International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

1
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
2
I. A NEW WAY TO SHARE PRACTICAL ARTIFICIAL
INTELLIGENCE AND INTERACTIVE MULTIMEDIA
KNOWLEDGE
HIS publication contains all research activities
performed by students during 2007/2008 academical
year. The editor's council has been reviewing and selecting
several outstanding research contributions related with the
following knowledge areas:
T
− Medical diagnosis: currently interactive multimedia
and artificial intelligence techniques suggest the
possibility to offer new approaches focused to discover
and obtain high precision diagnosis, merging technical
and medical knowledge. The experimental research
works show the high-potential provided by software
prototypes based on visual perception, fuzzy logic and
bayesian theories.
−
Semantic metadata: semantic web theories and formal
languages has been growing during last years due to
effort made by a high number of researchers. Besides, it
is important to point out an special area “Semantic
Metadata” with the main aim of provides several
benefits on metadata management. Geographical
Information Systems, Documental management systems
among others can be better off by the use of ontologies
(instead of natural language) as structural metadata
formal knowledge. This section includes the most
important contributions.
− Nature conservancy: we can help protect the world's
most beautiful and diverse places developing technology
that preserve threatened habitats around the world.
Artificial intelligence, robotics and real-time monitors
are essential tools used into our laboratory to build
prototypes which seek how we can design and deploy
real projects on demarcated environments.
−
Intelligence perception: this section includes two big
areas, intelligence indoor positioning systems and
intelligence visual identify systems. Learning algorithms
has been used to create adaptive positioning and
identify systems.
This journal has been created with the aim of publish,
share and bring together our experiences and work
performed during one academical year. However we hope
receive more contributions from IA and multimedia
researchers, in order to open and extend future next issues of
the journal. Finally, we would like to thank all unselfish
contributions made by the members of editor's council:
Dr. Oscar San Juan (Oviedo University).
Ing. Gloria García (Pontifical University of Salamanca).
Dr. Juan Luis Chulilla.
Lca. Pilar Azagra Albericio.
Lca. Raquel Ureña Joyanes.
II.OPENLAB PROJECT
OpenLab is an open laboratory of ideas and projects
related to application development and research using
artificial intelligence and interactive multimedia techniques.
It has been created from new practical laboratory approach
applied on artificial intelligent courseware.
The legacy education model focused on a teacher changes
around the nature’s education in the world of computer
science. Research produces new theories that make it
necessary adapting teaching to some more modern and useful
concepts. The student doesn’t learning the theory properly,
specially in our night-shifts groups (worker-students). It isn’t
enough to go into the classroom and teaching about
equations, reasoning and proofs in the last years of a degree.
The student needs practical applications that allows solving
in a more efficient way the real-life problems.
In certain cases it is not possible propose a set of practices
with direct application on coursework, since it is necessary
too much previous knowledge. However in artificial
intelligence and cognitive sciences it is possible, given the
wide range of existing tools and the student knowledge level.
The OpenLab project inspired by bologna process
guidelines, attempts to launch a set of artificial intelligence
projects every year. Each project is developed by a small
team (3-5 members) with an specific technique: bayesian
networks, fuzzy logic, neural networks. decision trees,
among others. Due to time constraints, each team is
specialized in a concrete field and exposes the results at the
end of the year.
After three successful years we conclude, that group
working combined with good personal tutor sessions allows
obtain high-quality projects that can be reused each year.
Thus if an student wishes to develop a game using XNA
technology and some artificial intelligence technique
(heuristic search, planning, etcetera) in following years
another student will be able to use the project finished. We
hope to continue improving our interactive multimedia and
artificial intelligence laboratory in order that students can
discover the practical application of artificial intelligence
techniques.
EDITOR'S NOTE
Editors Note, Jesús Soto Carrión and Elisa García Gordo
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

3
TABLE OF CONTENTS
Social maturity of WWW and AI feedback: opportunities for an additional human revolution................................4
ADT-3D Tumor Detection Assistant in 3D ............................................................................................................6
Computer-aided diagnosis of pancreatic and lung cancer ......................................................................................16
EvoWild: a demo-simulator about wild life ...........................................................................................................25
Intelligent Garbage Classifier ...............................................................................................................................31
Software fires detection and extinction for forest...................................................................................................37
Iris recognition using the JAVAVis Library .........................................................................................................43
Fuzzy Logic Indoor Positioning System ..............................................................................................................49
Nintendo DS Programming: DS MUS ................................................................................................................55
B-Network - Sport forecast .................................................................................................................................60
Imusic..................................................................................................................................................................66
General purpose MDA tools.................................................................................................................................72
Open Access Journal
ISSN: 1989-1660
CopyRight Notice
Copyright © 2008 ImaI. This work is licensed under a Creative Commons Attribution 3.0 unported License.
Permissions to make digital or hard copies of part or all of this work, share, link, distribute, remix, tweak, and
build upon ImaI research works, as long as users or entities credit ImaI authors for the original creation. Request
permission for any other issue from jesus.soto@imai-software.com
All code published by ImaI Journal, ImaI-OpenLab and ImaI-Moodle platform is licensed according to the General
Public License (GPL).
http://creativecommons.org/licenses/by/3.0/
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
4
In these days, Internet Adoption Curve in developed
countries belongs to an interval between early majority and
late majority. There is already a significant population
profile which can be defined as 'digital natives', more or less
isolated from their 'digital immigrants' thanks to the abyss
defined by the decisive integration of Internet in digital
natives' daily lifes. This situation is actually significant if one
takes in mind that digital persona of digital natives is much
more than a mirror of, let's say, actual persona and both
belongs to a new, multifaced entity still not well understood.
As Adoption Curve has progressed, the complete range of
social software has been under a harsh darwinist pressure.
The most important selection factor has been the social
success of each social software via personal, colective and
professional impact. Some social software has not resisted
'Net pace (i.e., gopher or archie); other types have achieved
and maintained the status of killer application; finally, new
types of social software has emerged in the last fifteen years.
From blogs and their metarreality (blogosphere/The
Conversation), wikis and other collaborative spaces, to the
last big social network, multidimensional products
(Facebook, MySpace, etc.). Second life and another previous
attempts (like VRML, X3D, etc.) have shown clearly that
main ingredient of success is not cutting edge technologies
(beyond the bare minimums for make things work, very close
to commodities in these days), but uses and images that the
users produces about each technology: what they can do with
each kind of social software by themselves alone or in
company of other peers. Twitter, del.icio.us or digg are just
three examples choosen from a huge pool which proof that
simple and clear interfaces are not obstacles toward success
at all.
Social evolution of Internet, its way to its maturity after teen
crisis (web 1.0), has carried with it specific successes utterly
unforeseables both because of their magnitude and potential.
One of the most clear examples would be Wikipedia. A
reference space generated freely (discharging maintenance
costs), feeded by huge numbers of amateurs/microexperts and
totally usable by any individual with connection to Internet.
Beyond casuistics in Web 2.0, true essence of social impact
of a social mature Internet would be the unending amount of
ways in which knowledge, communication and activism can
be shared and builded: all that can be done with and for other
people. Digital citizen, both native or immigrant, can make
use of possibilities of personal, professional and collective
development that just weren't foreseen, beyond the direst
dream of Science Fiction, wider and deeper than River
World or Foundation. It is possible that, in fact, this
circumstance has take part of science-fiction decadence.
A priori, there is no clear signs of a decceleration of the
change pace. Indeed, since web 2.0 has gained popularity as
a concept (more or less, half a year after O'Reilly conferences
when term was conceived), there has been no interruption of
announcement of new and then quickly popular new social
software, with a significant public and media impact.
However, there is a fact which is determining clearly the
evolution of all the social software: information overload.
This net of human for humans, this eclectic set of services
and documents has reached a point in its path in which
Internet power users adapt themselves as good as they can to
a quantity of information received at a pace without
precedents at all. There are clearer and clearer signs of that
there is no intelectual and even no biological tools for
copying with that magnitudes of information. Search engines
are just too good and efficient, and social software types are
just too atractives for some population profiles.
The worst fact is not the human impossibility of overcome
information overload, but the effects of that situation. The
abused term of multitasking (applied to human beings) is
coming with a serious price, the difficulty of maintain deep
thought. Some of the benefits of social software are quickly
evolving into created necessities, and this new condition
contributes to explain the problem mentioned just before.
There is no way to assure if definition and transition to
some web 3.0 compromise is an urgent needing or not. If
finally Internet obtains a framework in which software tools
(agents) can operate with information at its level, this would
mean a decissive step forward.
This scenario could evolve into a state of continuous
positive feedback. Until now, interaction between different
social software has permitted an evolution which pace is only
surpassed by its consequences. As I affirmed lines above,
unavoidably out of control growth of possibilities of access to
information has reach a state of use of Internet with clearly
negative consequences in differents aspects of daily life and
proffesional activity. If finally emerges a way of improve
user's capability of cope with The Problem of information
overload, it will have consequences both unforeseeable and
deep: if the user can cope adequatedly with information
overload and gains the possibility of access and operate with
quantities of information even bigger, it would produce
decissive effects both to individuals and communities.
If some part of natural intelligence is based in the
capatibility of capture and work with information and
produce new results as new information, an improved access
to Internet information would improve the intellectual
Dr. Juan Luis Chulilla and Lic. Pilar Azagra Albericio
Social maturity of WWW and AI feedback:
opportunities for an additional human revolution
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

