Volume Full-text Down load - ijabme.org
IJABME
ISSN 1906 - 4063
International Journal of
Applied Biomedical Engineering
Vol. 2, No. 2, July-December 2009
Editorial Board ……………………………………………………………………………
Message from the Editor-in-chief ……………………………………………………… iii
_______________________________________________________________________
INVITED PAPERS
Effect of Transcranial Magnetic Stimulation (TMS) on Visual Search Task
……………………………………………………………………………………………… K. Iramina 1
Development of Virtual Palpation System using Ultrasonic Elastography
………………………………………………………………………………………… K. Hamamoto 9
_______________________________________________________________________________
REGULAR PAPERS
Metal Guarder Could Prevent the Spread of Tissue Desiccation: A Preliminary in Vitro
Double Blinded Study
…………………………… K. Khampitak, T. Khampitak, S. Taechajedcadarungsri, K. Seejorn 14
Energy Packing Efficiency Based Threshold Level Selection for DTW ECG Compression
………………………………………………………………………… Y. Chompusri and S. Yimman 19
No Visual Mismatch Negativity (MMN) for Simultaneously Presented Audiovisual Stimuli:
Evidence from Human Brain Processing
…………………………..…………………………………………………………… W. Sittiprapaporn 29
A Diagnosis of Tonsillitis using Image Processing and Neural Network
………………………………………………………………… A. Leelasantitham and S. Kiattisin 36
Treatment Planning Technique Applying the Combined CT Imaging and Computational
Fluid Dynamics (CFD) Analysis
………………….……. K. Hemtiwakorn, N. Phoocharoen, V. Mahasittiwat and M. Sangworasil 43
Analysis of Quartz Crystal Microbalance Sensor Array with Circular Flow Chamber
.… ………… K. Jaruwongrungsee, T. Maturos, P. Sritongkum, A. Wisitsora-at, M. Sangworasil
and A. Tuantranont 50
Muscular-Contraction Classification : Comparison Study Between Independent Component
Analysis and Artificial Neural Network
…………………………. D. Sueaseenak, T. Chanwimalueang, W. Iampa, and M. Sangworasil 55
ii
_______________________________________________________________________________
Manuscript Submission Guideline
_______________________________________________________________________________
http://www.ijabme.org
COPYRIGHT
Thai Biomedical Engineering Society (ThaiBME)
International Journal of Applied Biomedical Engineering (IJABME)
International Journal of Applied Biomedical Engineering, Vol.2, No. 2 2009
i
International Journal of Applied Biomedical Engineering (IJABME)
Vol. 2, No. 2 (July-December 2009)
Editor in Chief:
Chuchart Pintavirooj, King Mongkut’sInstitute of Technology Ladkrabang (KMITL)
Associate Editors:
Somsak Choomchuay
Supaporn Kiattisin
Supareak Janjarasjitt
Supan Tungjitkusolmun
King Mongkut’sInstitute of Technology Ladkrabang (KMITL)
University of the Thai Chamber of Commerce (UTCC)
Ubonratchathani University (UBU)
King Mongkut’s Institute of Technology Ladkrabang (KMITL)
Editorial Board:
1. Somkiat Wattanasirichaigoon SWU, Thailand
2. Tohru Yagi TIT, Japan
3. John G. Webster Univ. of Wisconcin, USA
4. Ratko Magjarevic IFMBE, Croatia
5. Tsuyoshi Shiina Univ. of Tsukuba, Japan
6. Fernand S. Cohen Drexel U., USA
7. Manus Sanworasil KMITL, Thailand
8. Chusak Limsakul PSU, Thailand
Board of Reviewers:
1. Kazuhiko Hamamoto Tokai Univ., Japan
2. Kosin Chamnongthai KMUTT, Thailand
3. James Goh NUS, Singapore
4. Ian Thomas KKU, Thailand
5. Takafumi Suzuki U. of Tokyo, Japan
6. Kochi Ito Chiba Univ., Japan
8. Shozo Kondo Tokai Univ., Japan
9. Vitoon Leelamanit PSU, Thailand
10.Suradej Tritriluxana
KMITL, Thailand
11. Willis J. Tompkins Univ. of Wisconcin, USA
12. Olivier Adam France
13. Adisorn Leelasantitham UTCC, Thailand
14. Arthorn Sanpanich MU, Thailand
15. Chissanuthat Bunluechokchai KMUTNB, Thailand
16. Mana Sriyudthasak CU, Thailand
17. Sinchai Kamolpivong PSU, Thailand
18. Siridech Boonsang KMITL, Thailand
19. Nuttaporn Pimpha NanoTEC, Thailand
20. Pasin Israsena NECTEC, Thailand
21. Patamaporn Sripadungtham Thailand
22. Phensri thongnopnua CU, Thailand
23. Pornchai Phukpattaranont PSU, Thailand
24. Thurdsak Leuwhathong KMITL, Thailand
25. Saowapak Sotthivirat NECTEC, Thailand
26. Supot Sookpotharom Bangkok Univ., Thailand
27. Warakorn Charoensuk MU, Thailand
28. Watcharachai Wiriyasuttiwong SWU, Thailand
29. Surapan Airphaiboon KMITL, Thailand
30. Wongwit Senawong SWU, Thailand
ii International Journal of Applied Biomedical Engineering, Vol.2, No. 2 2009
Message from the Editor-in-Chief
The IJABME has increasingly drawn attention from many biomedical engineers. As editor-in-chief, I
would like to express my sincere thanks to all the researchers who have submitted their good research
works to be shared and their research result to be exchanged in the current issue of IJABME. Of course,
the success of IJABME would not have been possible without the support and dedication of our
Editorial Board including many colleagues, friends, and professor who have reviewed the manuscripts.
To all of those involved I give my sincere thanks for jobs well done.
Moreover, I would like to take this opportunity to welcome our next guest editor, Professor Kazuhiko
Hamamoto, a distinguish Tokai-University professor in Biomedical Engineering of IEEJ societies and our
long friend and colleague since 2000. The next issue will be distributed in BMEiCON2010 held in Kyoto
University, Japan. I hence encourage all biomedical engineers to wrap-up his/her research work and
prepare the manuscript submitting to IJABME by the end of April. It is a good opportunity to have
their research work shared Asian wide.
Chuchart Pintavirooj, Ph.D.
Department of Electronics
Faculty of Engineering
King Mongkut's Institute of Technology Ladkrabang, Thailand.
E-mail: kpchucha@kmitl.ac.th
Chuchart Pintavirooj was born in Bangkok, Thailand in 1962. He received the B. Sc. (Radiation
Techniques) and M.Sc. (Biomedical Instrumentation) from Mahidol University, Bangkok, Thailand in
1985 and 1989 respectively. In 1995, he received another master degree in Biomedical Engineering
from Worcester Polytechnic Institute, MA, USA. In 2000, he earned a Ph. D. in Biomedical
Engineering from Drexel University, Philadelphia, PA. After working as a research scientist at
Biomedical Instrumentation Department, Mahidol University, he joined Electronic Department,
Faculty of Engineering, King Mongkut’s Institute of Technology at Ladkrabang, Bangkok where he
is currently an associate professor. His current research is in Biomedical Image/ Signal Processing
majoring in Image reconstruction, Image Classification and Image restoration.
