J Med Syst (2009) 33:413–421
DOI 10.1007/s10916-008-9203-3
ORIGINAL PAPER
Cholangiocarcinoma—An Automated Preliminary Detection
System Using MLP
Rajasvaran Logeswaran
Received: 30 May 2008 / Accepted: 4 August 2008 / Published online: 12 August 2008
# Springer Science + Business Media, LLC 2008
Abstract Cholangiocarcinoma, cancer of the bile ducts, is
often diagnosed via magnetic resonance cholangiopancrea-
tography (MRCP). Due to low resolution, noise and
difficulty is actually seeing the tumor in the images,
especially by examining only a single image, there has
been very little development of automated systems for
cholangiocarcinoma diagnosis. This paper presents a
computer-aided diagnosis (CAD) system for the automated
preliminary detection of the tumor using a single MRCP
image. The multi-stage system employs algorithms and
techniques that correspond to the radiological diagnosis
characteristics employed by doctors. A popular artificial
neural network, the multi-layer perceptron (MLP), is used
for decision making to differentiate images with cholangio-
carcinoma from those without. The test results achieved
was 94% when differentiating only healthy and tumor
images, and 88% in a robust multi-disease test where the
system had to identify the tumor images from a large set of
images containing common biliary diseases.
Keywords Magnetic resonance cholangiopancreatography
(MRCP) . Cancer . Tumor . Bile ducts .
Computer-aided diagnosis
R. Logeswaran
Global School of Media, Soongsil University,
Seoul, South Korea
R. Logeswaran (*)
Faculty of Engineering, Multimedia University,
63100 Cyberjaya, Malaysia
e-mail: loges@ieee.org
Introduction
Cancer, in its many varieties, is on the rise in most parts of
the world. Increasingly common in modern medical
technology is the use of computers as an aid in the
diagnosis of diseases. Artificial neural networks (ANN)
are commonplace in such efforts, including in cancer
detection systems of the liver [1], brain [2], breast [3], lung
[4], thyroid [5] etc.
Bile is used in digestion for absorption of fat-soluble
nutrients, as well as in the removal of fat-soluble waste
products (such as cholesterol) from the body. Produced in
the bile ducts of the liver and stored in the gallbladder, bile
is transported via the biliary tract to the small intestines
during digestion. Diseases affecting the biliary structures
block the biliary system causing swelling, pain and a
buildup of toxins in the body. Cholangiocarcinoma or
tumor of the bile ducts is on the rise [6]. It is the second
most common primary malignant tumor of the liver after
hepatocellular carcinoma and comprises approximately 10
to 15% of all primary hepatobiliary malignancies [7].
Clinical diagnosis for this disease depends on appropri-
ate clinical, imaging, and laboratory information [7]. The
signs and symptoms include clay colored stools, jaundice
(yellowing of the skin and eyes), itching, abdominal pain
that may extend to the back, loss of appetite, unexplained
weight loss, fever, chills [8] and dark urine [9]. Follow-up
tests include non-invasive imaging such as computed
tomography (CT), magnetic resonance imaging (MRI) or
ultrasound; while invasive ones include endoscopic retro-
grade cholangiopancreatography (ERCP), endoscopic ultra-
sound (EUS), percutaneous transhepatic cholangiography
(PTC) or even bile duct biopsy and fine needle aspiration
[8]. MRI, or more specifically, magnetic resonance chol-
angiopancreatography (MRCP), has now become the
414
favorite follow up test, replacing the former golden
standard ERCP.
MRCP images are taken to cover the biliary system area,
as shown in Fig. 1 [10]. Some information relating to the
MRCP protocol, examination and other relevant informa-
tion are available online [11]. A sample MRCP image for a
patient with cholangiocarcinoma is given in Fig. 2. Auto-
mated computer-aided diagnosis (CAD) systems for MRCP
are very rare, more so for cholangiocarcinoma detection.
This is due to the high complexity in accurately identifying
such diseases in a relatively noisy image. Often, the tumor
is not seen clearly in a single MRCP image (refer to the
labeled area in Fig. 2) as being a solid, it just appears as a
low intensity area in the T2-weighted MRCP image (cf.
high intensity indicates liquid, thus the biliary structures are
represented by the high intensities). Furthermore, back-
ground tissue and organs, orientation and overlapping
structures, noise, body fat, signal strength, artifacts, partial
volume effect and parameter settings are some of the factors
affecting the quality of image acquired. They cause the
large inter-patients and even intra-patient image variations
in grayscale intensity, which makes computer-aided auto-
mated processing very difficult.