5
capabilities of the population. If some of the ways than AI
can offer for overcome information overload obtain certain
success, abyss between digital native, digital immigrant and
digital refugee would by a tiny step compared with the
(nasty) future that waits for those who cannot obtain benefits
from this developments.
This possible future is needing urgently solid analysis in
order of potentiate the possitive benefits and minimize the
negative ones. Social disadjustments can dwarf every other
previous generational shock. Intellectual impact can reach an
state of dystopia clearly comparable with "Logan's run" but
without the imaginery of Carroussel: all those people that
wouldn't be able to adapt and obtain the new condition of
"augmented human being" would be too much close to be
considered as living fossiles, parts of a past that still hasn't
learned that future is here right now and for stay.
As a conclusion, interaction between AI technologies and
the Web is close to be in a critical moment: every time is
clearer that AI is needed for overcome web 2.0 information
overload. However, if this combination obtain a remarkable
success, it can produce deeper social effects (positives and
negatives) by far compared with any other previous
technologies, keeping in mind that they would affect the core
of cognitive capabilities of human being. As digital divide of
today can be just a tiny step compared with the abysm
derived from AI success, I do believe that reflection and
analysis of it is an urgent necessity, specially if it reach the
sphere of public debate. Maybe this time we will be able of
maintain Pandora's box closed.
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
6
Abstract — The present document describes ADT-3D
(Three-Dimensional Tumor Detector Assistant), a prototype
application developed to assist doctors diagnose, detect and
locate tumors in the brain by using CT scan. The reader may
find on this document an introduction to tumor detection;
ADT-3D main goals; development details; description of the
product; motivation for its development; result’s study; and
areas of applicability.
Key Words — ADT-3D, Briain tumor, CT Scan, image
processing, Jaimes’ algorithm, Kmeans, Matlab, segmentation,
tumor, Visual Intelligence.
I.INTRODUCTION
A.A piece of history on Computerized Tomography
According to Tomografia Computerizada [1] on October
1st, 1971 the first brain scan was performed in a London
hospital. From then on, there has been a quick development
on the field. In 1973, the EMI Scanners, developed by
Electrical Musical Instruments Co. headed by Hounsfield,
expanded rapidly to the US and Europe. After which the
brain scan adopted different names such as CT Scanner
(Computerized Tomography) by Anglo-Saxons, TDM
(Tomodensiometry) or CAT Scan (Computerized axial
tomography); all of which were also known as X-ray
scanners.
Since Hounsfield begun the investigations, there have been
many changes on the discipline. The improvements try to
shorten the swap time and refine the image quality. Xavier
Vila [2], on its website, explains quickly, easily, and visually
the evolution of those changes. X. Vila states that there are
four different machinery ages.
At the early stages on tomography there was only one
detector, which rotated opposite to the X-ray tube, which is
how tomography was reached. Then, these machines were
upgraded to decrease rotation times and maximize the angle
of rotation from 180 degrees up to 360 degree.
B.Computerized Tomography
The CT technology was developed in order to carry out
head studies, but previous advances made possible the study
of the whole body. Nowadays, the CT can be used for any
region of the body.
The CT core are X-rays1. The main difference is how they
are used. For CT Scans the X-rays are applied over small
body slices. Therefore, instead of a body projection, CT
obtains a full resolution slice. Many manufactures have

1 X-rays: The oldest and best known of the medical imaging techniques. X-
rays, discovered by Röntgen more than 100 years ago, are high energy
radiation produced in a special type of lamp called an X-ray tube. X-rays are
also a form of electromagnetic radiation, but with a much shorter wavelength
and a higher frequency than visible light.[4]
developed different CT machines, building up the different
generations, from the first to the very last fourth scanner
generation. Generations differ from each others on the tube-
detector system. The most use CT scan generation is the 3rd
which is illustrated on the following picture.
Fig 1. “Schematic drawing of tube-detector system of 3rd generation CT
scanner. The X-ray tube emits a sharply collimated fan beam of X-rays which
passes the patient and reaches an array of detectors. Tube and detector
array rotate together around the patient; one exposure often comprises 360o
rotation.”[3]
The X-rays are emitted from the scanner as thin telescopic
fan shaped beam, making a regular angle with the patient
long side. The distance between the patient and the tube
must mach with the fan projection. The thickness of the fan
could be selected creating thicker or thinner slices, form one
to ten mm.
Each slice is divided into a constant number of volumes, so-
called voxels (Figure 2). Each single voxel value is obtained
by combining the values of several projections. Each
projection is done rotating the tube around the patient, for
each rotation the values are received by the array of detectors
which are on the opposite side the tube (Figure 1). When the
voxel is represented as a two-dimensional image, it becomes
a pixel. The pixel value, position, and size represent the
whole voxel data. As an image, the value will be a
represented in the grey scale, depending on the selected
brightness. The higher the attenuating voxel the less quantity
of rays have crossed through it, e.g. bone, which are brighter
in the grey scale, closer to white. On the other hand, the less
dense tissue, which more amount of X-rays cross through,
the darker the voxel representation is.
Fig 2. “The imaged slice of tissue divided into volume elements, voxels. The
attenuation in each voxel determines the brightness (shade of grey) of the
corresponding pixel in the final two-dimensional image.”[3]
The relationship of bones as white and air as black voxels is
obtained from the Hounsfield Scale. The Scale links each
voxel value with a grey shade scale. The default numbers are
-1,000 to represent air, 0 for water, and +3,000 for dense
Jaime García Castellot, and Jaime Lazcano Bello
ADT-3D Tumor Detection Assistant in 3D
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

7
elements, such as bones. The brain tissues average is between
60 and 120. The unit uses is called HU, Hounsfield Unit. In
fact, each manufacturer use its own scale but always referred
to Hounsfield Scale, see the picture below.
Fig 3. “Scale of Hounsfield units (HU). The approximate scale locations of
different substances are indicated. (By "tissue" is meant most fat-deficient
soft tissues and parenchymal organs.) Reference points: -1,000 HU for air, 0
HU for water.”[3]
CT scan resolution is far greater than older radiography
techniques, but moreover, the spatial resolution is lower. The
voxels thickness is as large as the slice. If the scanning set-
up asks for a normal study of the patient, the number of
slices to determinate the patient’s head will be lower than
forty. If the head is represented using computer programs
there will be many gaps between slices. The empty gaps are
fill by using the average value among the surrounding
voxels. The process is called interpolation, the added gaps
are in fact not real data. This is called the spatial loss of CT
scannig. This spatial loss could be resolved using thinner and
more slices. Consequently, longer CT examinations will be
needed.
A thin slice resolves the spatial loss, but also requires
higher amount of radiation to obtain a usable slice. The
process of thinner slices will multiply the radiation levels.
In CT scanning is crucial the relationship between quality,
time, radiation, and area covered.
C.Technological approach
Nowadays, the contribution of technology to any field is
fundamental but especially indispensable on medicine. It is
clear the close relationship between medicine and technology
since the manipulation of high complex medical equipments
is part of the technological progress made up to date.
A vertiginous advance on science has characterized last
fifty years, causing all technologies to advance rapidly. For
those whose professions involve the utilization of the
technologies, it’s being difficult to stay on top of the rapid
changes. Technological developments have positively altered
medicine, since it has enabled the better comprehension of
the multiple processes, which explain the symptoms of
several illnesses. Therefore, it has help on the understanding
of many symptoms and repercussions on the human body.
A classic classification of medical technologies could be:
diagnosis, preventive, therapy, rehabilitation, and
management.
D.Medical approach
A brain tumor is “a benign or malignant growth in the
brain. Primary brain tumors arise in brain tissue. Secondary
brain tumors are cancers that have spread to the brain
tissue (metastasized) from elsewhere in the body. Brain
tumors can and do occur at any age” [5].
A brain tumor is still very difficult to diagnose. According
to the University of California, San Francisco [6], the doctor
would first ask about family medical history and perform a
complete physical examination. Then, depending on the
results obtained, it may be possible that the doctor request a
CT scan, MRI, Brain Scan or other test. Unfortunately,
neither CT Scan nor MRI provides complete accurate data
about the presence of a tumor.
II.MAIN GOAL
By utilizing the group of slices obtained from a CT Scan of
the head, the ADT-3D is able to recognize a brain tumor. To
perform this task it will be necessary to generate a 3D
representation of the sections or slices. Once the three-
dimensional image has been generated the ADT-3D will
proceed to study the different regions using Artificial
Intelligence Vision methods and algorithms such as Kmeans
and Jaimes’ segmentation algorithms.
The result of all these processes should help to differentiate
visually the tumor from the rest of the brain mass. The visual
differentiation would make it easier to study the properties of
tumors.
III.DEVELOPMENT DETAILS
A.Functionality
The main goal of ADT-3D is to assist doctors during the
brain tumor detection and the possible diagnosis on patients
who undergo the CT Scan. ADT-3D fulfils the 3D analysis
of the possible tumors. The utilization of these three-
dimensional images enables tumor location on XYZ axis for
their later examination.
B.Development tools
1)MATLAB R2007b
MATLAB is a development environment based on
MATLAB language. “MATLAB® is a high-level language
and interactive environment that enables you to perform
computationally intensive tasks faster than with traditional
programming languages such as C, C++, and Fortran. ”
[7]. This language can be used in many different fields such
as signal or image processing; communications; test and
measurement, as well as other scientific fields of
investigation.
On this development environment it is possible to manage
code, files, and data. It includes many functions for statistics,
optimization, and numeric data integration and filtering.
There are also great amounts of functions that facilitate the
manipulation of bi-dimensional, and three-dimensional
images.
2)MATLAB Newsgroup [8]
Software developers may have to face problems to which
are not able to find a solution. The MATLAB Newsgroups is a
website tool offered by Mathworks. It allows software
developers to have a common point of contact to solve
programming problems. The Newsgroup has an extensive
database with hundreds of conversations, about the
programming language, concerning troubleshooting.
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
8
C.Data Manipulation
ADT-3D data input consist on a variable number of
DICOM images, which represent the brain mass in
consecutive slices. An example of this is shown on Figure 4.
The utilization of such technique on CT Scans allows first, to
store data about the patient’s head. Secondly, to manage a
3D body in a group of slices.
Fig 4. The figure represents one of the DICOM images that form the data
input of ADT-3D.
Then, each single image on the group of input images is
transform into a bi-dimensional matrix. The next step is to
create a three-dimensional matrix out of the group of 2D
matrixes. The created volume matrix is used for holding the
patient’s data, which is used as the numeric input to
ADT-3D.
The process of transformation from image into numeric
values is done through dicomread(), a function provided by
MATLAB environment. As long as the input argument of
the function is a DICOM image, then the output argument
will be an square matrix containing the voxel values. An
example of this transformation is seen on Figure 5.
Fig 5. It shows the numeric values corresponding to the given image when
MATLAB’s dicomread() is used.
During the process of converting images into figures, each
bi-dimensional matrix is added to a three-dimensional
matrix. The size of its third dimension will be equal to the
number of images conforming the brain scan. The outcome
it’s a 3D matrix, containing all data acquired during the CT
Scan, becoming the initial unit of work.
D.Segmentation
The segmentation is a computer technique “in computer
vision in which an image is divided into regions” [9]. Also
defined as: “the act of dividing or partitioning; separation
by the creation of a boundary that divides or keeps apart”
[10]. What about image segmentation? If the term
segmentation is applied to the word image, tImage
Segmentation is the act of dividing or partinioning an image
into regions defined by the creation of a threshold.
There are many different segmentation algorithms, each
one uses different characteristics from the studied object.
There are four common approaches used with intensity
images (e.g. Dicom images which use voxel values). The
four approaches are: threshold techniques, edge-based
methods, region-based techniques, and connectivity-
preserving relaxation methods.
ADT-3D uses Jaimes’ algorithm for segmentation. The
algorithm is explain in more detail below. Jaimes’ algorithm
uses Kmeans algorithm as part of the process, but since
Kmeans does not divide the data with the needed accuracy, it
applies other filters.
E.Jaimes’ Algorithm
1)Differentiate a tumor from brain mass
The main goal of ADT-3D is the identification of brain
tumors. How could it be done? The answer is on the
segmentation of the image. Image segmentation, as part of
Visual Intelligence, is responsible for discrimination on
different pictures. DICOM’s protocols differentiate tumors
from other brain masses by using the Hounsfield Unit Scale.
Based on Hounsfield Unit Scale, tumors score diverse values
than the rest of brain mass.
2)Kmeans
Kmeans algorithm divides any image into k number of
different classes. The k value can be determinate by using
other algorithms such as Chain-Map or Max-Min over the
image. Kmeans algorithm, applied to a DICOM data
collection, allows a maximum k value of five. It denotes that
Kmeans will give five different centroids, and each pixel
matrix will be related to one of these centroids.
Unfortunately, a classification using only five types of classes
is not enough to differentiate tumors from other brain
masses.
3)Variables
Mdata: It is a three-dimensional matrix that contains the
input data for both Kmeans segmentations.
Mclasses: Is the result of the application of Kmeans
to Mdata.
Mresult: The output data of the application of Jaimes’
algorithm is stored in this 3-D matrix.
4)Iteration I
Mdata is the input information to Jaimes’ algorithm, a
matrix containing the original numbers. Kmeans allows us to
classify raw data contained in Mdata based on a scale from 1
to k elements. As it was stated above k value is equal to 5.
The following figure is an example of this step:
Fig 6. The first table represents the data input values, where the four bottom-
right hand corner values represent a simulated brain tumor. Unfortunately, the
estimated outcome should be such that only those four values should have a
value of five. Moreover, there are other extra 5’s not related to a brain tumor.
Thus, the idea is to improve this first segmentation.
The application of Kmeans is executed using the following
code.
Mdat
a
Mclasses
kmeans
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