Dr. Pintavirooj is the acting chairman of Biomedical Engineering Society of Thailand affiliated with
IFBME.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 1
Effect of Transcranial Magnetic
Stimulation(TMS) on Visual Search Task
K. Iramina, Guest member
ABSTRACT
Transcranial magnetic stimulation (TMS) has
been applied as an important method to investigate
human cognitive process. In this study, we applied
TMS to investigate temporal aspect of the functional
processing of the visual search in the brain. Subjects
were required to respond as quickly and accurately as
possible by pressing a mouse button to indicate the
presence or absence of the target, and the reaction
times were measured. Subjects received four trials
which the TMS stimulus onset asynchronies (SOA)
were set as 100, 150, 200 and 250 ms after visual
stimulus presentation for easy feature, hard feature
and conjunction task, respectively. It was found that
there was a significant elevation in target-present response
time when the TMS pulses were applied 150
ms after visual stimulus presentation. However for
the other SOA cases there was no significant difference
between no-TMS and TMS conditions. Therefore,
we considered that the right posterior parietal
cortex was involved in visual search at about 150 ms
after visual stimulus presentation. When the TMS
applied to the primary visual area on occipital area
50 ms after visual stimulus presentation, there was a
significant decrease in response time. However, there
was no difference at 150 ms SOA.
Keywords: Transcranial Magnetic Stimulation
(TMS), visual search task, visual perception, posterior
parietal cortex(PPC)
1. INTRODUCTION
Transcranial magnetic stimulation (TMS) is the
application of a brief magnetic pulse or a train of
pulses to the skull, which results in an induction of
local electric current in the underlying surface of the
brain, thereby producing a localized axonal depolarization.
TMS has become an important tool for the
study of the functional organization of the human
brain [1]. Earier studies using transcranial magnetic
stimulation were related to the mapping of the cerebral
cortex. Mapping studies were carried out by a
method of localized and vectorial magnetic stimulation
using a figure eight coil [2]. The basic principle
is to concen-trate induced eddy currents locally
Manuscript received on December 16, 2009.,
K. Iramina, Kyushu University, 819-0395 Fukuoka, Japan
Telephone & Fax: +81-92-802-3581
E-mail addresses: iramina@inf.kyushu-u.ac.jp
near a target by a pair of opposing pulsed magnetic
fields produced by a figure-eight coil. This method
facilitates stimulation of the motor cortex of the human
brain within a 5 mm resolution [3][4]. As a noninvasive
and effective method to make reversible lesions
in the human brain, TMS has a long and successful
history. Now it has become a major tool of
cognitive neuroscience and a treatment method for
various neurobehavioral disorders. TMS can be used
to stimulate the cortex to produce visual percepts[5]
or movements[6]. It is more suitable to provide reversible
disruption of activity in the cortex with millisecond
accuracy. Because TMS has high temporal
and spatial resolutions, it can investigate not only
the spatial localization but also the time course of
mental processes. TMS is a noninvasive and effective
method to making reversible lesions in the human
brain. The application of TMS in the investigation
of neurological deficits provides an important method
for investigating human cognitive processes.
Visual search, as a traditionally visual neglect sensitive
measure, was studied by many researchers. Selective
spatial attention [7], memory for the target [8],
object-based atten-tion and identification of the target
are all required in visual search. Much is already
known about the involvement of the large areas of
the visual field [9] [10], the right parietal cortex [11],
right superior temporal gyrus [12] and right posterior
parietal cortex [12] [13].
In contrast to the knowledge of the location of particular
cortex areas involved in the visual search, the
study of temporal aspect of cortex areas is not sufficient.
Ashbridge et al. [14] applied TMS over the subjects’
right parietal visual cortex while they were performing
feature and conjunction visual search tasks.
The same authors found that TMS had no detrimental
effect on the performance of feature search, but
significantly increased the response time in conjunction
search when TMS was applied over the right
parietal cortex 100 ms after the visual stimuli presentation.
However, in that study, the temporal aspect
of the posterior parietal cortex (PPC) in visual
search was not discussed. In Ellison et al.’s repetitive
TMS (rTMS) study [12], they found that TMS
over the right PPC caused a significant increase in
response time in the landmark task and hard conjunction
search task. However, in their study, since
the rTMS was applied before the visual search stimuli
presentation, the involvement of the PPC in visual
search was confirmed but the time course was not in-
2 K. Iramina: Effect of Transcranial Magnetic Stimulation(TMS) on Visual Search Task (1-8)
vestigated. Fuggetta et al.’s study [15] reported a
significant increase in response time for conjunction
task when a single-pulse TMS applied over right PPC
100 ms after the visual stimuli presentation. Ellison
et al. [16] applied 5 pulse 100 ms apart TMS over
right PPC, they found PPC plays significant effect in
conjunction visual search proc-essing.
In this study, to investigate the temporal aspect
of the posterior parietal cortex involved in the visual
search, we used different TMS stimulus onset asynchronies
(SOA) and measured the visual search response
times. The relationship between the SOA and
the response time was investigated. We also investigated
the effect of TMS on primary visual area V1 on
visual search task.
slash (/). In the ’hard feature task’ the target was
a 90 ◦ counterclockwise rotated ‘L’ shape among ‘L’
shape. These two tasks which have only one different
feature between the target and distractor are called
‘pop-out’ tasks. On the other hand, the task which
has more than two different features is called ‘conjunction
task’. In this study, the target was a red
backslash among red and green slash. All items were
subtended in 4.7×4.7 ◦ visual angle on the screen.
Subjects were required to respond as quickly and as
accurately as possible by pressing a mouse button to
indicate the presence or absence of the target. The
target was present on 50% of the trials and there was
never more than one target. Each trial was preceded
by a central fixation cross for 1500 ms, followed immediately
by the stimuli array which was presented
for 1500 ms.
2.2 Magnetic stimulation
The TMS stimulator was a MagStim Super Rapid
Stimulator. Stimulus strength was set as the subjects’
individual resting motor threshold which required to
produce MEPs of >50 µV peak-to-peak amplitude in
at least 6 of 10 successive trials. A figure-of-eight
70mm coil was used. The double coil windings in
the figure-of-eight coil carry two currents in opposite
directions. At the central point of the coil, the
two loops meet and generate a localized summation
of current. The localized electric current induced by
the pulse has low attenuation and high spatial resolution.
When the TMS coil is discharged, a loud click
sound will be produced and this makes the interpretation
of TMS-evoked activity difficult. To investigate
the TMS effect without including auditory artifact,
a control study, no-TMS conditions were applied in
this study. No- TMS condition used two coils. One is
a dummy coil which does not produce the magnetic
field and another coil which are located vertically to
the scalp produces the click sound as shown in Fig.2.
Fig.1: Examples of visual search stimuli (Top: easy
feature, Middle: hard feature, Bottom: conjunction,
Left: target-absent stimulus, Right: target present
stimulus)
2. MATERIALS
2.1 Visual search task
Three visual search tasks were carried out (see Fig.
1). There were eight items in each array. In the ‘easy
feature task’ the target was a backslash (\) among
2.3 Subjects
Eight subjects (one female, seven males, and aged
21-31 years) participated in all the experiments of the
present study, who were experienced psychophysical
observers but ignored the purpose of the experiment.
All were right handed, had normal or corrected-tonormal
visual acuity. All the subjects had previous
practice before the experiment in order to be familiar
with visual search and to make stable response, and
reported a complete absence of epilepsy, or any other
neurological condition in themselves and their known
family history.
3. EXPERIMENTAL PRCEDURE
The time sequence of experiment is shown in Fig.3.
Each trial was preceded by a central fixation cross
for 1500 ms, followed immediately by a visual search
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 3
Fig.2: TMs coil was applied over the right posterior
parietal cortex. (a) TMS condition (b) No-TMS
condition
stimulus, which would be presented for another 1500
ms. Subjects were required to respond as quickly and
as accurately as possible by clicking a mouse button
to indicate the presence or absence of the target (left
button for target present and right button for target
absent). The time from the visual search stimuli
presentation till the button click was recorded as the
response time. Transcranial magnetic stimulations
were applied over the right PPC of the subject at
different time intervals after the visual search stimuli
presentation. These time intervals were called as
TMS stimulus onset asynchronies (SOA).