Methodology
As the objective of this work was to develop a system to aid
medical practitioners in their diagnosis, it was decided that
as far as possible, radiological diagnosis based ideas should
be followed in the methodology. This is further a necessity
as the nature of the images makes it a challenge for any
single conventional algorithm to perform well. In addition,
it was realized that a successful system would require
multiple stages, and some amount of dynamic decision
Fig. 1 Location of the biliary tract in the abdomen
J Med Syst (2009) 33:413–421
bile ducts
cholangio
carcinoma
Fig. 2 MRCP image containing cholangiocarcinoma
making capabilities would required in at least some of the
stages. The radiological diagnosis knowledge and experi-
ence, as well as the medical data in terms of MRCP images
and confirmed diagnosis, were obtained through collabora-
tion with a hospital that was the liver referral center for
liver diseases. From the information gained and consid-
erations above, the algorithm shown in Fig. 3 was
developed. The implementation details are given below.
Image preparation
The intensities between MRCP images can differ signifi-
cantly, even between images of the same patient. This is
further compounded by the fact that the images are usually
acquired at a 12 bits per pixel (bpp) resolution, providing a
large intensity distribution range. Normalization is required
but fixed values of intensities cannot be used. When a
radiologist analyses MRCP images with different intensity
distributions (e.g. change in brightness and/or contrast), it is
the relative information within the intensities are usually
used instead of the absolute values.
In a study [12], it was realized that the general histogram
distribution pattern in MRCP images were roughly similar
even though their intensities were not. From the study, it
was found that the first global maximum peak (P1) in an
Output: Optimized Decision
[Tumor /
Trained MLP
Normal]
Fig. 3 Proposed cholangiocarcinoma detection system
J Med Syst (2009) 33:413–421
MRCP image corresponded to the intensity of most of the
air in the body (e.g. those in the bowels). The second peak
(P2) corresponded to the intensity of the soft tissues, which
also overlaps with the lower intensities of the biliary
structure as the intensities are not uniform throughout the
biliary system. The minimum point between the two peaks
(i.e. the trough, X) could be used as a threshold to remove
the air. The remainder of the histogram then corresponds to
various parts of the biliary tract, along with some residue
noise, artifacts, background and other structures that were
not successfully thresholded. The image may have bright
spots (very high intensities), which make the rest of the
image appear dull in comparison. A point Y (about a third
down the slope) and Z (about two-thirds from Y to the
maximum intensity) are estimated as the beginning and
ending intensities of most of the biliary structures.
Identifying the intensities in a particular MRCP image
that correspond to points P1, P2, X, Y and Z, the image can
be dynamically normalized and enhanced. Figure 4 shows
the histograms before and after the image preparation stage.
The histogram is thresholded at X, and truncated at Z (to
minimize the effect of very high intensities). The remaining
parts are shifted left such that X′ is on the y-axis, stretched
Fig. 4 MRCP histograms for image preparation stage
415
between Y and Z such that Z′ is at the original Z intensity.
Finally, the whole resulting histogram is scaled down to
8 bpp (intensities 0–255) from the original 12 bpp, to make
them suitable for analysis on a normal computer monitor.
The area between X and Y is not stretched by design as it
contains most of the background tissue and should not be
enhanced. However, it cannot be eliminated as parts of the
biliary structures also fall within that intensity range and are
required in the following stages.
Segmentation
The previous stage employed intensity frequency but did
not take the location information into consideration, a
necessary criteria to segment an image into meaningful
homogenous sections. Although infamous for its over-
segmentation problem where images tend to be broken up
into too regions, the watershed algorithm [13] does produce
relatively accurate segmentation boundaries. Taking the
intensity gradient image, the watershed algorithm produces
segment boundaries at the steepest points of the gradient,
thus separating the image into semi-homogeneous seg-
ments. To overcome the over-segmentation problem,
caused by minor fluctuations in the intensity of the
background as well as throughout the biliary system, the
spurious and very small regions need to be consolidated
into the larger regions.