9
For nc=1:60
X = Mdata(:,:,nc);
k = 5;
[idx, partialClasses] = kmeans(X,k);
Mclasses(:,:,nc)= partialClasses;
End
Fig 7. Kmeans brings forward two values: idx and PartialClasses. The
former contains centroid values and the latter saves the data related to each
element’s classification.
5)The Adjustment
The previous step has achieved a decent approximation.
Nevertheless, it has confused certain brain masses also as
tumors. How could be possible to isolate the tumor? The
exponential function solves this problem by applying the


2
f x
x

function for each matrix pixel:










1
2
2
1
2
1
2
,
,
x
R
x
R
f x
x
x x
f x
f x

Þ
 ￎ


Þ 










1
2
1
2
:
92
130
92,130
38
92 ,
130
8464,16900
8436
Example
x
x
f
f





 

Fig 8. Two different numbers from the same class could be distanced by using
the exponential function .
The distance between numbers will be increase if the f
function is applied. It does not matter if the element belongs
to the same class. Once the function is applied, it is also
multiplied by the Mclasses number k. In this way, a
hypothetical second iteration will distinguish tumors from
brain mass values. The following function represents how the
adjustement is done for a selected group of data.
Fig 9. The Adjustment function is applied to Mdata after the First Iteration. Is
easy to appreciate how the tumor pixels have grown more than the brain mass
data pixels.
The Adjustment process is done by using the following
MATLAB code.
Mresult=Mclasses;
dim = size(Mclasses)
for i=1:dim(3)
for j=1: dim(2)
for k=1: dim(1)
Mresult(k,j,i)=
Mdata(k,j,i)*Mclasses(k,j,i)*Mdata(k,j,i);
end
end
end
Fig 10. The Mresult matrix is the weighting output. Mresult will be used on
the second Iteration.
6)Iteration II
Once the adjustment is done, the second Kmeans Iteration
may proceed. As well as on the first Iteration, the number of
centroids and classes will be five. The difference between
iterations is that class five is reserved for high tumor
probability pixels. The picture below shows how the second
Iteration leads the tumor to the fifth class.
Fig 11. As well as the First Iteration, the Second Iteration returns a new
Mclasses matrix. Now the accurary is higher. Anyway, it is needed the last step
to assure a precise result.
7)Smooth
The last algorithm’s step has a crucial effect on the final
results. Sometimes, depending on the working slice, other
non-tumor related five class pixels could be detected. For
sure, a tumor is a non-stoppable growing amount of cells.
When smooth is applied to a group of pixels the result is that
those isolated pixels loose importance. The smooth process is
done over the Mclasses matrix. The objective is to penalize
the spread pixels. It is done by giving more importance to the
grouped fifth class pixels. Thus, the final matrix solution will
increase the five class pixel groups using float numbers from
1 to 10. The smooth comparing data will be five. The higher
concentration of fives that are closer to each other, the higher
possibitily for that value to reach 10. The process will
produce a final matrix like the following one.
Fig 12. The final matrix is finally obtained. It could be appreciated how the
tumor has been finally found out. The tumor related pixels values are clearly
distant from the brain mass pixels.
The following matlab code represents how the smooth is
implemented:
%Create the Mresult matrix
Mresult = Mclasses;
%patch smoothing factor
rfactor = 0.125;
%isosurface size adjustment
level = .8;
%useful string constants
c2 = 'facecolor';
c1 = 'edgecolor';
p=patch(isosurface(smooth3(Mresult ==5),level));
reducepatch(p,rfactor)
set(p,c2,0.8*[1,0.5,0.5],c1,'none');
Fig 13. The smooth is applied for the pixel value of 5. The reducepatch
function with a rfactor lower than one perseveres the shape of the smoothed
zone. The set line configures the output representation. The result is shown on
Figure 3.
f
Mdat
a
Mclasses
Mresul
t
kmeans