Fig.3:
No-TMS condition
In the TMS condition, since paired-pulse TMS can
produce larger and longer effects on cortical activity
than single-pulse, paired-pulse TMS was used in the
present study. The highest frequency of TMS device
is 50 Hz, thus, we set inter-stimulus interval=20 ms.
Paired-pulse TMS were applied over the right PPC of
subject. The coil was placed tangential to the surface
of the skull and the centre of the coil was positioned
over electrode site P4 of the international 10-20 system.
The induced current in the brain was parallel
to the Oz-Pz midline, flowed from the occipital
to the frontal cortex. While in the no-TMS condition,
an open circuit coil was put over the subject’s
right PPC just like the TMS condition, while a figureof-eight
coil discharged paired-pulse near to, but directed
away from the subject’s skull. The application
sequences of TMS and no-TMS conditions were randomized
across subjects. Each subject received 60
trials for two times under TMS and no-TMS conditions
as one test. Subjects received four tests with the
TMS stimulus onset asynchronies (SOA) were set as
100, 150, 200 and 250 ms after the visual search stimuli
presentation. Subjects received these four tests
for the easy feature and hard feature search tasks,
respectively. Consequently, subjects received a total
of eight tests in this study. The stimulation protocol
was in accordance with published safety recommendations.
In considering of the safety of TMS, only one
test was performed within one day for each subject.
4. RESULTS
4.1 Visual search task stimulating on the right
PPC
For the easy feature search task, hard feature task
and conjunction task, the target-present response
times at each SOA for each subject are shown in the
Table 1. Obviously, the individual differences among
subjects were large. Moreover, there were differences
in response times between SOA cases even for the
same subject. Since subjects participated all the 4
tests in different days, the mental condition, concentration
and relaxation degrees were different even for
the same subject. This logically led to difference in
the response time. Therefore, it is not appropriate
to use the average of response time of each subject
as the typical value of the experimental data. In the
present study, in order to remove the individual difference
and the difference of experimental conditions, we
used the average of normalized response time as the
typical value of the experimental data. Thus, for each
subject, the average of target-present response time
in the TMS condition was normalized to the no-TMS
condition (set as the baseline=1). The difference between
the average of normalized response time of all
the subjects and the baseline was taken to demonstrate
the TMS effect. Fig.4 shows the target-present
response times and normalized response time at each
SOA cases. In order to investigate the influence of
TMS effects, paired t-test was used to analyze the
difference between the TMS and no-TMS conditions.
4 K. Iramina: Effect of Transcranial Magnetic Stimulation(TMS) on Visual Search Task (1-8)
Table 1: Respose times at each SOA for the easy
feature task, hard feature task and conjunction tasks
of each subjects
We found that, when SOA=150 ms, compared to the
no-TMS condition, there was a significant elevation
in response time when the TMS pulses were applied.
However, for the other SOA cases, there was no significant
difference between the TMS and no-TMS conditions.
For the hard feature search task, the targetpresent
response times and the difference between the
average of normalized target-present response time of
all the subjects is shown in Fig.5. Based on the paired
t-test analysis results, we found that, when SOA=150
ms, compared to the no-TMS condition, there was a
significant elevation in response time when the TMS
pulses were applied. However, for the other SOA
cases, there was no significant difference between the
TMS and no-TMS conditions. For the conjunction
task, the target-present response times and the difference
between the average of norma-lized targetpresent
response time of all the subjects and is shown
in Fig.6. When SOA=150 ms, compared to the no-
TMS condition, there was an elevation in response
time when the TMS pulses were applied. For the
target-absent condition, there was no significant difference
between the TMS and no-TMS conditions for
all three visual search task and the SOA cases.
Fig.4: (a) Target-present response times (b) Normalized
target-present response times at each SOA for
the easy feature visual search task when the right PPC
was stimulated.
4.2 Visual search task stimulating on the right
V1
In order to investigate the difference between the
role of right PPC and the primary visual cortex V1,
TMS was applied to the right V1 at 50ms SOA, 100ms
SOA and 150ms SOA. For the hard feature search
task, the tar-get-present response times and the difference
between the normalization of response time
ratio TMS to no-TMS condition is shown in Fig.7.
When the TMS applied to the right V1 after visual
stimulus presentation at 50 ms SOA, there was a significant
decrease in response time. However, there
was no difference at 150 ms SOA.
4.3 Simple reaction task
In this experiment, we investigated the simple reaction
time. No-targeted hard feature visual search
task was used in this experiment. The subjects re-
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 5
Fig.5: (a) Target-present response times (b) Normalized
target-present response times at each SOA for
the hard feature visual search task when the right PPC
was stimulated.
Fig.7: (a) Target-present response times (b) Normalized
target-present response times at each SOA for
the hard feature visual search task when the right V1
area was stimulated.
Fig.6: (a) Target-present response times (b) Normalized
target-present response times at each SOA for
the conjuncion visual search task when the right PPC
was stimulated.
Fig.8: (a) Response times (b) Normalized response
times at each SOA for the simple reaction task when
the right V1 area was stimulated.
6 K. Iramina: Effect of Transcranial Magnetic Stimulation(TMS) on Visual Search Task (1-8)
Fig.9: (a) Response times (b) Normalized response
times at each SOA for the simple reaction task when
the right PPC was stimulated.
spond the visual stimulation sa soon as possible without
judgement of the presenttation. Stimulus points
of TMS were right PPC or right V1. Fig. 8 shows
the normalization of response time ratio TMS to no-
TMS condition of all the subject when the right V1
was stimulated at 50 ms or 100 ms SOA. Fig. 9 shows
the normalization of response time ratio TMS to no-
TMS condition of all the subjects when the right PPC
was stimulated at 50 ms or 150 ms SOA. When the
right V1 was stimulated at 50 ms SOA, there was a
significant decrease in response time. However, SOA
is 100 ms, there was no difference. When the right
PPC was stimulated at 50 ms SOA, there was a significant
decrease in response time. However, SOA is
100 ms, there was no difference.
5. DISCUSSION
Two previous studies have investigated the involvement
of the PPC in visual search task using
TMS[12, 14]. However, the former used repetitive
TMS (rTMS) before visual search stimuli presentation.
That study indicated that the PPC is involved
in visual search, however the temporal aspect of the
PPC involved in visual search was not discussed. The
other one used single pulse TMS but reported TMS
over the PPC had no detrimental effect on the performance
of feature search. In the present study, to
investigate the temporal aspect of the PPC involved
in feature search, a paired-pulse TMS was applied
over the right PPC, and the time intervals between
visual search stimuli presentation and TMS stimuli
were set as four types, i.e., SOA=100, 150, 200 and
250 ms. Several past TMS experiments on the PPC
showed that TMS over the PPC does not induce eye
movement [14, 17]. Furthermore, some past studies
suggested that the PPC plays a role in the saccadic
eye movement beyond 200 ms after target presentation[18].
In our study, we found a significant influence
in response time when SOA=150 ms; Based on
the above-mentioned studies, we considered that the
saccadic eye movement was not bound at all at 150
ms after visual search task presentation. This lead to
the conclusion that the disruption of eye movement,
which is induced by the TMS, is not the key factor
for the elevation in response time when SOA=150 ms.
A common knowledge of TMS experiments is that
TMS can either excite the cortex or disturb its function.