Before merging, to counter inter-pixel variations within a
segment, the intensity of all pixels within the segment is set
to the average intensity of that segment. This allows the
segments (instead of pixels) to be treated as the minimal
unit in the remaining stages. The segment merging is then
performed. Mistakes in selecting the appropriate criteria for
merging could lead to premature elimination of parts of the
biliary structure, with dire consequences in the tumor
detection stage. In practice, using manually labeled test
set of 256×256 MRCP images obtained from the collabo-
rating medical institution, it was found that the segments
that generally fit the criteria for merging were those that
were less than 30 pixels in size (i.e. very small), and only
merged to other segments with intensities within a
difference of 10% from their own (i.e. similarity criterion).
Biliary structure identification
In the ideal scenario, the images should now be well
segmented and the biliary system should be easily
identified. In a realistic environment, this is rarely the case,
especially when dealing with biological objects. The nature
of the MRCP images, partial volume problems and
intensity inconsistencies, make the biliary structure identi-
fication process more complicated. Several strategies
including scale-space analysis and directional wavelets
416
known as contourlets also failed to produce sufficiently
accurate results. Finally, an adaptation of the conventional
region growing at the segment level, called segment
growing, was found to produce good results [14].
Growing strategies require the selection of initial seeds
to start the process. In an automatic system, this has to also
happen automatically. The criteria for the selection had to
be identified. Next, to proceed with the growing, criteria for
selecting appropriate segments to grow into have to be
identified. These criteria were identified through analysis
of a number of test MRCP images, which were prepared
and segmented using the strategies discussed in “Image
preparation” and “Segmentation” sections. The biliary
structures on the images were then manually labeled. Statistics
on the most significant parts of the structures in the images
were used as the criteria for seeds, whilst the statistics of the
other labeled areas, compared with the statistics of the
unlabeled background, were used as the growing criteria.
The analysis [15], found that only average intensity,
location and size influenced the correct selection of seed
segments. Seeds were found to be segments within the top
20% highest intensity, close to the middle of the image
(bright segments at the fringes of the image tend to be
bright spot noise or artifacts) and not too small (minimum
10 pixels in size). Multiple seeds are required to identify
disjoint biliary structures, so all seeds meeting the criteria
were grown in descending order. The growing criteria were
segments had to be of similar high average intensities (i.e.
intensity greater than 200 with an intensity difference of
less than 20%), small segment size (less than 100 pixels as
larger segments were usually background) and sharing a
Fig. 5 Result of image prepara-
tion (b), segmentation (c), and
biliary structure detection (d)
J Med Syst (2009) 33:413–421
small border (maximum 30 pixels; large border often
background).
Figure 5 [14] shows the effect of the biliary structure
identification on a sample 256×256 50 mm thick slab
MRCP image, along with the results of the previous
stages. It is observed that application of the various
stages of the proposed scheme successfully eliminated
most of the background in the image, although not all as
some background shared very similar characteristics with
the biliary structures. The resulting image also provides
better 3D information approximation, with the lower
intensities representing parts further away (or deeper in
the abdomen).
Tumor detection
Diseases of the bile ducts generally cause distention or
swelling of the ducts due to the blockage of the flow of bile
in the ducts and cholangiocarcinoma is no exception.
Although the tumor itself is not seen clearly in a T2 MRCP
image, where solid objects are represented by low intensi-
ties, the appearance of disjoint distended biliary ducts
provides clue to their presence. This is also the case in
radiological diagnosis, in which the suspicion is then
confirmed by examining the appropriate series of MRCP
images, other series (different orientations and sequences)
and possibly ordering more tests and/or images, if neces-
sary. In the interest of lower processing requirements, time
and complexity, the developed algorithm is meant for
preliminary detection of cholangiocarcinoma by examining
only a single 2D MRCP image.
J Med Syst (2009) 33:413–421
For a bile duct to be considered as swollen, it should be
sufficiently large. The common bile duct (CBD) is usually
less than 6 mm thick in a healthy person, so biliary
structures beyond that may be considered distended.
Ideally, the duct thickness (or a simple approximation of
area divided by length) would be used as a measure of
distension, but it was found that in many images, the
distension of the biliary structures in the case or cholangio-
carcinoma were not significant enough (see Fig. 2) to be
incorporated as a criteria. Instead, simplifying, it was
decided that the area would be used and structures less
than 50 pixels in size were not considered as a significant
biliary structure and would not participate in the detection
process. In the case of Fig. 5d, the small patch of
background tissue above the biliary structure could be
eliminated this way.
Normally, a tumor would not be very large and thus, the
disjoint areas would be in moderately close proximity.