 2
.
.
Mresult
Mdata
Mclass

ó


Mclass
kmeans Mdata

f
Mclasses
Mresul
t
kmeans


Mclass
kmeans Mdata

Mclasses
Smooth
Mresult


,5
Mresult
Smooth Mdata

International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
10
IV.ADT-3D COURSE OF ACTIONS
Thanks to the user interface, it is possible to monitorize
each step on the process of tumor detection, recognition and
isolation. The first step is Image load, which reads dicom
images from the hard drive and transforms them into a three-
dimensional image. The second step is brain mass isolation
consisting on two subprocesses such as noise reduction and
bone exclusion. Third, it is Tumor simulation, due to the
difficulty of finding head CT scans with tumors on them.
The simulation, as explained later on, consists on the
generation of a head tumor. The following step,
Segmentation, is the most important since it divides the
studied data (or head) into different classes for later envision
and study. The next step is Class Visualization allowing the
3D visualization of the different classes obtained from the
segmentation. Later on is Class Plotting; this step is used to
display on a bi-dimensional image all the different classes
resulting of the segmentation. Finally, it is
Tumor
Visualization, which is used to visualize the detected tumors
isolated from the rest of brain mass. Two possible
visualizations are offered to the user: a group of bi-
dimensional images that represents the diverse slices of the
detected tumors; and a three-dimensional representation of
those tumors.
A.Constant Functions
Here can be found the different functions, which are
present during the running time of the program. They can be
executed at any given time to obtain the desired result. These
results are display graphically, so the user of the system can
acquire a better knowledge of them. The quantity of feedback
given to users is vital. It simplifies the diagnosis process.
1)3D Visualization
The function’s name is self-explanatory. Provided that the
user of the system has executed correctly the step in which
this function is attached; then, the outcome will be a three-
dimensional representation of the results obtained.
Nevertheless, this process takes more time than users might
expect, due to the fact that the process of transformation
from a three-dimensional matrix into a 3-D representation
requires reading each single item on the matrix. Moreover,
since usually the dimensions of a matrix are 256x256x60, it
results on a composition of almost four million elements.
MATLAB handles the representation of any volumetric
data, but in order to perform a proper representation it
requires the programmer to establish certain information
related to color map, axis aspect, and other elements.
Once the user has the 3D illustration of the results, the
view can easily be rotated; amplified or reduce its scale; and
even change the color map. Automatically all these
functionalities are included by MATLAB. An example of
this type of representation can be seen on Figure 14.
2)Video Visualization
Is a function that makes possible a dynamic bi-
dimensional visualization of three-dimensional matrixes.
The process is simple and requires fewer time compared to
the 3D visualization function. In order to decide how the
data is going to be sliced, the function’s data input has to be
a three-dimensional matrix. The decision is done by sizing
the matrix and by selecting the smallest dimension. The next
step is to show the first 2D slices on the stack, then a counter
is set up to count a number of milliseconds. Once this is done
then the second 2D slice is shown and so on. The final
output is a dynamic movie simulation, which exhibits
correlating images.
Since this functionality is directly perform by implay(), a
MATLAB function. It is not strange to assume that color
map, velocity, and other aspects of the movie could be easily
changed.
3)Image Closure
The main purpose of this function is to close all other open
images but the interface, which is also considered a figure.
Some processes, such as 2D visualization of tumors, can
generate too many figures. Deleting the images one-by-one
could be boring and time-consuming for the user. Thus,
Image closure is provided to close all open images except the
user interface, making it fast and easy.
On the one hand, some users might find this function
unnecessary and useless. On the other hand, its users would
surely benefit from its features.
B.Image Load
If Image Load is not the first procedure ever to be executed
on the system, then any other action would have no reaction.
The reason for this is that this process is in charge of reading
the data input images. As stated above, Dicom images have
to be store in the computer or in a server and they have to be
reachable by this procedure.
It is not required that the input images enclose a common
string name including an identification number (e.g
“ImageName”+num+”.FileExtention”). However, the use of
a common string name would facilitate loading images into
the system, at least to the programmer. In any case this
would affect on response times from this process, but it
would help, on one hand, the organization of images
belonging to the same patient, and on the other hand, the
programming code used to read and load the images. This
way, it could be a more autometized and simplier process.
C.Brain Mass Isolation
The intention of this process is to isolate brain mass from
the rest. Now that images have been loaded into a matrix, it
is possible to get rid of all those parts that are not needed to
detect a brain tumor. The process is done in two steps: first,
Noise Reduction and second Bone Exclusion. Both
procedures will be explained below with plenty details.
1)Noise Reduction
The apparition of noise on images is usually unwanted and
random, causing a variation of values in the image’s pixels.
It can be originated, for example, on electronic devices such
as radios, sensor devices (e.g. cameras). Unfortunately, CT
Scans are not an exception to the apparition of noise. The
problem of noise reduction was deeply studied by Hanson
[11]. He differentiates three different types of noises:
random, artifactual and structural. He also classifies random
noise into statical, electronic and roundoff noise.
In relation, noise reduction was applied in two different
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

11
steps. First, the voxel values lower than twenty are assigned
a zero. A study of voxel values were done to check that those
values under twenty could be easily erase without the loss of
information. The second step is to delimitate brain mass
boundaries meaning that those values over the skull and
under the patients’ jaw are equalize to zero. This would
prevent those values to interfere on the system.
2)Bone Exclusion
The second step on Brain Mass Isolation is in charge of
minimizing bone tissue. The brain’s surrounding bone
complicates the process of tumor detection, recognition, and
isolation; therefore, it has to be minimized as much as
possible.
An extensive bone tissue values should be completed to
perform appropriately the bone exclusion. Such study
includes bone original Voxel values and the calculation of
the bone mass average. Then, the average bone values are
used to check that no other brain mass might be also
affected.
In the end, bone tissue is excluded from the CT Scan
three-dimensional matrix. If the matrix is now delineated
either onto two or three dimensions, the representation would
show the brain.
D.Tumor Simulation
As hard as it is to find a digital CT Scan example, getting
one with a tumor is even more difficult. For this reason,
tumors on ADT-3D have to be simulated. Simulating brain
tumors is not an easy task owing to the fact that they are
supposed to seem real, at least to the system. ADT-3D is
tested based on its capability to recognize brain tumors,
meaning that the most desirable features are accuracy and
effectiveness.
The tumors are simulated based on voxel density values.
Brain tissues were found to be between 60 and 120, raising
the question of which value had to be used. Various
alternatives such as above, below, and inside the stated
limits. The final decision was to simulate three different
brain tumors. The main characteristics for all three is
dissimilararity from each other. It can be accomplish by
using dispare voxel values, location, and size. Such
specifications intend to test the system to its maximum
capacity.
E.Segmentation
The segmentation algorithm used on ADT-3D is called
Jaimes. This algorithm is explained on further details on
section Jaimes’ Algorithm, included on Development
Details. This algorithm provides a better differentiation of
the original input data and thus a better segmentation is
achieved. Based on Kmeans segmentation algorithm
classifies the input data into five different classes. On the
fifth class is where detected tumors can be found.
Segmentation is the most time-consuming process due to
the fact that almost twelve million digits are used (three
element’s iteration for each item on a three dimensional
matrix of 256x256x60). It supplies information to Class K
Visualization, Class Plotting, and Tumor Visualization.
F.Class K Visualization
There is a simple but tedious requisite demanded by this
function. The requisite is the correct execution of
Segmentation. By using this function it is possible to
represent, in a three dimensional environment, the specified
K class given as an input value.
The data input is Segmentation’s data output, thus only a
matrix is given. In order to represent one of the classes
contained on the matrix all values different to K are
equalized to zero for the representation, but then the values
are restored. Otherwise, it could only be possible to delineate
the first class.
G.Class Plotting
Segmentation’s output matrix is used as input data, thus
Class Plotting uses the same matrix than Class K
Visualization. In this case the intention is to visualize a
determined matrix slice on a two dimensional draft. The
difference with Class K Visualization relays on the fact that
now all the different classes are shown on a 2D picture.
Class Plotting introduces the idea of individual segment
visualization. Such technique increases the feedback given to
the user of the system. As said before, is better to have an
informed user than an ignorant one.
H.Tumor Visualization
Segmentation supplies Tumor Visualization the necessary
information to envision tumors. Recalling Jaimes’ algorithm
outcome is a three-dimensional matrix containing the
belonging of each element to a K class, on a range of 1 to 5.
It is possible to say that tumors are localized on the fifth
class. Tumor detection, diagnosis, and recognition are done
on Segmentation. It is time to picture the result to the user,
usually a doctor, who would have the last word on this
matter. ADT-3D assist the process of detection, and under
that circumstance, offers the detailed information achieved.
1)Tumor’s 3D Visualization
Tumor’s 3D Visualization function is only in charge for
the 3D representation of the detected tumors. The only
difference with 3D Visualization is the data input.
2)Tumor’s 2D Visualization
Tumor’s 2D Visualization is in charge for generating a bi-
dimensional image of tumors location in a similar way to
Video Visualization. The idea is to display a variable number
of 2D images, not in video, each one of them containing
different parts of the tumors.
V.REASONS
ADT-3D (Three-dimensional Tumor Detection Assistant)
came up as a scientific proposal for tumor detection. The
main purpose was to offer information related to the possible
infected areas. In this matter it facilitates, complements,
reduces time, lowers cost, and sets up an entry point to other
further advanced systems.
A.Facilitates
Nowadays, CT Scans display for doctors the gained data
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
12
from patients. The data can be presented in different formats,
chosen by the doctors depending on the situation. Adding the
ADT-3D inside CT Scans would facilitate and greatly
improve the detection and diagnosis of tumor because it is
specifically designed for this mission.
B.Adds functionality
Doctors around the world would agree, "Medicine is not an
exact science". During the diagnosis process, most doctors
will contact colleges to discuss a series of events that point
towards a final judgment. The insertion of ADT-3D in a CT
Scan may lead to a greater efficiency detection tumors and
diagnosing each single patient. The improvement in
efficiency would utilize better the time doctors’ work
diagnosing the patient. Therefore, there would be more time
to expend curing the tumor.
C.Economizes
As a result of the implementation of ADT-3D system into
CT Scan there is a smaller delay for the results, but on the
other hand, since the data was processed and studied by the
ADT-3D, it saves time. The economization of time and
money is always seen as a positive asset. In any case, the
implementation of the ADT-3D would only improve and
advance the process. The addition of the new system would
never become a disadvantage. The possible high cost of the
new ADT-3D would be cover by the time savings on doctors
and patients’ time, and ever greatly it would be more than
justified by the lives that it would save.
D.Clarifies
For image diagnosis the most common method used by
doctors is the ocular identification of unknown shapes and
forms. Unfortunately, in few cases, this type of identification
cannot be used because the affected areas are particularly
small. If we take tumors as an example, it is expected that
first stages of the illness cannot be detected because of the
difficult location and the small size of the tumor. ADT-3D
assist doctors with the detection and diagnosis phase, which
includes early stages since it's a computerized system that
can access all the data from the patient, even the smallest
piece. The figure 3 represents how easily a tumor could be
shown with ADT-3D.
Fig 14. The figure represents the tumors detected in a brain example. There are
two cubes that represent two experimental tumors.
E.Sets up an entry point
At the moment there is no other product similar to
ADT-3D in the market. Its broadcasting could be
inspirational for people with especial interest, attracting
better facilities or more time and money. The utilization of
this system can, positively improve the quality of life for
many people. Present CT Scan or MRI's do not include any
kind of Artificial Intelligence algorithms. The expansion of
Artificial Intelligence to other fields has improved their main
features and client satisfaction; therefore, the inclusion of AI
methods and algorithm in Medicine would also help any
patient
VI.RESULTS
1)Introduction
First of all, it should be said that ADT-3D is a prototype. As
a prototype, the goal is achieved; the brain tumor is found,
located, and shaped.
ADT-3D objective is to find a brain tumor, excluding any
other CT scan elements such as bone, face, neck or skin.
ADT-3D works only with the brain mass.
There is a limitation in ADT-3D results because it has only
tested with data source in which the experiments were done.
As said before, in the experiment the tumors were simulated
through data.
The following test were run after the matrix was obtained,
delimitated, and the tumors were added. Unfortunately, it
was impossible to find any CT scan examples which had
tumors. Neither on the net, nor at any Madrid’s hospital
tried.