The observed excitatory effects are normally
muscle twitches or phosphenes, whereas in the “lesion”
mode, TMS can transiently suppress perception
or interfere with task performance. An appropriately
applied TMS pulse delivered in time and space can
transiently disrupt the function of a certain cortical
area at a given time, creating a temporary and localized
cortical “virtual brain lesion”. Thus, TMS can
be used to disrupt the brain activation at a specified
cortical area at a given time point. Therefore, TMS
can be used to investigate not only the spatial but also
the temporal characteristics of task performance. In
the present study, compared to the sham TMS condition,
since a significant elevation in response time
was only observed when SOA=150 ms, we concluded
that the feature search was probably processed in the
PPC at about 150-170 ms after stimulus presentation.
Based on the past conjunction search research[14] and
present study, it seems reasonable to consider that
the PPC plays a dominant role not only in the conjunction
search but also in the feature search tasks.
On the other hand, for the target-absent condition,
it was found that there was no significant difference
between TMS and sham TMS conditions for all the
SOAsettings in both easy and hard feature search
tasks. Based on the actual experimental experience,
we considered that the response of “target present”
or “target absent” is different to that of “where is the
target” in the visual search task. i.e., detection of
target and localization of target are two different processes.
When the primary visual area V1 was stimulated
in the hard feature task, response time shows
the significant decreasing in the case of 50 ms SOA
and 100 ms SOA. For the simple reaction experiment,
when the primary visual area V1 was stimulated, response
time shows the decreasing significantly in the
case of 50 ms SOA. There is no difference in 100 ms
SOA. When the right PPC was stimulated, There
was no significant difference of response time in the
case of 150 ms SOA. These results suggest that right
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 7
PPC plays a dominant role in spatial processing during
visual search. This TMS study opens up a new
possibility for understanding not only the physiological
role of the PPC in the visual search but also the
temporal aspect of the PPC involved in the visual
search. We expect that this study will contribute to
the theories about the visual search dynamics.
6. CONCLUSION
We concluded that it is most difficult to find the
target in the conjunction task. Furthermore, compared
with easy feature task, we found that it is
more difficult to find the target in the hard feature
task. In easy and hard feature tasks, when SOA=150
ms, compared to no-TMS condition the response time
of the TMS condition were increased. Especially in
the hard feature task, there is a significant difference
between TMS and no-TMS conditions. However, in
the other SOA cases of all tasks, we could not found
such TMS effect. Paired t-test was used to examine
whether there is a significant difference between TMS
and no-TMS conditions or not. There was a significant
difference between TMS and no-TMS conditions
when SOA=150 ms in easy feature tasks, hard feature
tasks (p
8 K. Iramina: Effect of Transcranial Magnetic Stimulation(TMS) on Visual Search Task (1-8)
K. Iramina received his Ph.D degree
in Engineering from Kyushu University,
Japan, in 1991. From 1991 to 1995,
he worked at the Department of Electronics,
Kyushu University. In 1996, he
moved to the University of Tokyo, Japan
and he has been an associate professor of
Institute of Biomedical Engineering, the
University of Tokyo. In 2005, he moved
o Kyushu University as a professor of
Graduate School of Information Science
an Electrical Engineering. He has been also a professor of
Graduate School of Systems Life Sciences in Kyushu University.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 9
Development of Virtual Palpation System
using Ultrasonic Elastography
K. Hamamoto, Guest member
ABSTRACT
“Virtual palpation system”, which can present a
tactile sense as user palpates a patient in virtual space
based on the patient’s elasticity information obtained
by an ultrasonic elastography is proposed. This system
uses 3-dimentional haptic rendering technique.
In this research, it is attempted to determine parameters
which are used to realize the same tactile sense
as one in actual palpation. The parameters are determined
experimentally using some breast phantom.
Keywords: Virtual reality; Palpation; Elastography;
Ultrasonic imaging; Haptic rendering
1. INTRODUCTION
ULTRASONIC imaging is widely applied and indispensable
technique in medical diagnostic field currently.
It has many advantages compared with X-
ray imaging, MRI and so on. Unfortunately, however,
it has a serious weak point, where ultrasonic
echo image is not quantitative. Therefore, there are
many researches about quantitative image construction,
which is called “Tissue Characterization”.
Ultrasonic elasticity imaging is one of the most
important tissue characterization researches recently.
HITACHI Medical Corporation has already produced
the technique on a commercial basis and the equipment
is mainly used to decide whether a tumor (for
example, in breast) is benign or malignant [1]. The
newest equipment enables to display elasticity image
in real time. As the technology advances, medical
doctors and ultrasonographers are expecting to display
the elasticity information as not only a visual
image on 2D monitor but “the sense of touch”. That
is to say, they expect “virtual palpation system”. Virtual
Palpation System enables medical doctor to get
the sense of touch of tissue even if it is deep seated
tumor. This system is provided as man-machine interface
between medical doctor and real-time tissue
elastography.
In recent years, virtual reality technique is applied
to medical field, especially, virtual surgery simula-
Manuscript received on December 29, 2009.,
Kazuhiko Hamamoto, Department of Information Media
Technology, School of Information and Telecommunication
Eng., Tokai University, 2-3-23 Takanawa, Minato-ku, Tokyo,
108-8619 Japan Telephone: +81-3-3441-1171 Fax: +81-3-3447-
6005
E-mail addresses: hama@keyaki.cc.u-tokai.ac.jp
tor [2][3]. And I have already introduced virtual reality
application for medical ultrasonic diagnosis in
ref[4]. In ref[4], virtual palpation system has been
also introduced. The virtual palpation system uses
PHANToM DeskTop TM , produced by Sensable Technologies
Inc, which is the most popular haptic device.
This device presents the sense of touch (reaction
force) on a point in virtual space with the point’s
elasticity by Spring-Damper theory. That means a
technique is strongly required which can calculate
accurate elasticity on the pressed point considering
surrounding three-dimensional elasticity information
and the depth of the press. This process is called
“haptic rendering”. This rendering process is very
important because human body is not homogeneous
and has non-linear characteristic. Three-dimensional
finite element method is usually used for the calculation
to get strict accuracy of human haptic sense
[5]. However, it takes high computational cost and
cannot realize a real-time virtual palpation system.
In this research, volume rendering technique in
Computer Graphics has been applied to haptic rendering
[6][7]. It enables to reproduce the human haptic
sense of real palpation approximately and easily
in real time without high computational cost. Of
course, finite element method, etc may be needed
for ideal and precisely accurate virtual palpation system.
However, they take high computational cost and
cannot provide real time system. In this proposed
method, real time system can be provided and it enables
to reproduce real haptic sense approximately
by setting some appropriate parameters. The parameters
are decided from phantom experiments.
In this paper, the setting of parameters is investigated.
Four phantoms which have different
elasticity are made from vinyl chloride material or
polyurethane material. The parameters are set as the
sense of touch in virtual space becomes the same as
the sense of touch of the phantom. Whether the parameters
are unique for the virtual palpation system
or depend on users’ characteristics are investigated. If
the parameters depend on users’ characteristics, calibration
process for each user might be needed. Such
problem for practical use is also investigated.
2. PROPOSED HAPTIC RENDERING
2.1 Transform of Elasticity for PHANToM
In this process, elasticity information which is
measured quantitatively is transformed. This pro-
10 K. Hamamoto: Development of Virtual Palpation System using Ultrasonic Elastography (9-13)
cess enables the elasticity information to be used in
PHANToM DeskTop TM . A parameter used in this
process is called “Elasticity Information Parameter
(EIP)”. This parameter is set by a user according to
the real sense of touch. A user adjusts EIP as the
virtual sense of touch becomes the same as the real
sense of touch.
2.2 Transform of Elasticity for depth
Reaction force which is felt when a user presses a
surface depends on the depth of the press. In addition,
elasticity information which is seated closely to
the surface affects the reaction force significantly and
effect of elasticity information which is seated deeply
to the reaction force is small.
