Expert Systems with Applications 36 (2009) 7242–7251
Contents lists available at ScienceDirect
Expert Systems with Applications
j o u r n a l h o m e p a g e : w w w . e l s ev i e r . c o m / l o c a t e / e s w a
Computer aided diagnosis system of medical images using incremental
learning method
M. Park a,*, B. Kang b, S.J. Jin a, S. Luo a
a School of Design, Communication and Information Technology, Faculty of Science and Information Technology, The University of Newcastle, University Drive, Callaghan,
NSW 2308, Australia
b School of Computing and Information Systems, The University of Tasmania, Australia
a r t i c l e
i n f o
a b s t r a c t
Keywords:
Computer aided diagnosis
Incremental learning methods
Medical images
Multiple classification
Ripple down rule
Chest radiography
Intracranial CT angiography
This paper is about CAD in the medical imaging domain. CAD stands for computer aided detection or
computer aided diagnosis and the authors argue that both are important in assisting radiologists inter-
pret abnormal features in medical images.
The main novelty of this paper is the introduction of multiple classification ripple down rule (MCRDR).
The goal of the present work is to extend the RDR approach to produce multiple conclusions for a given
input, hence multiple classification ripple down rules.
These theoretical advances are joined with the intelligent computer aided diagnosis (ICAD) interface
that consists of three parts: image analysis, inference and reclassification. Once a medical image is loaded,
the system automatically extracts image features and the system indicates the radiologic findings. The
system enables only those attributes with abnormalities. The radiologist can add or modify the annota-
tion of the image, using the attributes window, by simply selecting the value of image attributes using
pop down menus to annotate any abnormalities.
Results are reported for a diagnostic knowledge base with 34 cases of chest radiographs selected in the
radiology department of St. Vincent’s Hospital, Sydney. Throughout this study, the authors proved that it
is possible to integrate the detection system and diagnosis system by proposing a new CAD architecture,
which supports multiple disease diagnosis and the learning of new adaptation knowledge. We also
showed that the diagnosis system could prevent radiologists from making misdiagnoses because of the
complexity of the anatomy and the subtlety of features associated with some abnormalities.
Ó 2008 Elsevier Ltd. All rights reserved.
1. Introduction
Remarkable advances have been made in the medical imaging
domain to deliver more sophisticated and complex medical images
from modalities such as computer tomography (CT), magnetic
resonance imaging (MRI), positron emission tomography (PET)
and X-ray systems. Computer assistance is needed to make it easier
for the radiologist to handle the information overload. To address
this need, a new class of products delivering computer aided diag-
nosis (CAD) capabilities has become available. CAD systems have
been developed to automate the interpretation process of medical
images.
In order to interpret medical images, the CAD should (a) detect
abnormalities on images first and (b) make diagnoses using the de-
tected abnormalities. Most CAD research focuses on the first step
only, detecting abnormalities. In terms of abnormality detection,
the image processing tool (IPT) plays a key role. It enhances the
* Corresponding author. Tel.: +61 2 4985 4518; fax: +61 2 4921 5896.
E-mail address: mira.park@newcastle.edu.au (M. Park).
0957-4174/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2008.09.058
particular abnormality features that are important in clinical
decision-making and suppresses the ‘noise’ that confuses that deci-
sion-making. More sophisticated IPTs have techniques for quanti-
fying visual features and for providing information to measure
geometric, topologic, or other characteristics by which images
are described. These various techniques became the foundation
of CAD technology although the larger role for IPTs used in CAD
is segmentation, which is separating an image into regions of sim-
ilar attributes. Based on these attributes, most CAD algorithms
adopted classification schemes such as the K-nearest neighbour
rule (KNN), Bayesian classifier and artificial neural network
(ANN). Although these classifiers are useful to mark the suspicious
region in the medical image this still belongs within the detection
process. Therefore, CAD is a detection level tool. There are many
successful CAD algorithms up to this level including the detection
of breast lesions on mammograms (Chan et al., 1987, 1990; Freer &
Ulissey, 2001; Giger, Huo, Jupinski, & Vyborny, 2000; Huo et al.,
1998; Jiang et al., 1999; Jiang, Nishikawa, Schmidt, Toledano, &
Doi, 2001; Kegelmeyer et al., 1994; Schmidt et al., 1996; Warren-
Burhenne, Wood, & D’Orsi, 2000; Yin et al., 1991), the detection
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
7243
of lung nodules in chest radiographs (Ashizawa et al., 1999; Difazio
et al., 1997; Ishida et al., 2002; Li, Katsuragawa, & Doi, 2000; Shi-
raishi et al., 2003; Uozumi et al., 2001) and thoracic CTs (Armato,
Giger, & MacMahon, 2001; Brown, Goldin, & Suh, 2003; Lawler,
Wood, Paunu, & Fishman, 2003; McCulloch, Kaucic, & Mendonca,
2004; Suzuki, Armato, & Li, 2003; Ukai, Kiki, & Satoh, 2000; Wor-
manns, Fiebich, & Saidi, 2002), and the detection of polyps in CT
colonographies (Acar, Beaulier, & Gokturk, 2002; Gokturk, Tomasi,
& Acar, 2001; Jerebko, Summers, Malley, Franaszek, & Johnson,
2003; Park, Hoffsttate, Jin, & Luo, 2006; Summers et al., 2001; Yos-
hida & Dachman, 2005; Yoshida, Nappi, MacEneaney, Rubin, &
Dachman, 2002).