2)Tumor density equal to brain average.
Once known the source protocol, it depends on the CT
manufacturer; the non-meaningful value is selected. The test
consists on adding a simulated tumor value similar to the
brain tissue. If ADT-3D works properly, this kind of tumor
cannot be detected by the system. The following picture
shows where the simulated tumor was added..
Fig 15. The simulated tumor has a value of 90, which is really close to brain
tissue. It is not simple to distinguish where the tumor begins.
The simulated tumor could only be distinguished because of
the shape, but never by comparing it to the voxel density
levels. In order to recognize the tumor, another kind of
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

13
artificial intelligence algorithm could be use to target shape
oriented recognition.
An example of the detected tumors can be seen on Fig 14.
In the sample (Fig 14), the 33rd slice does not show the
tumor with value 90 because the density difference is almost
equal to the surrounding brain tissue.
The following picture represents how the tumor is
segmented by Jaimes’ Algorithm.
Fig 16. The tumor values are exactly the same, but due to the low value, for
each slice the tumor is classified by the segmentation algorithm on different
classes, instead of the same class.
Figure 16 shows how the Jaimes’ Algorithm perceives the
simulated tumor. For each slice the Kmeans class changes.
In this way, it is impossible to join all the squares into a 3D
image or a five class section for 2D recognition.
3)Intermediate Tumor Value
The Intermediate Tumor Value runs a simulation slightly
above the average brain tissue density values. The tumors
with density values above 90, the average, are easily
recognized by ADT-3D. Tumors will be distinguished but the
smoothen step is crucial here. Without image smoothing
many spread pixels would be perceive as tumors.
The next picture represents how is a 110 value tumor is
represented.
Fig 17. The tumor is seen better now because there is more distance comparing
to the brain average value.
Once our Algorithm is applied the tumor is automatically
detected. On Figure 14 the top part of the cube shows a
tumor. The 2D result is shown below.
Fig 18. The 2D solution represents with garnet color the possible tumors and the
number of pixels that satisfied the tumor condition.
Therefore, the case illustrates the ADT-3D objective, to
automatically find tumors otherwise undetected by
conventional CT Scans.
4)High Tumor Value
The last test is done to ensure the recognition of high-
density tumor values. The voxel density selected is 130,
which is hardly ever found merged into brain tissue. It could
be said that any other value higher than 130 will be surely
recognized, because the algorithm performs better in voxel
density values above and beyond 130.
It is important to emphasized how the simulated tumor is
contrasted differently to the brain mass. The following
images demonstrate the difference in contrast.
Fig 19. The tumor value is so high that comparing to the brain mass it is closer
to white values, such as bone.
The 3D result obtained is shown on the Figure number 13.
There are two cubes. This tumor is represented as the lower
cube. The 2D recognition representation for one slice is on
figure .
When comparing figures 18 and 20 is easily perceive that
the higher the tumor density the less number of elements will
be represented in a 2D picture. Figure 18 has more noise,
more irrelevant information, while Figure 20 has only the
tumor and two small data zones.
Slice 27/60
Slice 33/60



International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
14
Fig 20. The tumor on slice 15 is shown as a garnet square with 21 high chance
tumor pixels.
5)Results Analysis
These three simulated tumors were use to check ADT-3D
performance. The first test verified ADT-3D detection
capabilities for tumors with voxel density values equal to
brain tissue. The second test tested ADT-3D capacity to find
tumors with density voxel values slightly above brain tissue
average. Finally, ADT-3D was tested for tumors with voxel
density levels above 130.
The three tests reinforced the usage of Jaimes’ algorithm on
ADT-3D for brain tumor detection. Anyways, the prototype
has to be further tested and CT scan voxel density values
have to be collected to better understand and improve the
process.
VII.PRODUCT LIMITATIONS
Unfortunately, any prototype has its drawbacks and
limitations. The positive aspect is that it is possible to learn
from them so that future updates can overtake them. Once
ADT-3D was finalized, it was studied and tested in further
detail to analyze its outcome. The negative side carries
some limitations that could be divided into two independent
categories: Physical and Software limitations.
A.Physical Limitations
Physical limitations concern the hardware devices in which
ADT-3D is launched. Ram memory, processor speed, and
similar features affect directly to the velocity of execution
but never to the effectiveness. As a matter of fact ADT-3D is
a very demanding project in terms of hardware
requirements. It needs at least 1 Gb of RAM, and a
processor equal or higher to 2,5GHz to work properly. The
results of investing in hardware would mean better response
times, which is always desirable.
B.Software Limitations
Software limitations are related to the effectiveness of the
program. ADT-3D uses density values to discriminate
classes during the segmentation process. There could be
tumor areas which densities are too similar to densities on
the surrounding areas, this would make impossible to detect
that tumor. This is an important software limitation of
ADT-3D but doctors already face this problem. Doctors
many times detect tumors using their visual perception on a
CT Scan, which also shows density values.
C.Area of Applicability
ADT-3D is a remarkably peculiar product which requires
specific data from a CT Scan. Its main goal is to assist with
tumor detection. For this reason the only possible application
in the field could be Medicine. It is in radiology where the
product is more feasible because its implemented into CT
scans. Computer applications have revolutionized every area
of Medicine, but they have to be updated.
VIII.CONCLUSION
Although additional studies are necessary to provide
ADT3D more capabilities, it is clear that this project
represent the very start of computing software applied to
medical area.
Nowadays it is known that there are many applications
dealing with medical stuff recognition using Artificial
Intelligence. Each IT company focused on TC develops not
only in the hardware, but also in software to deal with the
achieved data. ADT3D tries to be cross-platform software
able to work with any kind of raw data acquired by a CT
machine, no matter its firm.
The current ADT3D is a prototype, an introduction of what
could be done. The whole project could be improved, from
the GUI to the core. The Jaimes' Algorithm, or the core as it
was said, is a mathematical approximation to the tumor
recognition which must be studied and improved. The
present algorithm only works for simple cases where the
whole tumor has a constant value.
Visual Intelligence Computing branch is getting more and
more significant, acquiring new people day by day. The
applicability of this computing area is enormous because it
could be used easily as a human tool, not only for medical
stuff, but also, for car identification numbers, people
identification, optical character recognition ...
Developing ADT3D has been a great challenge during the
last eight months. During the development of this project,
MATLAB has been our third coworker. Matlab is a great
prototype oriented application with a clear advantage: find
much information on the Internet. Internet research has been
also important to achieve the goal of recognition, so, as well
as people provide their free code on the internet, ADT3D
will be for free for all on the Internet.
Upgrades are an important part on each project or
prototype. ADT-3D particularly could improve the user
interface. The GUI was developed to perform the desired
task but user interaction with it was not taken into account.
User experience is key to many applications, upgrading the
actual interface could make easier and simpler the
understanding of the process.
New functionality could be added so that ADT-3D
performs new tasks as acquiring data from the computer or
servers to treat it as input data. Many additional functions
could improve the productivity and capabilities of the
system.
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

15
IX.REFERENCES
[1] J. Gonzalez, J.A. Vara del Campo, and J.C. Vazquez Luna. Tomografia
computarizada, 1992.
[2] X. Vila. Obtención de imágenes a través de un TC. [Online]. Available:
http://www.xtec.es/~xvila12/funciona.htm.
[3] Computed tomography, GE Healthcare Glossary, [Online], Available:
http://www.medcyclopaedia.com/library/radiology/chapter04/4_2.aspx
[4]
X-Ray,
GE Healthcare Glossary,

[Online],
Available:
http://www.medcyclopaedia.com/Home/library/glossaries/x_ray.aspx
[5] MedicineNet, Inc. Brain Tumor definition. [Online]. Available:
http://www.medterms.com/script/main/art.asp?articlekey=2519
[6] University of California, San Francisco Medical Center, Brain Tumors.
[Online]. Available:
http://www.ucsfhealth.org/adult/medical_services/cancer/brain/conditions/brain
_ tumor/diagnosis.html


[7] Matworks TM, MATLAB product information. [Online] Available:
http://www.mathworks.com/products/matlab/
[8] Matworks TM, MATLAB Newsgroups .[Online]. Available:
http://www.mathworks.com/matlabcentral/newsreader/
[9] Department of Computer Vision, University of Bristol, Segmentation
definition [Online]. Available:
http://www.cs.bris.ac.uk/Teaching/Resources/COMS11200/jargon.html
[10] Wordnet 3.0, Princeton University, Segmentation definition. [Online]
Available: http://wordnet.princeton.edu/perl/webwn?s=segmentation
[11] Kenneth M. Hanson: Noise and contrast discrimination in Computed
tomography. Radiology of the Skull and Brain, Vol. 5: Technical Aspects of
Computed Tomography, T. H. Newton and D. G. Potts, eds. (C. V. Mosby, St.
Louis, 1981) ISBN 0-8016-3662-0 (v. 5) (p 3943-3943)
Other sources
• General Electrics Healthcare
Computed

Tomography.

[Online]
http://www.gehealthcare.com/eues/ct/products/products_technologies/prod
ucts/lightspeed-rt16/index.html
ListhSpeed

rt16

product

information.