However, although a maximum distance may be prescribed,
the actual distance covers a large range dependent on the
3D orientation of the structures when captured in the 2D
plane. For example, two disjoint sections as seen in the
coronal (front) view may appear closer and even joined in
the sagittal (side) view, with decreasing distance between
the disjoint sections in the rotational angles between the
coronal and sagittal views. The case of joined structures is
beyond the scope of this work as even a trained radiologist
would not be confident of a cholangiocarcinoma diagnosis
by examining such an image. The developed system will be
limited to excluding such images as normal (thus, unhealthy)
in the case of distention of the bile ducts are present. All the
tumors in the test cases analyzed were less than 50 pixels
wide, and this is used as an additional maximum distance
criterion to eliminate potential noise, artifacts or even other
organs from influencing the tumor detection.
Decision-making
The assumption of the size of disjoint structures can vary
depending on the orientation of the acquired image, and is
greatly influenced by the size of the focus area in the
image. To add to the complication, not all disjoint structures
necessarily indicate the presence of a tumor. As such, there
may be more subtle criteria that need to be used in the
decision-making, which has not been identified as yet.
ANN have been used successfully in a large number of
medical system (and a variety of other fields), allowing
automatic learning, error tolerance and generalization in
handling unseen data. As such, an ANN is used in the final
stage of the proposed system to the boost decision-making
performance.
There are a large number of ANN topologies available,
with different abilities to handle specific problems and data
417
[16]. Although there have been critiques by some [17], the
feedforward multi-layer perceptron (MLP) is still arguably
the most popular ANN with a long and successful track
record. Easily set up, with the ability to be instructed on
expected outcomes through supervised training, the MLP in
Fig. 6 was used in the decision-making stage. In the case of
multi-layered architectures, the number of hidden layers
and corresponding hidden neurons relate to the complexity
of the problem. Decision-making information is usually
stored as weights (coefficients) and biases affecting the
neurons, set (“learned”) during a training phase.
The quality of training is determined by a learning rate
and learning goal (i.e. error rate). In many cases, over-
training may also be prevented by setting a maximum
number of iterations (epochs) limit. Over-training an ANN
makes it rigid and hinders its generalization ability to
handle new data. A good distribution of training data,
covering typical and atypical scenarios, improves the
ANN’s generalization ability.
The input layer consists of a neuron for each of the
features (f) presented to the MLP, namely, average segment
intensity and sizes of the individual biliary structures
detected. Therefore, a minimum system would require four
input neurons as there should be at least two disjoint
structures. Additional sets of input neurons could be used if
more than two biliary segments are to be analyzed. Further
input neurons to incorporate parameters including distance
between the midpoint of the structures, thickness as well as
shape characteristics, could be added. As the priority was
low complexity and fast processing, only one hidden layer
was implemented. Ten hidden neurons were found to be
adequate to delineate the network and produce the best
performance. The output layer required only a single
neuron for the binary decision, i.e. ‘0’ indicating negative
tumor detection and ‘1’ indicating possible tumor detection.
The activation or transfer functions used in the neurons
were sigmoidal and linear, in the hidden and output layers,
respectively.
Input Hidden layer Output layer
layer [sigmoidal] [linear]
f1 Tumor /
not
tumor
fn
Fig. 6 MLP architecture used in the proposed system
418
Being a supervised network, a set of training data and
confirmed diagnosis results were required to train (i.e.
configure the weights and biases) the MLP. The training set
consisted of approximately 20% of the 593 MRCP images
obtained from Selayang Hospital, Malaysia. The training
images were selected pseudo-randomly to include a
distribution comprising proportionately of images consider
as normal (healthy), diagnosed as cholangiocarcinoma, and
diagnosed with other biliary diseases. As all the images
were those of actual patients, the radiological diagnoses
were available and further confirmed by a senior radiolo-
gist. The standard backpropagation learning algorithm was
used to configure the coefficients (weights and biases) of
each neuron during training. Due the large variations within
the images, the error rate was high. Thus, the stopping
criterion for the training was the maximum iterations,
where it was set to 30,000 epochs as there was no
convergence of the achieved error beyond that. Based on
past experience, a learning rate of 0.15 was used in this
work as the system was not sensitive to minor changes in
the learning rate.