We discovered that researchers did not notice, or overlooked,
the difference between computer aided detection and computer
aided diagnosis. This difference is critical to the design of CADs.
Previous CAD research defined CAD as a diagnosis made by a phy-
sician who takes into account the result of the computer output as
a ‘second opinion’ while the final diagnosis regarding the possible
state of the disease is left to the radiologist (Doi, MacMahon, Katsu-
ragawa, Nishikawa, & Jiang, 1997). In this definition, there are two
different levels of ‘second opinion’ for these domain experts: low is
a detection level of recommendation from CADs; and, high is an
interpretation level of diagnosis. However, if CAD provides an indi-
cation of a suspicious region and assists in recognising patterns in
medical images, it is better to defer to a computer aided detection
system not computer aided diagnosis. For the domain expert, it is
important to have the high level of ‘second opinion’ since the med-
ical image modalities are dramatically improved and IPTs based on
these images detect a larger number of, and more complex, abnor-
malities. It is important to differentiate between these two levels
and, because of this, we will redefine the low and high levels as
CADx, for computer aided detection, and CAD, for computer aided
diagnosis. The authors’ interest is in the high level CAD and the
main goal of this chapter is to propose a method to make this avail-
able in the best way. We believe the Multiple Classification Ripple
Down Rule (MCRDR) offers the means to achieve this and will pro-
vide the radiologist with the tools to produce multiple conclusions
for a given input. Note that in this chapter, we deal with the med-
ical imaging domain only although CAD can be applied to many
other domains.
2. Background
In the medical imaging domain, a diagnosis is the process of
identifying or determining abnormal features and the opinion de-
rived from that process. Diagnoses can be diseases, interpretations
or suggestions by radiologists. In CAD, this is considered a classifi-
cation task. In classification systems, acquiring knowledge from ex-
perts to classify objects, in this case medical images, is a necessary
process and is known as knowledge acquisition in the expert sys-
tem area. Therefore, knowledge acquisition is essential in develop-
ing successful CADs for classifying images from domain experts.
The acquisition of this knowledge is very difficult because the im-
age interpreting process cannot be easily explained by domain ex-
perts. This is caused by the nature of human knowledge which is
not systematically organised and documented, a concept discussed
later in this section.
In order to discuss diagnosis systems for medical image inter-
pretation, it is necessary to examine expert system technology be-
cause the authors’ CAD uses theories from this area. In expert
systems, there are two major components: inference engine and
knowledge base (Giarrantano, 1989). The inference engine gener-
ates interpretations using the knowledge base. Rules are used as
the representation for knowledge in the knowledge base and the
interpretations by the inference engine are diagnoses, classifica-
tions or conclusions. High quality rules within the knowledge base
are built by domain experts using the knowledge acquisition meth-
od. This method is the key to the success of expert systems as they
are bound to the quality of acquired knowledge. However, knowl-
edge acquisition is a difficult process within expert systems be-
cause domain experts usually provide incomplete, even incorrect,
knowledge as they are unable to articulate it. This is called ‘knowl-
edge acquisition bottleneck’ (Churcharoenkrung, Kim, & Kang,
2005).
There are few CADs which attempt to diagnose based on ex-
tracted abnormal features. Abe, Ashizawa, Katsuragawa, MacMa-
hon, and Doi (2002) designed an ANN to differentiate between
11 types of interstitial lung disease by using up to 10 clinical
parameters and 16 radiologic findings. The radiologic findings
were converted to numerical values from 1 to 1.0, which were used
as input to the ANN. The output of the ANN also ranged from 0 to
1.0, which corresponds to the likelihood of each disease. However,
this approach is not free from the problem of knowledge acquisi-
tion bottleneck and it is important to study why this bottleneck
has been a problem in expert system areas and whether there is
a better method we can use in CAD.
Machine learning and ANN are the most popular methods used
in expert system development. In these approaches, the quality of
the knowledge base is completely dependent on the number of
training cases. A large number of training cases can give the knowl-
edge base higher quality but this does not necessarily mean that it
will cover all domain knowledge. (Additionally, the quality within
these training cases depends on the noise level, which is the ratio
of misclassified cases from the training set.) This knowledge base
quality issue is relevant in the medical imaging domain both in
acquiring the expert knowledge and maintaining this knowledge.