[Online]
http://www.gehealthcare.com/eues/ct/products/dedicated_systems/products
/ctmr.html
• Bruno Alcouffe (SES Leader) & Olivier Segard (NP&S)
Advantage Workstation DICOM Connectivity Validation
Specification with Other DICOM Products
Revision 1, 1999
• Siemens Medical Solutions
Somatom Definition
[http://www.medical.siemens.com/webapp/wcs/stores/servlet/CCategor
yDisplay~q_ProductDisplay~a_fvwe~a_catalogId~e_-11~a_catTree~
e_100010,1007660,12752,1008408~a_langId~e_-11~a_productId~e
_168189~a_storeId~e_10001.htm]
Syngo Neuro Perfusion CT
[http://www.medical.siemens.com/webapp/wcs/stores/servlet/ProductD
isplay~q_catalogId~e_11~a_catTree~e_100010,1007660,12752,100
8408*4222936910*1902206816~a_langId~e_11~a_productId~e_14
5942~a_productParentId~e_168189~a_relatedCatName~e_Clinical+
Applications~a_storeId~e_10001.htm]

X.AUTHORS
J. García Castellot, Madrid, 1983, Bachelor of Computing
in Business Information Technology, University of Wales,
Wrexham, Wales.
5th Year Computing Engineering Student at Universidad
Pontificia de Salamanca, Madrid, Spain.
e-mail: jaimegaca@gmail.com
J. Lazcano Bello, Madrid, 1983, Bachelor of Computing in
Business Information Technology, University of Wales,
Wrexham, Wales.
5th Year Computing Engineering Student at Universidad
Pontificia de Salamanca, Madrid, Spain.
e-mail: lazcano.jaime@gmail.com
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
16

Abstract — when we talk about cancer diagnosis the most
important thing is early diagnosis to prevent cancer cells from
spreading. We may also consider the high cost of diagnostic
tests. Our approach seeks to address both problems. It uses a
software based on Bayesian networks that simulates the cause-
effect relationships and gets the chance of suffering a
pancreatic cancer or lung cancer. This software would support
doctors and save a lot of time and resources
Keywords— Bayesian networks, pancreatic cancer, lung
cancer and computer-aided diagnosis
I.INTRODUCTION
ancreatic cancer is the cancer with a higher mortality. Is
the fourth leading cause of death in the United States of
America. In 2006 were predicted 33,700 cases of which
32,300 were fatal [1]. In Spain (year 2004) the situation is as
follows [2]: 2,438 deaths (4% of all cancer deaths in men)
and 2098 (5.8% of all cancer deaths in women).
P
Lung cancer is the most widespread cancer (27.4% in
Spain). In USA accounted for more deaths than breast
cancer, prostate cancer, and colon cancer combined. In year
2004, 108,355 men and 87,897 women were diagnosed with
lung cancer (89,575 men and 68,431 women died).
Both lung cancer and pancreatic cancer persist as a
challenge to medicine, because despite advances in
diagnostic techniques and therapeutic resources, the statistics
reflect the few dressings in relation to the number of patients
cared for. This is due largely to the fact that in recent years
there have been no meaningful responses to attempts to
reduce exposure to carcinogens, in the case of lung cancer,
and also that did not improve the detection of the tumour at
an early stage, which would have been able to increase the
possibility of healing, but especially not yet have an
appropriate solution for these pathologies.
In an organization as the Spanish health, oncology
institutes are not designed so far to leave in search of any
persons concerned. Patients who received are
overwhelmingly referrals with confirmed or presumptive
diagnosis and usually at a relatively advanced stage of
evolution.
Hence the need arises to create a web application that
calculates instantly and reliably the likelihood of suffering
from cancer. This reliability will ultimately depend on the
human factor, since it will be the oncologists who through
his experience and statistical data to determine the values of
Manuscript created May 27, 2008. A. Álvaro Núñez Díaz is now studying
at the Pontifical University of Salmanca (e-mail: alvand9@gmail.com).
B. Luis Lancho Tofé is now studying at the Pontifical University of
Salmanca (e-mail: elluiso_7@hotmail.com).
probability tables that use the system to infer the likelihood
of final diagnosis. The medical knowledge is stored in a
bayesian network (see Fig. 6) that models the relationships
between symptoms, risk factors, etc.
The project aims to provide an early warning system and
determine the likelihood of developing now or in the future
some of the, previously mentioned, types of cancer,
according to risk factors, symptoms and other diseases.
Given the seriousness of these types of cancer and the
known fact that a large percentage of cases are discovered in
advanced stages of the disease, a tool aimed to inform and to
alert the population is a great help. This type of tool can be
further enhanced in the current social context, where a
growing sector of the population has access to the Internet,
from where access to this software can be easy and
instantaneous, increasing considerably the potential utility of
the application for physicians and patients.
Besides the usefulness purely informative, this project
would have a preventive character because:
• If the application concludes that a person is at risk of
cancer due to certain risk factors, this person could increase
their awareness of the problem and how to avoid them.
• A sufficient risk of suffering from cancer could push the
individual to carry out periodic reviews, which could lead to
earlier detection of the disease.
II.CANCER DIAGNOSIS
A.Pancreatic cancer
Pancreatic cancer is presented with an incidence of 8 to 10
cases per 100,000 inhabitants per year in many industrialized
countries of Europe [1]. This incidence is higher than that
recorded a few years ago [2]. In Spain, the incidence of
pancreatic cancer has risen dramatically in the second half of
the twentieth century, with an increase in the same close to
200% [3]. Pancreatic cancer is the fourth leading cause of
cancer deaths in the United States, both men and women
[1,4,5]. Until the last evolutionary stages, when the majority
of diagnoses are introduced, pancreatic cancer is a disease of
evolution silent and insidious. Since the disease is usually
diagnosed when it is no longer confined to the pancreas, has
a survival rate at 1 year of
15-20%, and when survival is valued at 5 years, that figure
drops to 2-4% [1,6]. Only when the pancreatic cancer is
small (less than 2 cm.) sits at the head of the organ, and
there are no lymph node metastases and invasion of
neighbouring organs, can be expected survival rates of 20%
to 5 years.
A. Álvaro Núñez Díaz, B. Luis Lancho Tofé
Computer-aided diagnosis of pancreatic and
lung cancer
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

17
B.Lung cancer
It is a very aggressive and deadly cancer, most patients die
before the first year after diagnosis. It is the most common
cancer in men over 35 countries, especially developed or
industrialized countries [22], including Spain, being the
leading cause of cancer death in both men and women. More
people die of lung cancer than of colon cancer, breast and
prostate combined. It represents 12.5% of all malignant
tumors in Spain. [23]
The number of cases has increased since the beginning of
the century, doubling every 15 years. The incidence has
increased nearly 20 times between 1940 and 1970. At the
beginning of the century, it was considered that most were
metastatic lung tumors, and that cancer was rare primitive.
III.SYMPTOMS, RISK FACTORS AND PATHOLOGIES
risk factor is anything that affects the likelihood of
developing a disease. Having a risk factor, or even
several factors, does not mean that a person can contract the
disease. In addition, many people who develop the disease
may have no known risk factor [9].
A
A.Pancreatic cancer
Over the years it has investigated the possible existence of
risk factors that influence the emergence of pancreatic
cancer. In this sense, risk factors have been invoked since
demographic factors to personal terms with a greater or
lesser risk could be related to pancreatic cancer [10].
It is unknown the cause of emergence of pancreatic
cancer; it is more common in smokers and in the obese
people and almost 1/3 of cases is due to cigarette smoking.
There is controversy over whether diabetes type 2 is a risk
factor. In addition, it is known that a small number of cases
are related to symptoms that are transmitted through families
[11].
Now we are going to show a list of symptoms, risk factors
and pathologies related with pancreatic cancer. After, we
will develop the most significant facts and will put the range
of values used in the bayesian network. We have chosen the
range of values depending on the relevancy of the fact in
relation with the diagnosis.
Symptoms [11,15,16,17]:
• Weight Loss (Fig. 2)
• Loss of appetite (Fig. 2)
• Abdominal distention (Fig. 3)
• Edema in lower extremities (Fig. 3)
• Backache (Fig. 3)
• Mass palpable abdominal (Fig. 3)
• Abdominal pain (Fig. 3)
• Clots or fatty tissue abnormalities (Fig. 3)
• Uneven texture of fat tissue (Fig. 3)
• Relaxation of the gallbladder (Fig. 3)
• Nausea or vomiting (Fig. 3)
• Diarrhea (Fig. 3)
•
Indigestion (Fig. 3)
• Eyes and skin yellowish (Fig. 3)
• Urine of dark colour (Fig. 3)
•
Faeces of clear colour (Fig. 3)
Risk factors [11,13,14,19,20,21]:
•
Sex (Fig. 4)
• Age (Fig. 4)
•
Smoking (Fig. 4)
•
Food (Fig. 4)
• Consumption of coffee (Fig. 4)
• Consumption of alcohol (Fig. 4)
• Race (Fig. 4)
• Obesity and physical inactivity (Fig. 4)
• Exposure to asbestos, pesticides, dyes, petroleum
(Fig. 4)
Pathologies [10,17,18]:
• Pernicious anemia (Fig. 5)
Fig. 1 Immunohistochemical detection of KIT and SCF expression in pancreatic
cancer. (A) KIT-positive pancreatic cancer cells (× 200). (B) SCF-positive
pancreatic cancer cells (× 200).
[Credit:
Yasuda et
al.
Molecular
Cancer
2006 5:46
doi:10.1186/1476-4598-5-46]
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
18
• Endocrine tumours (Fig. 5)
• Esteatorrea (Fig. 5)
• Anorexia (Fig. 5)
• Ascites (Fig. 5)
• Cystic fibrosis (Fig. 5)
• Chronic pancreatitis (Fig. 5)
• Tonsillectomy (Fig. 5)
• Peptic ulcer surgery (Fig. 5)
• Cholecystectomy (Fig. 5)
• Allergies (Fig. 5)
• Thrombophlebitis (Fig. 5)
• Long-term diabetes (Fig. 5)
• Diabetes mellitus of recent beginning (Fig. 5)
a) Age
The risk of pancreatic cancer increases with age. Almost
all patients are older than 45 years. More than 90% are over
55 and almost 70% are over 65. The average age at the time
of diagnosis is 72 years. (see Fig. 4)
In the bayesian network we used the following range of
values: children, young or adult
b) Sex
Men are slightly more likely to develop pancreatic cancer
than women. This difference was most pronounced in the
past, but has declined in recent years. Perhaps that
difference, at least partly, is due to higher consumption of
tobacco. (see Fig. 4)
Range of values: male or female
c) Race
Black people and Orientals in this order are more likely to
develop pancreatic cancer compared to white people. The
cause is unknown, but may be due to higher rates of smoking
and diabetes among men, and overweight in women of those
races. (see Fig. 4)
Range of values: white, black or oriental
d) Tobacco
Smokers have a risk of two to three times greater of
suffering pancreatic cancer. Scientifics consider that this
may be due to chemical agents that exist in cigarette smoke.
The chemical agents enter in the blood causing damage to
the pancreas. It is thought that between 20 and 30 percent of
all pancreatic cancer cases are smoking cigarettes. (see Fig. 4)
Range of values: former smoker, non-smoking, moderate
or high
e) Diet
Some studies have found a relationship between pancreatic
cancer and a diet high in fats, or that includes the
consumption of lots of red meat, pork and processed meats
(such as sausage and bacon). Others have discovered that a
high consumption of fruits and vegetables can help reduce
the risk of pancreatic cancer. But not all studies have made
these associations, and the true value of the food is still under
study. (see Fig. 4)
Range of values: good or poor
f) Obesity and physical inactivity
People with overweight are much more likely to develop
pancreatic cancer, like those with little physical activity. (see
Fig. 4)
Range of values: mild, moderate or serious
g) Diabetes
The pancreatic cancer is most common in people who
suffer from diabetes mellitus. The reason for this relationship
is unknown. Most risk is detected in people with diabetes
type 2, which usually starts in adulthood and is often linked
to overweight and obesity. There is no certainty if there is an
increased risk in people with diabetes type 1 (in younger
people). (see Fig. 5)
Range of values: negative or positive
h) Chronic pancreatitis
The chronic pancreatitis is a long-term inflammation of the
pancreas. This condition is associated with an increased risk
of pancreatic cancer, but most patients with pancreatitis
never develop this cancer. Perhaps the main reason for this
association is that patients with pancreatitis also are more
likely to have other risk factors such as smoking. Very few
cases of chronic pancreatitis are caused by inherited gene
mutation (see family history). People with this hereditary
form of chronic pancreatitis seem to have a high risk for a
lifetime of suffering pancreatic cancer (approximately 40 to
75 percent). (see Fig. 5)
Range of values: negative or positive
i) Exposure to working conditions
The intense exposure at work to certain pesticides, dyes
and chemicals used in metal refinery could increase the risk
of pancreatic cancer. (see Fig. 4)
Range of values: negative or positive
j) Family history
It seems that the pancreatic cancer occurs more frequently
in some families. Perhaps up to 10% of pancreatic cancers
are related to changes inherited DNA (mutations). These
changes often increase the risk for other cancers as well.
Some examples include:
Mutations in the BRCA2 gene: this mutation also
increases the risk of ovarian cancer and breast cancer
Mutations in the p16 gene: it also increases the risk of
melanoma.
Mutations of the gene PRSS1: cause severe pancreatitis
during an early stage.
Hereditary non polyposis colorectal cancer (HNPCC or
Lynch syndrome): also increases the risk of colorectal cancer
and endometrial cancer.
Peutz-Heghers Syndrome: also has partnered with polyps
in the digestive tract and other cancers.
Scientists have found some of these changes in DNA and
can be identified through genetic testing. (see Fig. 4)
Range of values: negative or positive.
k) Stomach Problems.
The stomach infection with the bacterium called
Helicobacter pylori, which causes ulcers, could increase the
risk of pancreatic cancer. Some researchers believe that
excessive stomach acid could also increase the risk. (see Fig.
3). Range of values: mild, moderate or serious
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