In this process, elasticity information is transformed
according to the depth of the press by eqn.(1).
Fig.1: A phantom with thin strings embedded in a
lattice for investigation on effectiveness of surrounding
elasticity information of the pressed point.
E ′ (x, y, d) =
E(x, y, d)
p d (1)
d : the depth of the press
E ′ (x, y, d) : transformed elasticity at depth d
E(x, y, d) : elasticity at depth d after II(A) process
p : depth parameter (DP )
Fig.2:
A part of Displacement Filter
DP is also set by a user in the same way as EIP.
2.3 Displacement Filter
The reaction force is affected by not only elasticity
information along the direction of the press but
the surrounding elasticity information of the pressed
point. The effectiveness of the surrounding elasticity
information is investigated by using a phantom made
from vinyl chloride material whose elasticity is close
to one of human body. Some thin strings are embedded
in the rectangular prism phantom in a lattice
shown in Fig.1(a). The deformation of the strings
is measured when the phantom is pressed shown in
Fig1(b). The effectiveness is determined by the deformation
experimentally. The estimated effectiveness
is shown in Fig.2. Z direction is the direction of the
press, and the left end column of the top row is the
position of the press. This effectiveness table is called
“Displacement Filter (DF)”.
DF consists of 2 filters. One is for x-z plane and
the other is for y-z plane. That is to say, DF is expressed
in DF(x, z) and DF(y, z). If the depth of the
press is “4”, 4 rows and 5 columns of DF are used.
The accurate elasticity E” on the pressed point considering
surrounding three-dimensional elasticity information
and the depth of the press for PHANToM
DeskTop TM is calculated in eqn.(2).
E ′′ =
4 ∑
4∑
|x|=0 |y|=1 z=0
d∑
{E ′ (x, z) • DF (x, z)
+E ′ (y, z) • DF (y, z)}
(2)
E ′′
: elasticity at the position of the press
for P HANT oM Desktop T M
2.4 Calculation of the reaction force
Finally, the reaction force F presented by PHAN-
ToM DeskTop TM is calculated in eqn.(3).
F = E ′′ × d (3)
This haptic rendering method can produce the accurate
touch of the sense approximately by controlling
parameters, EIP and DP in real time.
3. EXPERIMENTS
3.1 Experimental condition
It is required to set two parameters, EIP and DP
properly to present the touch of sense for virtual
palpation system, which is the same as one of real
palpation. In this paper, the setting of EIP and
DP are investigated experimentally. In this experiment,
PHANToM DeskTop TM is used as haptic device,
which is shown in Fig.3
First, 4 phantoms whose elasticities are different
each other are made. Next, a subject of the experiments
touches the phantoms by a stylus which
is similar to one of PHANToM DeskTop TM . At a
time, the subject touches virtual object by PHAN-
ToM DeskTop TM . The subject adjusts EIP and DP
as the virtual sense of touch becomes the same as the
real sense of touch of the phantoms. Results obtained
by some subjects are compared and investigated.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 11
3.4 Experiment-1
In this experiment, both EIP and DP are variable.
The number of subject is 15 which consist of 10 males
and 5 females in 20 years old generation.
The result is shown in Fig.5. The right side longitudinal
axis shows DP and the left side one shows
EIP.
Fig.3:
PHANToM DeskTop TM
3.2 Phantoms for experiment
Four kinds of phantom are made for this experiment.
The specifications are shown in Table 1.
Fig.4 shows a comparison of elasticity between the
phantoms and tissues of female breast. Phantom1,2
and 3 correspond to normal tissue or benign tumor.
Phantom4 corresponds to malignant tumor. Therefore,
it is important problem that a subject can distinguish
phantom4 from phantom1,2 and 3 by the
proposed virtual palpation system or not.
Fig.5: The right side longitudinal axis shows DP
and the left side one shows EIP.
Fig.6:
DP setting in the case where EIP is fixed.
Fig.4: Comparison of elasticity between phantoms
and female breast tissues.
3.3 Virtual object
The size of virtual object is 34×34 in x-y plane
and 18 in depth. The resolution (voxel size) is 2.5mm
cube. The size of DF in depth is 10. That means a
subject can press the virtual object until 25mm in
depth. The resolution (2.5mm) is determined from
a human sense where the distance in which human
sense can distinguish two points is 2-4mm.
All subjects could set EIP and DP as the virtual
sense of touch of PHANToM DeskTop TM becomes the
same as the real sense of touch of phantoms. However,
the averaged value doesn’t correlate to elasticity
of phantoms in the case of both EIP and DP. Furthermore,
the standard deviation of EIP is large. That
means the parameter setting depends on user. This
result causes difficulty to set the parameters and to
realize the proposed virtual palpation system.
3.5 Experiment-2
The reason why such result is obtained in
Experiment-1 is estimated to be high correlation between
EIP and DP and different standard value for
each subject. Therefore, standard value for either
EIP or DP is defined and fixed in Experiment-2. Realization
of the proposed virtual palpation system is
attempted by one parameter setting.
12 K. Hamamoto: Development of Virtual Palpation System using Ultrasonic Elastography (9-13)
Fig.7:
1.168.
EIP setting in the case where DP is fixed in
DeskTop TM becomes the same as the real sense of
touch of phantoms as well as previous experiments.
The result of EIP setting is shown in Fig.7. As the
result shown in Fig.7, EIP setting corresponds to elasticity.
Larger elasticity phantom has, larger EIP setting
user selects. And its variance is also small. This
result suggests that the proposed virtual palpation
system can be realized by EIP setting according to
elasticity measured by ultrasonic elastography without
a parameter optimization (calibration) for each
user.
However, the variance in the case of phantom4 is
significantly larger than others. The border of benign
tumor and malignant tumor is seated between
phantom3 and phantom4 shown in Fig.4. Therefore,
whether this border can be recognized or not is very
important. This concern is investigated by using t-
test.
As the result of t-test, it is confirmed that there
is significant difference between phantom3 and phantom4.
The result of comparison of a confidence interval
in t-test is shown in Fig.8. This result shows that
phantom 3 and phantom4 are obviously distinguished
on the confidence interval in 95%.
Fig.8:
Comparison of confidence intervals in t-test.
3.5.1 The standard value for EIP
In this experiment, EIP is fixed for each phantom
and only DP can be adjusted by a subject. The subjects
consist of 6 males and 2 females in 20 years
old generation. The standard values of EIP for each
phantom are follows:
phantom1 : 0.0215
phantom2 : 0.0303
phantom3 : 0.0521
phantom4 : 0.1093
All subjects could set DP as the virtual sense of
touch of PHANToM DeskTop TM becomes the same
as the real sense of touch of phantoms. The result of
DP setting is shown in Fig.6.
Unfortunately, the DP setting doesn’t correlate to
elasticity of phantoms as well as Experiment-1. However,
since the DP setting tends to be constant value
and not to depend on elasticity and subject, it is estimated
that DP should be fixed in a constant value.
3.5.2 The standard value for DP
In this experiment, DP is fixed in a constant value,
1.168 and only EIP can be adjusted by a subject. The
constant value for DP doesn’t depend on elasticity
and subject. The subjects consist of 6 males and 3
females in 20 years old generation. All subjects could
set EIP as the virtual sense of touch of PHANToM
3.6 Discussion
As the results of these experiments, it is shown
that the proposed virtual palpation system can be
realized by fixed DP and setting EIP from elasticity
information measured by ultrasonic elastography
without dependence on user.
4. CONCLUSION
The virtual palpation system has been proposed.
