to be resolved. The accurate estimation of the pullback
path in 3-D space is important for the correct positioning and orientation of the
IVUS images in 3-D space. The pullback path in the biplane angiograms can be
approximated either by the vessel centerline or by the location of the ultrasound
transducer in the vessel. In the first case, problems of overshadowed catheters are
overcome, but an angular error occurs whenever the catheter centerline is off the
Copyright 2005 by Taylor & Francis Group, LLC