19
Fig. 2 Detail of the bottom left corner of the bayesian net modeling the patient's
symptoms. Both common symptoms and the specific symptoms of lung cancer
are included.
Fig. 3 Detail of the specific symptoms of pancreatic cancer. They are divided
in: symptoms related with distension of the gallbladder, symptoms related with
digestive problems, symptoms related with jaundice and general symptoms of
pancreatic cancer.
Fig. 4 Detail of the center of the bayesian net. We can observe risk factors
grouped in four categories: diseases in the family history, quality of food,
exposure to carcinogens and general risk factors. In the last category we have
sex, smoking, etc.
Fig. 5. Detail in the upper right corner. Now we have pathologies. They are
divided in two big groups: pathologies related with pancreatic cancer and
pathologies related with lung cancer. It is remarkable the two types of diabetes.
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
20
Fig. 6. Overview of the bayesian network that models the system´s medical knowledge. The color code of the nodes is:
- Yellow: it expresses the input (risk factors, symptoms and diseases). These nodes have an associated probability value. Depending on the response of the patient (or defaults in case of unknown
response) will be established the evidences of the network.
- Pink: from left to right, symptoms and pathologies related with lung cancer. Top down, the node of lug cancer that determines the final positive/negative probability (the output)
- Sea green: from left to right, symptoms and pathologies related with pancreatic cancer. Top down, the node of pancreatic cancer that determines the final positive/negative probability (the output)
- Red: from left to right, the general nodes that determine symptoms, risk factors and pathologies
- Light green: node that determines the quality of food knowing consume of alcohol, café, vegetables and fruit
- Lavender: node that represents exposure to carcinogens. The node depends on: asbests, radiation, pesticides, dyes, petroleum, sulphur,coal and arsenic
- Dark blue: node that represents the incidence of diabetes in the final diagnostic. It depends of two types of diabetes: long-term diabetes and diabetes mellitus of recent beginning
- Brown: jaundice. It depends on three symptoms: eyes and skin yellowish, urine of dark colour and faeces of clear colour
- Grey: digestive problems. This node depends on three symptoms: nausea or vomiting, diarrhea and indigestion
- Orange: distension of the gallbladder. Depends on two symptoms: clots or fatty tissue abnormalities and uneven texture of fat tissue
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