Difficulty in acquiring human knowledge from domain experts in
a systematic way is well known in expert system studies as it is
not possible to collect all cases within the domain. As mentioned
previously, human knowledge is not systematically organised
and documented. Knowledge is always context dependent and
constructed ‘on the fly’ when it is needed (Compton, Peters, Ed-
wards, & Lavers, 2006). This theory is already well-established in
the cognitive science and philosophy communities. In the cognitive
field, Clancey (1997) discussed the concepts of ‘situated cognition’
and ‘classification task’ and proved how knowledge was acquired
in the real world. (Within the medical imaging domain, these con-
cepts are why radiologists are proficient at making a diagnosis but
find it difficult to explain how they reach a decision or solve a
problem.) However, many scientists and knowledge engineers in
expert system research overlooked, or did not notice, these two
concepts when they developed their knowledge acquisition meth-
ods. This oversight created problems when knowledge engineers
interviewed domain experts and attempted to compile the knowl-
edge base for the expert systems because the knowledge was
incomplete and to increase, change and maintain this knowledge
was expensive.
In addition, domain knowledge in the human expert is not static
but dynamic and this is particularly relevant in maintaining the
knowledge base. The main task of knowledge base maintenance
is to change and evolve the knowledge and this maintenance is
expensive. In general commercial software development life cycle
(SDLC), the maintenance cost is two-thirds of the SDLC outlay
(Meilir, 1980). In terms of software development, the maintenance
of the knowledge base is recognised as the most costly part of ex-
pert system development and this is increasing as the size and
complexity of the knowledge is increasing. The most popular
method of maintenance in the expert system is verification and
validation (V&V). V&V ensures that a knowledge base system per-
forms as it is intended (Preece & Shinghal, 1992). Verification is
normally concerned with determining the internal consistency of
7244
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
a knowledge base. The normal approach in verification is to at-
tempt to reduce the knowledge base to pathways from data to con-
clusions and then look at
the relationships between these
pathways, the data they use, the intermediate conclusion they
establish, etc. (Mattler, 1987). Validation, in terms of maintenance
or incremental acquisition, is concerned with testing whether
other cases that were correctly classified previously will be mis-
classified by a new rule, as well as ensuring the new rule covers
the new case. In summary, V&V is a process to check for conflicts
within the rules (verification) and to check that the outcome of
the corrected knowledge base works properly for all input cases
(validation). In most other expert systems, the V&V process is car-
ried out after knowledge acquisition. Therefore, the cost to do V&V
is closely related to the efficiency and time of the maintenance pro-
cess. That is, the more expensive V&V becomes, the more expen-
sive the maintenance process of the knowledge becomes overall.
This problem gets worse as the size of the knowledge base grows.
The knowledge maintenance problem caused by the V&V process is
understood to be the main cause of knowledge acquisition bottle-
neck. Because we adopted the concept of the expert system into
CAD, the CAD should overcome these presented issues by applying
MCRDR.
However, the version before MCRDR was the ripple down rule
(RDR) approach and it worked in a similar way by eschewing any
notion of extracting or mining the domain expert’s knowledge.
RDR grew specifically from the experience gained in maintaining
an early medical expert system, the GARVAN-ES1, for a number
of years (Compton & Horn, 1989; Compton & Jansen, 1990). Obser-
vation of experts during maintenance suggested that experts never
provide information on how they reach a specific judgment.
Rather, the expert provides a justification that their judgment is
correct. The justification they provide varies with the context in
which they are asked to provide it (Compton & Edwards, 1992; Ed-
wards & Compton, 1993). The Pathology Expert Interpret Report
System (PEIRS) expert system, used to add clinical interpretations
to chemical pathology laboratory reports (Compton & Edwards,
1992; Edwards & Compton, 1993), is the major success with RDR.
PEIRS went into service with about 200 rules with the rest of the
rules (at last count up to about 1800) added while in routine use.
Most importantly, all rules were added by a domain expert with
no special computer skills and without any assistance from a
knowledge engineer or computer programmer. A knowledge engi-
neer/programmer was required for the initial data modeling only.
After the PEIRS system successfully used RDR, MCRDR, a general-
ised version of RDR, was introduced (Kang, 1996). PEIRS was rede-
veloped for MCRDR to use and it has been successful as the
commercial medical expert system (Compton et al., 2006). Through
this development, MCRDR has proven that domain experts can
maintain the knowledge base without help from a knowledge engi-
neer and the maintenance cost can be kept relatively low when
compared with other expert systems (Bindoff, Tenni, Peterson,
Kang, & Jacson, 2007; Hoffmann, Kang, Richard, & Tsumoto,
2006; Kim et al., 2006; Kim, Kang, Compton, & Motoda, 2007).