Results
Cholangiocarcinoma may be present in different parts of the
bile ducts. As with any biological system, even the images
of healthy patients can differ considerably. The choice of
parameters, intensity thresholds, values of the segment sizes
for merging and growing, and biliary segment size for
tumor detection were varied and tested. The configuration
and results presented in this paper are for the setup that
performed the best for the acquired MRCP images.
Two types of tests were conducted. The first was a proof
of concept, where the algorithm was evaluated on its ability
to differentiate images containing tumor from those of
healthy patients. Table 1 shows the details of this test, using
55 test images. The high accuracy level indicates that the
proposed system methodology is valid and the system is
indeed capable to differentiating cholangiocarcinoma from
normal images. The inaccuracies where the tumor was not
Table 1 Results of the tumor detection validation testing
Normal Cholangiocarcinoma
Medically diagnosed 27 28
Correctly detected 26 26
Accuracy 96.3% 92.8%
Error 1 2
Description Large biliary 1 normal, 1 other
structures dilation
J Med Syst (2009) 33:413–421
detected were: due to lack of distention which made the
structures appear to be that of a healthy patient, and
orientation that made the distended biliary structures appear
joint indicating other biliary diseases. The inaccuracy for
the normal image that was due to the biliary structure
appearing unusually large (possibly zoomed during image
acquisition) and merely indicated not normal (i.e. diseased)
as opposed to tumor. Such cases are unavoidable through a
simple detection system using a single image, given the
large inter-patient variations as well as different parameter
settings and practices by a number of radiographers who
acquired the images over the years.
The second test was to undertake a more realistic
simulation applicable in the clinical environment. The
MRCP images with other biliary diseases present (such as
stones, cyst and non-tumor swelling in the bile duct) were
included in the testing. Furthermore, patients may suffer
from more than one biliary affliction (which may have
similar visual characteristics), thus some test images were
influenced by more than one biliary disease. This multi-
affliction test set of 593 images contained 248 healthy
(normal), 61 tumor and the remainder images with other
biliary diseases such as stones, cyst, residual postsurgery
dilation, Klatskin’s disease, Caroli’s disease, strictures, and
obstruction from other liver diseases.
Table 2 shows the results of this robust testing. The
evaluation criterion used, overall accuracy, is calculated as
the percentage of the sum of the true positives and true
negatives over the total test set, taking into account how
well the system performed in detecting the correct disease
as well as in rejecting the wrong diagnosis. For lack of
much benchmark literature, a comparison with only one
prior automated tumor detection system using MRCP [18]
is given in the table. From the results, it shows that the
proposed MLP system’s performance was superior for the
detection of both cholangiocarcinoma (88.03%) as well as
healthy images (83.64%), even if only by a small margin.
In the case of identifying the presence of identifying biliary
diseases in general, the MLP system performed much
better. It achieved a detection rate of 90% for non-
cholangiocarcinoma diseases, an increase of 10% as
compared to the results obtained by the reference system.
The specificity (non-tumor images correctly identified as
not containing tumor) results was also calculated. Once
again, the MLP system dominated achieving 94.70% as
compared to 89.85% of the previous system.
Figure 7 presents examples of the images which were
detected incorrectly by the proposed system. The images in
Fig. 7a and Fig. 2 were acquired in the same MRCP
examination. However, although the biliary structures in
Fig. 2 are clearly disjoint, the orientation in Fig. 7a makes
them look joint and thus makes it impossible for chol-
angiocarcinoma to be diagnosed even by the radiologist.
J Med Syst (2009) 33:413–421
Table 2 Results of robust
testing using multi-disease
clinical MRCP images
Test images
Overall accuracy
MLP system
Reference system [18]
Fig. 7b suffers from poor contrast and low signal strength,
causing the boundaries and inhomogeneous intensities of
the liver tissue and biliary structures to overlap indistin-
guishably. In both examples, the inaccuracies of the
proposed system were unavoidable due to the poor quality
and non-standardization of the acquisition process, which
were beyond the control of this work.
Timing performance is not presented in the results as the
system was developed in IDL (Interactive Data Language)
for fast development with the available libraries, and not for
fast performance. However, without any significant optimi-
zation and running on a standard Pentium IV desktop
personal computer with the IDL virtual machine run-time
environment, the proposed system took less than a minute
to process an MRCP image. The timing performance could
be dramatically enhanced if the system is developed to be
compiled into machine code and executed on a high
performance workstation.