21
l) Consumption of coffee
Some previous studies indicated that coffee consumption
might be a risk factor; more recent studies have been unable
to confirm this. (see Fig. 4)
Range of values: little or much
m) Consumption of alcohol
Most studies have found no relationship between alcohol
consumption and pancreatic cancer. However, consumption
of alcohol in large quantities may increase the risk of
developing diabetes and chronic pancreatitis. (see Fig. 4)
Range of values: little or much
B. Lung cancer
Lung cancer is so far the deadliest in the developed world.
One of the reasons that make it so deadly is that lung cancer
is often not detected until it is in advanced stages.
However, some lung cancers are diagnosed in its early
stages as they are detected by tests like bronchoscopy
(visualization of the interior of the bronchi through a flexible
lighted tube), or sputum cytology (microscopic examination
of the cells contained in phlegm that is ejected with cough).
Symptoms:
• Nail anomalies (Fig. 2)
• Weight Loss (Fig. 2)
• Loss of appetite (Fig. 2)
•
Facial swelling (Fig. 2)
•
Swallowing dificulties (Fig. 2)
• Chest pain (Fig. 2)
• Hoarseness (Fig. 2)
• Wheeze (Fig. 2)
•
Shortness of breath (Fig. 2)
• Bloody sputum (Fig. 2)
• Cough (Fig. 2)
Risk factors:
•
Sex (Fig. 4)
• Age (Fig. 4)
•
Smoking (Fig. 4)
•
Food (Fig. 4)
• Consumption of coffee (Fig. 4)
• Consumption of alcohol (Fig. 4)
• Race (Fig. 4)
• Obesity and physical inactivity (Fig. 4)
• Exposure to asbests, pesticides, dyes, petroleum (Fig.
4)
Pathologies:
• Chronic bronchitis (Fig. 5)
• Tuberculosis (Fig. 5)
• Pulmonary infarction zones (Fig. 5)
a) Tobacco
The tobacco reaches the alveoli and water-soluble
components are absorbed by the mucous, not being absorbed
the fat-soluble ones (tar or pitch) that contain carcinogenic
polycyclic aromatic hydrocarbons. The pitch is phagocytes by
alveolar macrophages and eliminated in the sputum, but not
all alveolar macrophages will be eliminated in the sputum,
many of them break in their journey towards the glottis
leaving the pitch. This will be deposited in the carina
(confluence of the bronchi), especially in the more upper and
peripherical ones, producing its irritation. The pitch irritates
the mucous membrane and destroys the surface of the
respiratory epithelium forcing the basement membrane to
increase their proliferative capacity. The basement
membrane increases both his "turn over" (basal cell
hyperplasia, stratification of the epithelium), ending
producing squamous cell metaplasia, which will evolve to
dysplasia (carcinoma in situ) and finally to anaplasia as
carcinoma in situ and invasive carcinoma. [24] . (see Fig. 4)
Range of values: former smoker, non-smoking, moderate
or high
b) Sex
Lung cancer is the most common malignancy in men. The
relationship between the sexes nowadays is 4 men for every
woman: (4:1). The woman continues to have a lower
incidence, but has already been located in some countries in
second place after breast cancer, or even in the first place.
Several studies have found that the lung cells of women
have a higher chance of contracting cancer when exposed to
tobacco. (see Fig. 4)
Range of values: male or female
c) Family history
If you've suffered a lung cancer, there is an increased risk
of having another lung cancer. Brothers and sons of people
who have had lung cancer may have a slightly higher risk. If
the father and grandfather of a person died from lung cancer,
and this person smokes, the most likely cause of his death
will be a lung cancer. . (see Fig. 4)
Range of values: negative or positive
d) Exposure to work conditions
Asbests: This is another risk factor for lung cancer. People
who work with asbests have a higher risk of suffering from
lung cancer and, if also smoke, the risk increases greatly.
Although asbests has been used for many years, Western
governments have almost eliminated its use at work and in
household products. The rate of lung cancer linked to
asbests, mesothelioma, often begins in the pleura.
Radon: Radon is a radioactive gas that is produced by the
natural decay of uranium. Radon is invisible and has no taste
or smell. This gas can concentrate inside houses and become
a potential risk of cancer.
Some workers industry-related to asbests, arsenic, sulfur,
(the three Aes) vinyl chloride, hamatita, radioactive
materials, nickel chromates, coal products, mustard gas,
chloromethyl ethers, gasoline and diesel derivatives, iron,
beryllium, etc.., whereas the non-smoker has a probability of
1 suffering from lung cancer, the smoker has 30 or 40,
employees of these industries have up to 70 times more at
risk. All types of radiation are carcinogenic. The uranium is
weakly radioactive, but lung cancer is four times more
prevalent among non-smoker miners of uranium mines than
in the general population and ten times more prevalent
among smoker miners. [25, 26] (see Fig. 4)
Range of values: negative or positive
International Journal of Interactive Multimedia and Artificial Intelligence, Vol.I, Nº 1, ISSN 1989-1660
22
e) Pollution (place of residence)
It is conceivable that pollutants from the atmosphere
(especially urban) play a role in increasing the incidence of
bronchogenic carcinoma today. Lung cancer is most common
in the city than in rural areas (1,3-2,3 times higher among
men with similar smoking costumes) because:
• Engine gases (cars and motor vehicles in general)
and heating systems. The sulfur dioxide is one of
the most important carcinogenic substances.
•
Particles of pitch pavement of the streets.
• Radioactive Particles.
• The radon gas and natural radioactivity are very
abundant in certain geographical areas.
These four factors have an action mechanism like the
tobacco. While most authors acknowledge the existence of a
small urban factor in the incidence of lung cancer, the main
culprit, with overwhelming numerical difference is the
tobacco. (see Fig. 4)
Range of values: field or city
f) Diet
Some studies conclude that diets with little vegetables
could increase the risk of lung cancer in people who are
exposed to tobacco. It is possible that apples, onions and
other vegetables contain substances that offer some
protection against lung cancer. It is believed that certain
vitamins, especially vitamins A and C, are protective of the
bronchial walls, by their ability to inactivate free radicals,
carcinogens, or their ability to precisely regulate certain cell
functions, across different mechanisms. But it has been
shown that β-carotene was ineffective as chemoprevention of
lung cancer and also two large studies like the Carotene and
Efficacy Trial (caret) and Alpha-Tocopherol, Beta-Carotene
(ATBC) Lung Cancer Prevention Study, 18000 and 29000
participants, respectively, showed that consumption of beta-
carotene increased (rather than reducing) the risk of
contracting lung cancer (18% on average), and particularly
for smokers of 40 or more cigarettes daily , Who experienced
an increase in the incidence of 42%. (see Fig. 4)
Range of values: good or poor
g) Prior bronchial pathologies
Chronic bronchitis: It is accepted that can cause lung
cancer.
TB: It is a chronic irritation on the lungs parenchyma that
leaves a scar that favors the emergence of lung cancer
(carcinoma or scar scar on cancer, especially
adenocarcinoma).
Areas of pulmonary infarction, inclusions of foreign
bodies, idiopathic pulmonary fibrosis (10% dying from
cancer bronchogenic), scleroderma and other kinds of scars.
(see Fig. 5) Range of values: negative or positive
h) Viruses
The relation of viruses with lung cancer has two different
bases:
In the experimental field, the incidence of bronchial
epithelial metaplasia is caused by paramyxovirus.
As for the human pathology, there is a link with bronchus-
alveolus carcinoma.
i) Age
Mainly this cancer affects people between 55 and 65 years.
80% of cases occur in patients over 50 years of age. Lung
cancer is very uncommon in people younger than 40 years.
The average age of the patients is 60 years but more cases are
being diagnosed in young subjects. (see Fig. 4)
Range of values: children, young or adult
IV. THE APPLICATION
B.Description of the system
The main engine of the system is artificial intelligence.
Medical knowledge is built through Bayesian networks that
simulate disease diagnosis. The system uses the following
software: Genie & Smile. (Decision Systems Laboratory,
Department of Information Science and Telecommunications
and the Intelligent Systems Program, University of
Pittsburgh, USA). The project has been developed in the C#
language under the platform .NET using the library
smilenet.dll to access network capacities of the Bayesian
networks software previously mentioned.
The network nodes are distributed in three general
categories: symptoms, risk factors and diseases. Each one of
this general nodes is influenced by several, which are
grouped depending on the type of cancer they are related to
(pancreatic, lung or both), except in the case of risk factor,
which are all considered as global.
Each node contains tables of probability, which reflects the
importance of the evidences that each node describes. This
probability values must be defined by experts and/or using
statistical information. In the case of this prototype, the
values have been established according to statistical
information found in previous studies. With some reference
values, we created a numerical progression from the lower-
probability-cases to the higher ones, but this is only an
approximation of which it should be. In order to obtain
consistent diagnostic values, these tables must be reviewed
by an expert who can, according to his knowledge, set more
detailed and reliable values for these tables.

Fig. 7. Colour-enhanced image of a lung cancer cell dividing. [Credit: Anne
Weston]
International Journal of Interactive Multimedia and Artificial Intelligence , Vol. 1, Nº 1., ISSN 1989-1660

23
When application is being run, the doctor will ask the
patient about the evidences (above nodes of the network) and
will insert the appropriate value in each field of the form
(each field has as much predetermined answers as rows in its
bayesian network’s node. The doctor must select one
depending on the patient’s answer).
If some evidence is unknown, the system will take values
of equally likelihood (for example, abdominal distention:
will produce an evidence value of 0.5) in
order not to distort the outcome.
Once the form has been completed, the diagnostic can be
started. The system will return the likelihood of suffering
from lung cancer and pancreatic cancer.
This system has been developed as a web application
(which allows easy updates) in order to be adaptable to the
complex structure of the modern medical organizations or
research groups. Many of these organizations have a
distributed architecture which does not allow isolated
elements to work properly. The web application allows an
expert to modify the bayesian network in the application
server (due to new discoveries), either the structure of the
bayesian network or the values of the tables, and users will
have the latest version available immediately via web
(Medical Center Intranet, Internet with VPN, LAN ...).
C.Structure of the application
Each functional element of the object oriented application
is implemented by two files: the .aspx file, that shows the
interface with the user and the .aspx.cs file, that will control
the actions of the functional element. Following this schema
we can divide the structure of the application in three parts:
1.
Input:
a) AgenteDeInterfaz.aspx: this file shows a form with
fields that should be filled in by the user. The fields are drop
down lists with several options depending on the content. For
easy understanding, the fields are organized in: symptoms,
risk factors, family history, exposure to agents and
pathologies. At the end of the form there is a button to make
the diagnosis.
b) AgenteDeInterfaz.aspx.cs: this file recovers the values
of the fields selected by the user and stores them in a web
session (a web session is a technical resource for keeping
information between different web pages). Then the control
flow of the program redirects to the next functional element
(the Filter).
2. Filter:
a) AgenteDeFiltrado.aspx:
this file does not show
anything. It is needed due to structure of the application.
b) AgenteDeFiltrado.aspx.cs: this file creates a new object
for managing the bayesian network. At this point we can
comment that the bayesian network (see Fig. 6) is a file called
‘diagnosticoCancer.xdsl’ (this file is created previously with
the Genie´s software and represents the medical knowledge
of the system) and stored in the root directory of the
application. After creating the bayesian network
management object we read the network and now we have
access to its values. Then we recover the values stored in the
web session and we check if the fields have been filled in. If
a field is void we set neutral values of likelihood depending
on the type of field. We have three types: with two options
(0.5), with three options (0.3333) and with four options
(0.25). When all checks are finished we write the new values
of the network and the control flow of the program redirects
to the next functional element (the Output).
3. Output:
a) AgenteRedBayesiana.aspx: this file shows the final
results of the diagnostics. It points out through labels: the
positive/negative probability of pancreatic cancer and the
positive/negative probability of lung cancer.
b) AgenteRedBayesiana.aspx.cs: this file is responsible for
the calculations of probability that express the diagnosis.
First of all it makes the same work as ‘AgenteDeFiltrado.cs’.
But, instead of doing the checks, it sets the values of
evidence of the network (for a better understanding see the
reference [27]). If a value has not been selected we set some
default one. Then we update the beliefs of the bayesian
network. And now we make up the calculations of
probability with the two final nodes (pancreatic cancer and
lung cancer, see Fig. 6) and show the results to the user.
D.Efficiency analysis
In computational complexity theory, big O notation is
often used to describe how the size of the input data affects
an algorithm's usage of computational resources (usually
running time or memory).
The application uses simple instructions and merely
carries out checks on the entry of information. As already
commented, application filters those data that are not
introduced and provides default values. The complexity of
these actions is minimal: O(1).
However, the most complex actions are the two ‘for’ loops
that estimate the probability of suffering pancreatic or lung
cancer. Each loop has a complexity of O(n).
So, the complexity of the application is O(n), technically it
is said linear complexity. Hence we can say that the
application is quite efficient.
Of course this analysis only covers our application. For a
complete estimate, calculations of Genie´s software on the
network should be considered.
For more information read information related to the
Bayesian networks, inference, the Bayes theorem ... We
suggest as an introduction the following article in the
reference [27].
V.CONCLUSION AND FURTHER WORK
nformation and communications technologies are an
essential tool in the field of health sciences both in
research and management. The health system will work
better if we introduce them gradually.
I
This project would produce benefits to pancreatic cancer
research (the project is useful as a basis of a Bayesian
network more realistic) as well as in a more agile working of
medical consultations.
For future developments it would be a chance to add to the
Bayesian network results of diagnostic tests (images,