3. Multiple classification ripple down rule
MCRDR consists of two operational modules, inference engine
and knowledge acquisition engine, and two storage modules,
knowledge base and cornerstone cases. The inference engine uses
the knowledge base to interpret the input cases. In this instance,
the input case is the detected abnormal features from a medical
image. The knowledge acquisition engine is used to acquire or
maintain the knowledge base when cases are misclassified by the
inference engine. The knowledge base is a set of rules in the n-
ary tree format. The knowledge acquisition engine is required by
domain experts to correct the knowledge base when the case is
misclassified. After the knowledge base is, the misclassified case
should be classified correctly by the inference engine. The case
then becomes one of the cornerstone cases and each case is as-
signed to the rules that were added during knowledge acquisition.
These cornerstone cases are used in a similar way to the V&V case
set (see Section 2) in other expert system developments. However,
the role of cornerstone cases in the knowledge acquisition of
MCRDR is more important than the V&V set in other expert sys-
tems. In MCRDR, cornerstone cases are used for domain experts
to select conditions for the new rule and guarantees that the
V&V process is not required after knowledge acquisition. This is
one of the main strengths of MCRDR and why the cornerstone
cases are important in the knowledge acquisition process. The de-
tailed algorithms for MCRDR components will be explained in the
following sections.
3.1. Inferences
For each new input case, MCRDR evaluates all the rules in the
first level of the knowledge base. It then evaluates the rules at
the next level of refinement for each rule that was satisfied at
the top level and so on. The process stops when there are no more
rules to evaluate or when none of these rules can be satisfied by
the case at hand. It thus ends up with multiple paths, with each
path representing a particular refinement sequence and, hence,
multiple conclusions. The structure of an MCRDR knowledge base
can be drawn as a n-ary tree with each node representing a rule.
Fig. 1 shows such a structure and also shows the inference for a
particular case.
3.2. Knowledge acquisition
When a case has been classified incorrectly or classification is
missing, knowledge acquisition is required. This can be divided
into three parts. Firstly, the system acquires the correct classifica-
tion from the expert. Secondly, the system decides on the new
rule’s location. Thirdly, the system acquires new rules from the ex-
pert and adds them to correct the knowledge base. It is likely that
experts will find the system more natural if the order of steps two
and three are reversed, therefore hiding in a better way the implicit
knowledge engineering that is going on. However, the order is not
crucial in terms of the algorithm.
3.2.1. Acquiring new classifications
Acquiring new classifications is straightforward; the expert
simply needs to state them. For example, referring to Fig. 1, if the
system produces classification class 2, class 5 and class 6 for a gi-
ven problem, the expert can decide that class 6 does not need to
be changed but class 2 and class 5 should be deleted and class 7
and class 9 added.
3.2.2. Locating rules
The system should automatically find the location for the new
rules. If a new classification is the refinement of a wrong classifica-
tion, the system attaches the new rule to the classification that
produced that wrong conclusion. In other words, the system adds
a new rule to a different place from the current conclusion in the
tree. As well as attempting to decide whether a classification is
best seen as a refinement only or an independent classification,
we note that in some ways it does not matter – both are workable
solutions for any classification. After the system decides all new
rule locations for new conclusions, it attaches a stopping rule
(which has a null conclusion) to every rule that produces a wrong
conclusion that is not as yet refined by the new rule. Stopping rules
play a major role in MCRDR in preventing wrong classifications
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
7245
The highlighted boxes represent rules that are satisfied for the case {a, c, d, e, f, h, k}. Pathways
athrough the knowledge base. The rules producing conclusions are highlighted. Info n […]
indicates other rule numbers with the same classification.
Fig. 1. Knowledge structure and inference.
being given for a case. The inference process can be understood in
terms of capturing ‘paths’, as shown in Fig. 1. Where paths are pro-
duced there are a number of questions about whether the path
produces a classification, whether the classification is redundant
because it is produced elsewhere, etc.
3.2.3. Acquiring rule conditions – rule validation
In MCRDR, we are mainly interested in validating a knowledge
base system by testing it. A standard technique is to use a database
(Buchanan & Barstow, 1983) of typical cases. In this situation, one
depends on the cases being representative of the cases the system
is meant to cover. With RDR, one case is associated with one rule
because the rule is added to deal with that particular case. A new
rule must distinguish between the case that caused its creation
and the case associated with the rule that gave the previous incor-
rect classification. With MCRDR, a number of cases, cornerstone
cases, can associate with a new rule and the higher the rule is in
the tree, the higher the number of cases that can associate with
that rule. The new rule should distinguish between the new case
and all of these cornerstone cases. In other words, MCRDR has mul-
tiple cornerstone cases for a rule, compared to RDR where there is
one case per rule.