Discussion
The results obtained show good overall accuracy for the
MLP system, and an overall improvement over the reference
system. The accuracy achieved is impressive, taking into
account that even visual diagnosis using individual 2D
MRCP images is very difficult and impossible in certain
cases. This is further compounded by the fact the robust test
data consisted of a large number of images with other bile
duct diseases. Even experienced specialists are reluctant to
Fig. 7 Example problem cases
(a) Orientation problem (Fig. 2 patient)
419
Normal Cholangiocarcinoma Others TOTAL
248 61 284 593
83.64% 88.03% 90.14% –
76.90% 86.17% 80.99% –
make a confident diagnosis by just examining a single
MRCP image, as they usually use series of images in the
examination. Overcoming such problems requires strict
adherence not just to image acquisition standards and
uniformity, but also with a realization of the limitations of
the system as it is only presented a single image. As such,
some selection criteria for the images to be presented to the
system have to be set, including standardized acquisition
parameters and good orientation.
The proposed system is merely a tool to aid in the
diagnosis of cholangiocarcinoma, the responsibility of which
resides with the medical practitioner. Issues on the misuse of
ANN technology in oncology have been highlighted in the
past [19] and the user should be mindful of the purpose and
usage of the tool. The proposed system is a simulation
prototype, requiring further in-depth study and testing
before it could be used in the clinical environment.
The proposed system was developed with the handicap of
only analyzing a single image in the detection stage. This
restriction was placed in the effort of producing quick
preliminary detection with minimal resource requirements
and complexity. A typical MRCP examination consists of a
number of series of images, composed of locator slices, axial
T2 sequence series, axial in-phase (fat saturation) series,
MRCP thin slice series and MRCP thick slab images [11],
and additional series as deemed necessary by the radiolo-
gist. Multiple images could result in improved detection,
especially in cases of non-optimal orientation. Such efforts
could be taken as future work in developing a more
comprehensive diagnosis system. Additional information
(b) High background intensity
420
would be required from the image headers, 3D volume data
may be approximated and an interactive interface could be
provided for an improved user-friendly application.
The proposed MLP system, as with most applications,
treats the ANN as a “black box”. Although training is
controlled by the data, expected outcomes, error goal etc.,
the “rules” set within the MLP, by way of the actual
coefficient values of the weights and biases, are left
unknown. Furthermore, training is usually started by
assigning random values to these coefficients and allowing
them to converge. Thus, the exact setup may not be
repeatable due to the random values. Recently, there have
been efforts towards rule extraction from trained ANN.
These may be used as a basis to understand the necessary
criteria for the decision-making and allow for further
optimization of the network.
Some of the possible improvements to the system
include upgrading the MLP into a hybrid network. Such
networks are becoming popular by combining specifically
selected architectures and algorithms that allow their
strengths to be combined into a single network. This in
turn allows for the optimized network the ability to better
handle the type of data and complexity required, improving
overall performance. In addition to the typical ANN
architectures and fuzzy networks, other high performance
systems such as genetic algorithms (mimics population
growth), particle swarm optimization (mimics social be-
havior of birds flocking), ant colony algorithm and many
more, could be incorporated in a hybrid network. These
state-of-the-art algorithms often provide solutions to diffi-
cult problems, in this case, maybe even overcoming the
other complications present in the poorly acquired MRCP
images and diverse anatomical shapes.
Conclusion
Cancer is becoming an increasingly common fatal illness.
Cholangiocarcinoma or cancer of the bile ducts is also on the
increase and is often detected once at a late stage. MRCP
images are the defacto non-invasive diagnosis mechanism
for diagnosis affecting the bile ducts. This paper proposes a
system using MLP to perform automated preliminary
detection of cholangiocarcinoma in 2D MRCP images.
The multi-stage system consists of image enhancement,
segmentation, biliary structure identification and MLP
decision making for detection. The stages are inspired by
the clinical and radiological diagnosis conducted for the
disease, while the algorithms used at the different stages
tend to be inspired by natural systems as well. The test
results, both controlled (detection of cholangiocarcinoma
from a test set with normal healthy images) and robust
J Med Syst (2009) 33:413–421
(with a variety of biliary diseases), produced very good
accuracy in excess of 80% in all categories. The system
used in the simulation was a simple prototype. A discussion
of the issues and recommendations for improving the
system for clinical use is also given towards the end of
the paper.
Acknowledgment This paper is supported by the Soongsil
University Research Fund.