The system uses elasticity information measured by
ultrasonic elastography and can present the virtual
sense of touch which is the same as the real sense of
touch in practical palpation. In this paper, a new
haptic rendering method is proposed for the virtual
palpation system. The haptic rendering method enables
to realize re-producing elasticity of the point
of press for PHANToM DeskTop TM in real-time with
considering the depth of press and effect of three dimensional
distribution of elasticity information. In
this method, two parameters setting are needed, EIP
and DP. The results of experimental investigations
show that the proposed virtual palpation system can
be realized by fixed DP and setting EIP only from
elasticity information measured by ultrasonic elastography
without dependence on user.
There is remained evaluation of the proposed system
by medical doctor in the future.
References
[1] J. C. Bamber, P. E. Barbone, N. L. Bush, D.
O. Cosgrove, M. M. Doyely, F. G. Fuechsel, P.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 13
M. Meaney, N. R. Miller, T. Shiina & F. Tranquart,
“Progress in Freehand Elastography of
the Breast,” IEICE Trans. Inf. & Syst., Vol.E85-
D, No.1, pp.5-14, 2002.
[2] N. Suzuki, A. Hattori, A. Takatsu, A. Uchiyama,
T. Kumano, A. Ikemoto & Y. Adachi, “Virtual
surgery simulator with force feedback function,”
Proceedings of the 20th Annual International
Conference of the IEEE Engineering in Medicine
and Biology Society, Vol.20, No.3, pp.1260-1262,
1998
[3] M. Nakao, M. Komori, H. Oyama, T. Matsuda,
G. Sakaguchi, M. Komeda & T. Takahashi,
“Haptic Reproduction and Interactive Visualization
of a Beating Heart Based on Cardiac Morphology,”
2001 International Medical Informatics
Association, MEDINFO 2001, eds. V. Patel,
et al., IOS Press. Amsterdam, pp.924-928, 2001.
[4] K.Hamamoto, “Virtual Reality as Human Interface
and its application to Medical Ultrasonic
diagnosis,” International Journal of Applied
Biomedical Engineering, Vol.1, No.1, pp.14-
19, 2008.
[5] S. Sarama, “Virtual Haptics,” Proceedings of the
9th IVR seminar, pp.75-104, 2001.
[6] K. Hamamoto, M. Nakanishi & T. Shiina, “Investigation
on Haptic Display System for Medical
Ultrasonic Elasticity Imaging,” Modelling in
Medicine and Biology VI, Editors : M.Ursino,
C.A.Brebbia, G.Pontrelli and E.Magosso, WIT
PRESS, pp.591-598 Sep. 2005 [Proceedings of
6th International Conference on Modelling in
Medicine and Biology].
[7] K.Hamamoto, “Investigation on Virtual Palpation
System using Ultrasonic Elasticity Imaging,”
Proceedings of the 2006 IEEE Engineering
in Medicine and Biology 28th Annual Conference,
pp.4873-4876, 2006
Kazuhiko HAMAMOTO was born
in Nagasaki prefecture, Japan in 1966.
He received B.D, M.D and Ph.D from
Tokyo University of Agriculture and
Technology in 1989, 1991 and 1994 respectively.
He was assistant professor in
Dept. of Communications Eng., Tokai
University in 1994, Associate Professor
in 1999, and Professor in Dept. of
Information Media Technology, School
of Information and Tele-communication
Eng., Tokai University in 2009 and now. His research interest is
information architecture, especially, medical image processing,
human interface and virtual reality. He joins IEICE, IEEJ,
IEEE, VRSJ, etc. He is a committee member of Technical
Committee of Medical and Biological Engineering, Society of
Electronics, Information and Systems, IEEJ
14 K. Khampitak et al: Metal Guarder Could Prevent the Spread of Tissue Desiccation ... (14-18)
Metal Guarder Could Prevent the Spread of
Tissue Desiccation: A Preliminary in Vitro
Double Blinded Study.
K. Khampitak ∗ , T. Khampitak,
S. Taechajedcadarungsri, and K. Seejorn, Guest members
ABSTRACT
Nowadays the bipolar electro-coagulator is widely
used in gynecologic surgery. A major problem with
using this instrument was the uncontrolled extension
of tissue desiccation, causing organ injuries. This experiment
was designed to study the effect of metal
guarder on the spread of tissue desiccation compared
with the non guarding side. The experiment was divided
into two parts according to method of measurement.
The first was use of fresh raw meats and measurement
by visual inspection which mimicked the
real practice. The ten specimens were desiccated by
5-mm bipolar electro-coagulator (40 W). The mean
spread of uncontrolled tissue desiccation in the metal
guarding side was 1.4 mm. (range 1-3 mm) and non
metal guarding side was 4.9 mm. (range 2-7 mm.)
which the difference was statistically significant by
Wilcoxon Signed-Rank test. The second part was
made in the same way as the previous except for the
animal specimens and measuring method. Ten pieces
of very fresh rat meat, 3 mm thick, were prepared
and desiccated. The spread of tissue desiccation was
measured on histological section. The mean spread of
uncontrolled tissue desiccation in the metal guarding
side was 1.1 mm. (range 0.6-1.5 mm.) and non metal
guarding side was 4.1 mm. (range 2.5-9.1 mm.). The
differences was statistically significant by Wilcoxon
Signed-Rank test. We believe that this result can lead
the surgeon’s technique for the thermal safe electrosurgery
or can be a new concept for innovation of a
safety bipolar electro- coagulator.
Keywords: bipolar coagulation, thermal damage,
electro-surgery.
1. INTRODUCTION
Recently, bipolar electro-coagulator has been introduced
as an electrosurgical vessel sealing device.
Nowadays, it has been widely used in abdominal,
vaginal and laparoscopic hysterectomies with the advantages
of reducing surgical blood loss and operating
* Corresponding author.
Manuscript received on January 8, 2009.,
K. Khampitak, Department of Obstetrics and Gynecology
Faculty of Medicine , Khon Kaen University, Thailand.
E-mail addresses: kovit@kku.ac.th
period [1, 2]. The newly bipolar electro-coagulator
with audible tone indicating complete of the process
is able to seal vessels up to 7 mm. in diameter [3-6].
However, the risk of uncontrolled thermal injury from
high temperatures at the instrument tip has been the
major problems in iatrogenic thermal injury.
The living tissue is slowly heated when the temperature
above 50 ◦ C. If the tissue is heated to 90 ◦ C,
irreversible cellular water completely evaporates (desiccation).
The cell walls are ruptured at 100 ◦ C (vaporization)
and tissue begins to carbonize and char at
250 ◦ C.[7,8]. In the operative field, the uncontrolled
tissue desiccation could spread (above 60 ◦ C) up to 10
mm. from the tip of the instrument [9, 10].
However we have found atraumatic grasping forceps
adjacent to bipolar forceps can minimize thermal
injury during electrosurgery. To investigate this
effect, we designed a study by using fresh animal tissue
and measurement by visual inspection and histological
analysis.
2. MATERIALS AND METHODS
2.1 Experimental design
The experiment was divided into two parts according
to the measurement methods. Each animal tissue
was grasped by Kleppinger bipolar electro-coagulator
and the laparoscopic grasping forceps, which grasped
laterally as experimental side while the other was left
as control side. The data was then subjected to statistical
analysis.
2.2 Procedures
In the first experimental part, the fresh raw meat
with 3 mm. thickness was grasped by Kleppinger
bipolar forceps while laparoscopic grasping forceps
was being grasped laterally. The tissue desiccation
started when 40 W electric power of bipolar electrocoagulator
was turned on. The electricity was run
until no air bubble on the tissue surface (tissue desiccation).