A rule at a level can be hit by all the cases associated with rules
at the same level and sub-rules lower in the system. Therefore, the
rule has to be made sufficiently specific so that none of these other
cases satisfy it. However, it does not matter if other cases that in-
clude the same classification reach this particular rule. If a rule is
added at a level below the top level, only cases that satisfy the par-
ent rule above need to be considered as cornerstone cases. Note
that as the system develops, cases may arise that correctly satisfy
a rule, but may be added to the system because a rule is needed
elsewhere to add a further classification. Such a case will become
a cornerstone case for new rules below the rule it satisfies and
for which the classification is correct. As the tree develops, the
rules lower down will naturally have less cornerstone cases associ-
ated with them.
Therefore, the aim is to make a new rule sufficiently precise so
that it satisfies only the case it is being added for and no other
stored cases, even though it does not matter if it satisfies cases that
include the same classification. The algorithm for selecting condi-
tions to make the rule sufficiently precise is very simple but some
discussion is needed as to why a more sophisticated approach was
not chosen. Consider a new case A and two cornerstone cases B and
C. In creating a new rule, one may imagine that the domain expert
should choose at least one of the conditions from (case A (case
B [ case C)) or the negated condition of ((case B \ case C) case A).
However, as seen in Table 1, the result may be nothing, lead-
ing to the situation where no rule conditions can be found. Alter-
natively, the difference list may contain only trivial conditions
that are irrelevant. In other words, there are no common condi-
Table 1
1-A has some conditions in the difference list between A and (B and C), however, this
is not always true as is shown in 1-B
A
1-A
a
b
c
f
B
c
d
e
f
C
b
c
d
g
Difference list between A and (B and C)
a
Not d
A
B
C Difference list between A
and (B and C)
Difference list
between A and B
Difference list
between A and C
1-B
a
b
c
f
a
c
d
e
f
b
c
d
g
b
Not e
a
f
Not g
7246
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
tions that distinguish the presented case from all the other cor-
nerstone cases, but a number of different conditions distinguish
different cases and these conditions must all be included in the
new rules.
The cornerstone cases are recorded during the addition of new
rules. When a new rule is added to the system, the rule should con-
tain the differences among the cornerstone cases that satisfy the
rule and the input cases. In order to exclude a further case when
other stored cornerstone cases satisfy the rule, additional condi-
tions must be added. The process is repeated until there is no
stored cornerstone case satisfying the rule. If the conclusion of
the rule is new, then no override statement is added. It should be
noted that the above is the logical description of MCRDR since in
real implementation of the system all the rules are ordered in a
n-ary tree hierarchy.
4. Intelligent computer aided diagnosis system
Intelligent computer aided diagnosis (ICAD) consists of two
phases: abnormal feature detection and diagnosis (see Fig. 2).
The detection phase has two processes: image processing and sym-
bolic processing. Image processing includes all image analysis
schemes that obtain information from the images such as lung
boundary, abnormal texture, rib density and hila size. Symbolic
processing uses fuzzy function to convert the numeric data ob-
tained from the image processing into symbolic descriptions. Diag-
nosis phase interprets the abnormal features obtained from feature
detection phase and this phase has a strong resemblance to what is
generally meant by the term ‘intelligent cognition’.
The authors’ test medical imaging modality was posteroanterior
(PA) chest radiography. PA chest radiographs, often called ‘the mir-
ror of health’ (Mattler, 1987), make up a large portion of all radio-
graphs because of the relatively low cost in terms of both resources
and time. The complexity of the anatomy and the subtlety of fea-
tures associated with some abnormalities make this image modal-
ity highly suitable to evaluate the ICAD system proposed above.
4.1. Feature detection system
4.1.1. Image processing
Image processing produces image features in the form of nu-
meric data from an input chest radiograph. There are four mod-
ules: lung segmentation, texture analysis, rib detection and hila
detection.
4.1.1.1. Lung segmentation. The lung field is extracted to obtain the
boundary information of the lung. The knowledge based lung field
extraction method, developed by Brown, Wilson, Doust, Gill, and
Sun (1998) and extended by Park, Wilson, and Jin (2001), is applied
to segment and analyse the lung boundaries. Information obtained
includes the presence of edges, the cardiothoracic ratio, the posi-
tion of the diaphragm, the costophrenic angle and lung volume.
4.1.1.2. Texture analysis. The lung texture is analysed using quasi-
Gabor filter (Park, Jin, & Wilson, 2002a, 2002b) and 3D structure
classification with Score Block Operation (Park, Jin, & Wilson,
2002c) to solve the problem of identifying patterns and distin-
guishing the complex background of superimposed structures in
chest radiographs. The quasi-Gabor filters are capable of maintain-
ing low computational cost while keeping the important informa-
tion of the power spectrum, such as band-pass frequency and
direction of texture, of the 2D-discrete Fast Fourier Transform.