References
1. Meyer, C. R., Park, H., Balter, J. M., and Bland, P. H., Method for
quantifying volumetric lesion change in interval liver CT
examinations. Med. Imaging. 22(6):776–781, 2003.
2. Cobzas, D., Birkbeck, N., Schmidt, M., Jagersand, M., and
Murtha, A., 3D Variational brain tumor segmentation using a
high dimensional feature set. Workshop on Mathematical Methods
in Biomedical Image Analysis. 1–8, 2007.
3. Degenhard, A., Tanner, C., Hayes, C., Hawkes, D. J., Leach, M.
O., and Study t.U.M.B.S., Comparison between radiological and
artificial neural network diagnosis in clinical screening. Physiol.
Meas. 23(4):727–739, 2002.
4. Marchevsky, A. M., Tsou, J. A., and Laird-Offringa, I. A.,
Classification of individual lung cancer cell lines based on
DNA methylation markers use of linear discriminant analysis
and artificial neural networks. J. Mol. Diagnostics. 6(1):28–36,
2004.
5. Ippolito, A. M., Laurentiis, M. D., Rosa, G. L. L., Eleuteri, A.,
Tagliaferri, R., Placido, S. D. et al, Immunostaining for Met/HGF
receptor may be useful to identify malignancies in thyroid lesions
classified suspicious at fine-needle aspiration biopsy. Thyroid. 14
(12):1065–1071, 2004.
6. Patel, T., Worldwide trends in mortality from biliary tract
malignancies. BMC Cancer. 2:10, 2002, PMID 11991810.
7. Yoon, J-H., and Gores, G. J., Diagnosis, staging, and treatment of
cholangiocarcinoma. Curr. Treatm. Opt. Gastroenterol. 6:105–
112, 2003.
8. UCSF Medical Center. Cholangiocarcinoma. Retrieved May 24
2008, from http://www.ucsfhealth.org/adult/medical_services/
gastro/cholangiocarcinoma/conditions/cholang/signs.html, 2008.
9. Chari, R. S., Lowe, R. C., Afdhal, N. H., and Anderson, C.,
Clinical manifestations and diagnosis of cholangiocarcinoma.
UpToDate. Retrieved May 24 2008, from http://www.uptodate.
com/patients/content/topic.do?topicKey=gicancer/23806, 2008.
10. NIDDK—National Institute of Diabetes and Digestive and Kidney
Diseases. Digestive diseases dictionary A–D: biliary track. National
Digestive Diseases Information Clearinghouse (NDDIC). Retrieved
February 21 2008, from http://digestive.niddk.nih.gov/ddiseases/
pubs/dictionary/pages/a-d.htm, 2008.
11. Prince, M. R. MRCP Protocol. Retrieved 2008 May 25,
from http://www.mrprotocols.com/MRI/Abdomen/MRCP_Dr.
P_Protocol.htm, 2000.
12. Robinson, K. Efficient pre-segmentation. PhD thesis, Dublin City
University, from http://www.eeng.dcu.ie/∼robinsok/pdfs/Robinson
PhDThesis2005.pdf, 2005.
13. Vincent, L., and Soille, P., Watersheds in digital spaces: an
efficient algorithm based on immersion simulations. Trans-
actions on Pattern Analysis and Machine Intelligence. 13:583–
598, 1991.
J Med Syst (2009) 33:413–421
14. Logeswaran, R., Neural networks aided stone detection in thick slab
MRCP images. Med. Biol. Eng. Comput. 44(8):711–719, 2006.
15. Logeswaran, R., Scale-space segment growing for hierarchical
detection of biliary tree structure. International Journal of Wave-
lets, Multiresolution and Information Processing. 3(1):125–140,
2005.
16. Ripley, B. D., Statistical ideas for selecting network architectures.
Neural Networks: Artificial Intelligence and Industrial Applica-
tion. Springer, 183–190, 1995.
421
17. Roy, A., Artificial neural networks—a science in trouble. Special
Interest Group on Knowledge Discovery and Data Mining. 1:33–
38, 2000.
18. Logeswaran, R., and Eswaran, C., Discontinuous region growing
scheme for preliminary detection of tumor in MRCP images. J.
Med. Syst. 30(4):317–324, 2006.
19. Schwarzer, G., Vach, W., and Schumacher, M., On the misuses of
artificial neural networks for prognostic and diagnostic classifica-
tion in oncology. Stat. Med. 19:541–561, 2000.