Then, the extension of tissue desiccation was
evaluated by visual inspection using measuring ruler
which mimicked the real tissue inspection in the operative
field. (Figure 1 and 2)
The second part was made in the same way as the
previous except for the animal specimens and measuring
method. Since the visual measurement alone
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 15
Fig.1: A piece of meat was desiccated by bipolar
electro-coagulator (40 W) while grasping forceps were
grasping at the side of the bipolar electro-coagulator
as a metal guarder.
Fig.3: Normal muscle tissue is shown on the right
side of the picture. On the left side, the desiccated
area is obviously eosinophilic with glassy appearance.
Fig.2: The spread of uncontrolled tissue desiccation
in the metal guarding side (arrow) was lesser than
non metal guarding side.
was not sufficient to fully represent the histological
differences, then we decided to measure on microscopically
histological section and used fresh rat meat
as the experimental specimens. After ending of each
one, the tissue was immediately placed into a 10%
formalin-buffered solution. After 24 hours, the samples
were embedded in paraffin, sectioned and stained
with hematoxylin/eosin. Then, the section was evaluated
microscopically for the whole area where thermal
effects were evident-was measured, in mm, using
a built-in stage micrometer calibrated.
The experiment was performed at the Srinagarind
Hospital and the study was approved from the Animal
Care and Use Committee of Khon Kaen university,
Thailand.
2.3 Statistical analysis
On the study design, each specimen was measured
for both control and experiment sides, then it could
Fig.4: The cytoplasm of desiccated muscle cells
(Right Pict.)is glassy and eosinophilia with partial
loss of striation. The nuclei are condensed and deep
basophilic. The normal muscle cells (Left Pict.), on
the contrary, retain striation and detailed nuclei.
prevent intervariable influences that might have induced
a confounding bias. To abolish of personal bias,
we use double blinded technique for our pathologist.
Finally, the spread of uncontrolled tissue desiccation
in the metal guarding side and the non metal guarding
side of each experimental part was statistical analyzed
by Wilcoxon Signed Rank test.
3. RESULTS
Ten pieces of fresh raw meat were desiccated by
Kleppinger bipolar electro-coagulator (40 W) while
the grasping forceps were grasping beside the bipolar
electro-coagulator as a metal guarder. The length
of laterally uncontrolled tissue desiccation was measured
in both sides.
The results of first part were showed in the Table 1.
The mean length of uncontrolled tissue desiccation in
16 K. Khampitak et al: Metal Guarder Could Prevent the Spread of Tissue Desiccation ... (14-18)
the metal guarding side (group 1) 1.4 mm. (range 1-3
mm) and non metal guarding side was 4.9 mm. (range
2-7 mm.). There was significant difference between
the two groups according to Wilcoxon Signed Rank
test.
Ten pieces of fresh rat meat were desiccated in the
same way as the previous The results of laterally uncontrolled
desiccated length were showed in the Table
2. The mean length of uncontrolled tissue desiccation
in the metal guarding side (group 1) was 1.1 mm.
(0.6-1.5 mm.) and non metal guarding side (group 2)
was 4.1 mm. (2.5-9.1 mm.). There was also significant
difference according to Wilcoxon Signed Rank
test.
Table 1: The results of first experimental part,
group 1 was the spread of uncontrolled tissue desiccation
data of the metal guarding side and group was
the data of non metal guarding side, using visual inspection
(measuring ruler).
Number of Group Group
Experiments 1(mm.) 2(mm.)
1 1 4
2 2 5
3 1 2
4 1 6
5 1 7
6 1 5
7 3 5
8 1 5
9 2 5
10 1 5
Statistical significance P< .001.
4. DISCUSSION
The development of laparoscopic surgery has stimulated
substantive research and development in the
medical devices industry for the progression in the
technologies of energized dissection. With the advantages
of reducing surgical blood loss and operating
period, nowadays, electro-coagulator has been widely
used in all routes of surgery, including laparoscopic
hysterectomies. [1, 2] ,[11]
The incidence of iatrogenic complications caused
by laparoscopic electro-surgery is 2 to 5 per 1,000
[12,13]. Most are thermal injuries result from electrocoagulator.
Major of these are caused by the
stray current injuring adjacent structures, such as
the bowel, bladder, and ureter. Bipolar electrocoagulator
can better control this unwanted stray current
[14,15].
The mechanism of bipolar electro-coagulator is to
form vapor pockets coalesce and then become the vaporized
zones at the areas of highest power density
along with a proximal thrombus, causing hemostasis.[16]
However, thermal damage may occur well beyond
their confines. The unabated application of current
can propagate a thermal bloom that disruptively
bubbles steam through the surrounding parenchyma,
damaging tissue at some distance from the target
site.[17] (Figure 5)
Table 2: The result of second experimental part,
group 1 was the spread of uncontrolled tissue desiccation
data of the metal guarding side and group 2 was
the data of non metal guarding side, using microscopically
histological measurement.
Number of Group Group
Experiments 1(mm.) 2(mm.)
1 1.5 3.5
2 0.8 2.5
3 1.2 3.2
4 0.6 2.5
5 1.1 6.1
6 1.2 4.2
7 0.9 9.1
8 1.5 3.5
9 1.2 3.2
10 1 3.1
Statistical significance P< .001.
Fig.5: The mechanism of bipolar electro-coagulator
was to form vapor pockets coalesce and then become
the vaporized zones at the areas of highest power density.
The desiccation caused thermal effect spreading
through the adjacent tissue
A recent study conducted by the Dundee group
on detection of the spread of thermal energy in bipolar
electro-coagulator by using an infrared imaging
system, suggested that a maximum temperature of
150 ◦ C was detected at the instrument tip and a thermal
spread (above 60 ◦ C) was extended up to 10 mm.
from the tip. [7-10]
Remorgida V [18] report that tissue thermal damage
caused by bipolar forceps can be reduced with
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 17
outer plastic shield. However, in practice, plastic
shield can concentrate the heat inside the jaws,
but cannot neutralize the electrical current causing a
thermal bloom from the bubbles steam.
So that we designed this study under the hypothesis
that metal guarder could both neutralize the
spread of electromotive energy and absorb the spread
of thermal energy. Then, it could better reduce the
spread of uncontrolled tissue desiccation which this
experiment showed a better effect of metal guarder
compared to the control part.
Finally, we recommend conducting the experiment
on alive animals to gain data on the effect of this
technique in living tissue.
5. CONCLUSION
Metal guarder could prevent uncontrolled tissue
desiccation from bipolar electro-coagulator. We believe
that this procedure could modified to the thermal
safe surgical technique and lead the inventor to
modify the newly safety bipolar electro-coagulator.
6. ACKNOWLEDGMENT
The authors wish to thanks the Head of the Department
of Obstetrics and Gynecology, Faculty of
Medicine, Khon Kaen University for giving permission
to conduct this study and Ms.Dararat Khampusaen
for checking the English manuscript.
References
[1] H.S. Cronje, E.C. de Coning, “Electrosurgical
Bipolar Vessel Sealing during Vaginal Hysterectomy,”
Inter J of Gynecol Obstet, vol. 91, pp.
243-245, Aug. 2005.
[2] I.E. Petrakis, K.G. Lasithiotakis, G.E. Chalkiadakis,
“Use of the LigaSure Vessel Sealer in
Total Abdominal Hysterectomy,” Int J Gynaecol
Obstet, vol 89, no 3, pp 303-4 Jun. 2005.
[3] B. Levy, L. Emery, “Ramdomized Trail of Suture
Versus Electrosurgical Bipolar Vessel Sealing in
Vaginal Hysterectomy,” Obstet Gynecol, vol. 102,
no. 1, pp. 147-151, July. 2003.