The 3D classifier is able to capture not only local texture but also
the global distribution of lung texture. Lung field is divided into
right- and left-lung and each field is sub-divided into upper-, mid-
dle- and lower-lung. Lung texture is categorized as normal, dots
and grape-like.
4.1.1.3. Rib detection. To make rib edges clear, an expiration lung
field using ‘hemi-elliptical cavity’ method is produced. Based on
this, the rib edges are located using canny edge detector and a
new connectivity method called ‘4 way with 10-neighbours con-
nectivity’ (Park, Jin, & Wilson, 2002d). The ribs and clavicles are de-
tected for labeling and checked for density.
Expert:
Radiologist
User interface
Medical
Image
Features
Interpretations
Knowledge Acquisition
Interface
Image
Processing
Symbolic
Processing
Inference Engine
KA Engine
Feature Detection System
MCAD System
Knowledge
Base
Cornerstone
Cases
Diagnosis System (MCRDR System)
Fig. 2. ICAD system overview.
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
7247
4.1.1.4. Hilar region detection. A hilar region is a depression or fis-
sure where vessels or nerves enter an organ. This region is not easy
to find because it is superimposed with other structures of the lung
creating a complex structure that is difficult to define. We use
threshold scheme to detect and measure the hilar region (Park,
Jin, & Wilson, 2003).
4.1.2. Conversion of numeric data into symbolic description
In order to apply MCRDR, we should define the description of
medical image features to allow high level symbolic processing.
For this problem we adopt fuzzy set theory pertaining to linguistic
variables (Zadeh, 1975) allowing representation of vague natural
language such as big, middle, and normal in mapping between
numerical parameter values and linguistic descriptions, and vice
versa.
4.1.3. Linguistic uncertainty
Anatomical descriptions used by experts contain uncertainty
because of the imprecision of natural language. By using familiar
linguistic variables, uncertainty can be introduced intuitively at a
symbolic level. For each linguistic expression, a fuzzy set is created
which provides compatibility (fuzzy membership) scores for possi-
ble parameter values. Fuzzy sets have previously been used to link
numerical values to symbolic representations for high level match-
ing in medical images (Buckley, Siler, & Tucker, 1986).
Typically, the same set of linguistic variables is applied to a gi-
ven feature for all anatomical structures. For example, the linguis-
tic variables, lost, not_sharply_defined and sharply_defined may be
used to represent the edges for all organs. The authors’ approach
is different in that we redefine linguistic variables for each anatom-
ical structure in the model because, for example, a reduced lung
represents different numerical values to those for a reduced hilar
region. Furthermore, rather than modeling a structure as being re-
duced or enlarged we believe that a more natural description can be
derived relative to normal with structurally unusual variations
being modeled as, for example, abnormally high, abnormally low,
sharply defined or lost. Such a representation is consistent with
the assumption that the radiologist makes a diagnosis largely by
recognising normal anatomy and excluding it from their attention,
and therefore the radiologist is readily able to supply knowledge in
this process.
4.1.4. Fuzzy logic
Fuzzy logic is a method that permits a gradual representation of
likeness between two objects. It is based on Zadeh’s theory of fuzzy
sets (Zadeh, 1975). He uses a membership function to assign a
grade of membership between 0 and 1 to each element in the range
of all possible elements under consideration. This grade can be
thought of as a measure of compatibility between the element
and the concept represented by the fuzzy set. Formally, the mem-
bership function for a fuzzy set A, written lA(x), is a real valued
function defined as the application lA : X ? [0,1] for all x in a uni-
versal set X.
The application of fuzzy sets is fundamental if one wishes to
model common terms involving vagueness or imprecision. Modifi-
ers, such as very and more, control the amount of vagueness and
such terms are frequently used in natural language. For example,
we can say a person is tall or very tall.
In fuzzy logic there are a number of approaches that allow infer-
ence and we used the generalised modus ponens (GMP), one of the
most popular in literature. Through the GMP technique, the prop-
osition y is B can be derived from the rule if (x is A) then (y is B)
when the proposition x is A is true. The GMP can also be employed
when the two propositions x is A and y is B are defined imprecisely.
Thus, if a proposition x is A0 close to x is A is true, the principle of
the GMP is to derive another proposition, written y is B0 close to
y is B. This proposition is generated by taking into account both
the underlying semantics of the implication of the rule and a mea-
sure of the likeness between A and A0. With all, the inference con-
sists of defining a fuzzy set B0, which is as close to B as A0 is to A.
More formally, Dubois and Prade (1988) have computed the mem-
bership function of B0, written lB0
8y 2 Y; lB0ðyÞ ¼ supx 2 xTðlAðxÞ;ðx ! yÞÞ
where Y is the universal of y, X is the universal of x, T is a triangular
norm that makes the GMP compatible with the classical modus
ponens, lA0 is the membership function of A0 and (x ? y) represents
an implication denoting the kind of causal link involved in the pro-
duction rule.
For example, suppose that TðlA0ðxÞ;ðx ! yÞÞ is the minimum
operation between two numbers and (x ? y) is defined as min(x,y),
the notation m/x standing for ‘the value x has degree of member-
ship m0. Then, suppose that the rule IF (x is A) THEN (y is B), where
A = (0/100,0.5/125,1/150,0.5/175,0/200) and B = (0/10,0.5/20,0.5/
30,0.5/40,0/50).
In this chapter, we implemented the GMP by adopting the fuzzy
expert system (FES) concept introduced by Buckley et al. (1986).
According to them, every FES must hold the following:
It can manipulate fuzzy terms.
Its input is comprised of imprecise attributes, which can be rep-
resented by either discrete fuzzy sets or continuous fuzzy sets.
Its rules are defined so that they can operate with fuzzy data.
The final result is a fuzzy set.
Q
Q
We used the triad formed by the Z-function, the
-function and
the S-function for defining fuzzy membership functions (i.e., fuzzy
sets). These are particularly interesting when the usual values gi-
ven to fuzzy terms employed by experts can be represented by
means of the triple (T) hZ-function,
-function,S-functioni. Exam-
ples of this are hlow,normal,highi, hsmall,normal,bigi, etc. We will
assign a term to each component of T as Tlow, Tnormal and Thigh,
respectively. For example, in Fig. 3, a height of 2 cm would be de-
scribed as normal (l(x) 1.0), a height of 3 cm would be described as
abnormally high (l(x) 1.0) and a height of 2.3 cm has l(x) of 0.6 in
the normal set and has l(x) of 0.4 in the abnormally high set, so the
system might return slightly high as the linguistic value.
4.2. Diagnosis system
We have discussed why the diagnosis system is required in
CADs and we joined the MCRDR method to the CAD to develop
the ICAD system. This section describes how the diagnosis system
works in ICAD (Fig. 2).
4.2.1. Inference engine
The inference engine in ICAD requires an input case and a set
of detected abnormal features from the medical image as ex-
low(x)
T
T
normal
(x)
high(x)
T
µ(x)
1.0
0.6
0.4
0
1
2 2.3
3
(height of left hilum
above the right (cm)
Fig. 3. Fuzzy sets for the asymmetry between the LHS of the diaphragm (lower) and
the RHS (higher).
7248
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
plained in the previous section. As the outcome of the feature
detection system, the abnormal features will show on the inter-
face of ICAD (Fig. 4). The user then asks the system for inference
by clicking the ‘infer’ button. The system will infer the knowledge
base with the given case (detected abnormal features) and pro-
duce the interpretation. ICAD shows the conclusions/interpreta-
tions for the input case from the last satisfied rule in the
knowledge base. This is a core part of implementation of the
MCRDR concept; the root rule is always set as true regardless of
the values of the input cases. The conclusion of the root rule is
a default interpretation and it will be presented if no rules are sat-
isfied other than the root rule. In ICAD, the conclusion of the root
rule is normal. Hence, there is always at least one rule to be sat-
isfied and the system will show at least one interpretation. Note
that there is always at least one interpretation from the system
and it may not be a root rule. The mechanism to find the conclu-
sion from the last satisfied rule in a path (see Section 2) is always
possible because there is at least one rule which is satisfied using
the cases. This concept is important to simplify the knowledge
acquisition process.
4.2.2. Knowledge acquisition engine
The knowledge acquisition engine is utilised when the radiolo-
gist does not agree with the interpretation from the system. There
are three different types of disagreement noted in the ICAD knowl-
edge acquisition process: adding more interpretations for the case;
replacing (or correcting) some of the provided interpretations from
the system; and, removing some of the interpretations from the
provided interpretations. In ICAD knowledge acquisition, these
three types will be managed by the same steps with MCRDR: find-
ing a new rule location, finding conditions for the rule using cor-
nerstone cases, adding a new rule, and, adding the current case
to the cornerstone cases. This is shown in Fig. 2.
4.3. User interface
The ICAD user interface consists of three tasks – image analysis,
inference and reclassification – and three windows – image win-
dow, attribution window and conclusion window. Once an image
is loaded, ICAD automatically extracts abnormal image features
and indicates the radiologic findings. ICAD enables only those attri-
butes with abnormalities on the attribute window. The radiologist
can add or modify the annotation of the image using this window
by using pop down menus and simply selecting the value of image
attributes to annotate any abnormalities. This type of communica-
tion reduces the misunderstanding between the radiologists and
the system engineers. The very last step is to type the conclusion
in the conclusion window.
MCRDR is applied to infer the conclusion based on the enabled
attributes to indicate the abnormalities. If a radiologist agrees with
the conclusion, the radiologist merely references the conclusion for
the final decision. However, if the radiologist does not agree with
the conclusion, the case is reclassified by adding a new conclusion
to the knowledge base.
Fig. 4 shows an example of ICAD. It displays a chest radiograph
with lung boundaries, hilar region and lung texture. The attribute
window enables the abnormalities found by the system and sug-
gests the possible disease states. Knowledge-base viewer shows
the history of the learning for case 11. The initial conclusion was
‘normal’ since case 11 was a new case and there was no conclusion
(see the root node of the tree). ‘Mucoviscidosis’ was added as a
conclusion for case 11 and ‘other disease 11’ and ‘Mucoviscidosis’
were added as the new conclusion. The new conclusion has been
added for case 11, but the final node of the tree is the current
conclusion.
Fig. 5 shows the ICAD interface for the intracranial CTA. It dis-
plays axial, sagittal and coronal views and 3D view. The red color
Fig. 4. ICAD application for chest radiograph.
M. Park et al. / Expert Systems with Applications 36 (2009) 7242–7251
7249
Fig. 5. ICAD application for intracranial CTA.
indicates the segmentation of the vessels and the purple color indi-
cates the aneurysms. This works the same as the chest radiograph
system discussed above.
5. Experiment results
We constructed the prototype of ICAD and used 34 cases of
chest radiographs collected in the radiology department at St. Vin-
cent’s Hospital, Sydney, to evaluate the system performance. The
total number of possible features detected in the medical images
by the feature detection system was 45. Approximately 29 rules
were entered by the radiologist using the knowledge acquisition
system. The average number of conditions used in the rules was
13 – the minimum was 3 and the maximum was 21. The average
estimated time to enter a rule was approximately 15 min and the
estimated total time to make a rule set was 7 h and 25 min. How-
ever, this includes the time to enter the interpretations by domain
experts because the knowledge base was developed from scratch.
The time to enter the interpretations is not normally counted as
part of knowledge acquisition because they only need to be en-
tered once. According to the other commercial system, LabWizard,
the pathology interpretation system takes about 1 min per rule
(Compton et al., 2006). The knowledge in LabWizard should not
be too different from the medical imaging domain because the fea-
tures are all symbolic descriptions such as text values. In fact, the
rule for entering time in most MCRDR systems in other expert do-
mains is always measured in the low minute range because the
system only asks experts to identify the features to justify the gi-
ven case, not explain the theory.
It is impossible to know how much of the domain was covered
by these 34 cases. However, they were selected by the domain ex-
perts from a large set of medical images for different possible inter-
pretations. Note that most other medical image interpretation
systems deal with a constrained knowledge base because all cases
cannot be entered, but the coverage of the authors’ system is rela-
tively wider than others because ICAD learns incrementally.
6. Discussion
To evaluate ICAD, the abnormal feature detection system and
the diagnosis system need to be assessed separately and then over-
all performance of the whole system needs to be judged against
other systems. The evaluation of ICAD as a whole against other
CADs is difficult because this is a relatively new approach and there
are very few that integrate the abnormal feature detection and
diagnosis systems.
There are a few CADS using the classic AI technology in their
diagnosis systems and comparing ICAD with these can be consid-
ered, although the evaluation result will not show a meaningful
comparison. The final interpretations from the diagnosis system
in CADs are dependent on the feature detection system. The detec-
tion systems in all these CADs take different approaches and the
outcomes from the feature detection system will influence the per-
formance of the diagnosis system. Therefore, the accuracy of final
interpretations is influenced by both components.
In earlier studies, the authors evaluated feature detection sys-
tems (Brown et al., 1998; Park et al.,2001, 2002a, Park, Jin, & Wilson,
2002b, 2002c, 2002d, 2003) and the superiority of MCRDR, which is
the main diagnostic system method in ICAD, was proven in other
medical expert system areas such as pathology and medication re-
view. The type of clinical knowledge in these areas using MCRDR is
very similar to the medical imaging domain (Bindoff et al., 2007;
Compton et al., 2006). However, it is not possible to compare
MCRDR method directly to other machine learning methods be-
cause it learns incrementally while the majority of other ap-
proaches try to learn the complete knowledge base at once.
MCRDR, itself, is evaluated by using the statistics from logs of rule
acquisitions by experts. There were good empirical evaluation
studies done against machine learning methods using the simu-
lated experts (Bindoff et al., 2007) and it shows the strength of
MCRDR very well. According to this evaluation study, MCRDR
method learns domain knowledge with a small number of cases
much better than machine learning approaches and shows a
similar performance after enough number of cases are available.