[4] B.T. Heniford, B.D. Matthews, R.F. Sing, C.
Backus, B. Pratt, F.L. Greene. “Initial Results
with an Electrothermal Bipolar Vessel Sealer,”
Surg Endosc, vol 15, no 8, pp 799-801, Aug. 2001.
[5] A.S. Gozen, D. Teber, J.J. Rassweiler, “Principles
and Initial Experience of a New Device
for Dissection and Hemostasis,” Minim Invasive
Ther, vol 16, no 1, pp 58-65, 2007.
[6] A.I. Brill, “Bipolar Electrosurgery: Convention
and Innovation,” Clin Ostet Gynecol, vol 51, no
1, pp 153-158, 2008.
[7] A. Luciano, R. Soderstrom, D. Martin, “Essential
Principles of Electrosurgery in Operative Laparoscopy,”
J Am Assoc Gynecol Laparosc, vol
1, pp. 189-195, 1994.
[8] A.I. Brill, “Energy Systems for Operative Laparoscopy,”
J Am Assoc Gynecol Laparosc, vol
5, pp. 333-349, 1998.
[9] C. Song, B. Tang, P.A. Campbell, A. Cuschieri,
“ Thermal Spread and Heat Absorbance Differences
Between Open and Laparoscopic Surgeries
During Energized Dissections by Electrosurgical
Instruments,” Surg Endosc, Mar. 2009.
[10] J. Fleshman, “Advanced Technology in the Management
of Hemorrhoids: Stapling, Laser, Harmonic
Scalpel, and lLgasure,” J Gastrointest
Surg, vol 6, no 3, pp 299-301 May-Jun. 2002.
[11] R.M. Soderstrom, “Electrosurgical injuries durling
laparoscopy: prevention and management,”
Curr Opin Obstet Gynecol, vol 6, pp. 248-250,
1994.
[12] M.P. Wu, C.S. Ou, S.L. Chen, E. Yen, R.
Rowbotham, “Complications and Recommended
Practices for Electrosurgery in Laparoscopy,”
Am J Surg, vol.179, pp.67-73, 2000.
[13] C.C. Nduka, P.A. Super, J.R.T. Monson, “Cause
and Prevention of Electrosurgical Injuries in Laparoscopy,”
J Am Coll Surg, vol.179, pp.161-170,
1994.
[14] M.A. Hefni, J. Bhaumik, T. El-Toukhy, P. Kho,
I. Wong, T. Abdel-Raz,ik, A.E. Davies, “Safety
and Efficacy of Using the LigaSure Vessel Sealing
System for Securing the Pedicles in Vaginal
Hysterectomy: Randomized Controlled Trial,”
BJOG, vol. 112, pp. 329-333, Mar. 2005
[15] S. Dessole, G. Rubattu, G. Capobianco, S.
Caredda, P. C. Cherchi, “Utility of Bipolar
Electrocautery Scissors for Abdominal Hysterectomy,”
Am J Obstet Gynecol & Gynecology, vol.
102, no. 1, pp. 147-151, July. 2003.
[16] B. Sigel, M.R. Dunn, “The Mechanism of Blood
Vessel Closure by High Frequency Electrocoagulation,”
Surg Gynecol Obstet, vol 121, pp. 823-
831, 1965.
[17] J.H. Phipps, “Thermometry Studies with Bipolar
Diathermy During Hysterectomy,” Gynaecol
Laparosc, vol 3, pp. 5-7,1994.
[18] V. Remorgida, “Tissue Thermal Damage Caused
by Bipolar Forceps can be Reduced with a Combination
of Plastic and Metal,” Surg Endosc, vol
12, pp. 936-939, 1998.
18 K. Khampitak et al: Metal Guarder Could Prevent the Spread of Tissue Desiccation ... (14-18)
Kovit Khampitak M.D. Department
of Obstetrics and Gynecology Faculty of
Medicine , Khon Kaen University, Thailand.
Sirivit Taechajedcadarungsri Ph.D.
Department of Mechanical Engineering,
Khon Kaen Univarsity, Khon Kaen,
4002, Thailand
Tueanjit Khampitak M.D. Department
of Biochemistry Faculty of Medicine
, Khon Kaen University, Thailand.
Kanok Seejorn M.D. Department of
Obstetrics and Gynecology Faculty of
Medicine , Khon Kaen University, Thailand.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 19
Energy Packing Efficiency Based Threshold
Level Selection for DTW ECG Compression
Y. Chompusri ∗ and S. Yimman, Guest members
ABSTRACT
In this research, the compression method of Electrocardiogram
(ECG) signal is proposed. This
method is based on Wavelet Transform with EPE
threshold level selection. Also, the advantage of Dynamic
Time Warping (DTW) is applied in this algorithm
for efficiently compression of ECG signal, particularly
in irregular period signal. There are many
researches working on compression algorithm of ECG
signal. Almost of them have good result in compression
of regular period signal whereas the result is not
satisfied when those techniques are applied to irregular
period signal. DTW technique is applied for this
purpose which replaces of Period Normalization process.
DTW maps ECG signal to the most correlate
point of reference beat. Therefore all beats seem to
have equal length while all signal beats still remain
the same. From DTW, the residual difference value
is decreased that helps to reduce the error which occurring
from Period Normalization process. Discrete
Wavelet Transform (DWT) is applied to decompose
residual signal into wavelet coefficients after that the
redundant information is eliminated by thresholding.
To higher the ratio of compression while remain the
significant clinical information, the threshold level is
chosen based on Energy Packing Efficiency (EPE).
This threshold technique helps to remove the insignificant
information depending on the conserved energy
in each frequency band. The performance is evaluated
on both regular and irregular period records
of MIT-BIH arrhythmia database. This method improves
compression error and moreover, the peak information
which is essential information of ECG is
always reconstructed at the exact position. Also,
the setting of threshold value based on EPE can efficiently
increase the compression ratio higher than the
conventional method.
* Corresponding author.
Manuscript received on September 29, 2009 ; revised on December
28, 2009.
Yotaka Chompusri is with the Instrumentation and Electronics
Engineering Department, King Mongkut’s University
of Technology North Bangkok, 10800, Thailand.
Surapun Yimman is with the Industrial Physic & Medical
Instrumentation Department, King Mongkut’s University of
Technology North Bangkok, 10800, Thailand.
E-mail addresses: ycps@kmutnb.ac.th (Chompusri Y.) and
sym@kmutnb.ac.th (Yimman S.)
Keywords: ECG, Compression, Dynamic Time
Warping, Discrete Wavelet Transform, Energy Packing
Efficiency
1. INTRODUCTION
ECG is one of very importance clinical information
for cardiologist. The disorder of electrical conduction
in heart can be diagnosed as the risk of cardiac arrhythmias
[1]. For diagnosis, the ECG signal will
be recorded, transmitted, monitored and processed.
The memory consuming for ECG processing is high.
Therefore, many methods were invented to reduce
the size of memory consuming for efficiently ECG
processing. Compression methods are done in many
techniques which classified into three main categories,
Parameter Extraction such as vector quantization,
Direct Time Domain technique such as AZTEC and
Transform Domain technique such as wavelet transform
[2-5]. The objective of compression is increasing
Compression Ratio (CR) meanwhile essential information
is not lost in reconstructed signal. Percent
Root Means Square Difference (PRD) is one parameter
to evaluate how well of quality in reconstructed
signal. There is a tradeoff between CR and PRD
that increasing of CR will reduce the quality of reconstructed
signal which is the increasing of PRD. One
of algorithm resulting in good performance both in
CR and PRD is average beat subtraction and residual
differencing method [6]. This method did well on regular
period signal but not for irregular period signal.
That is affected from Period Normalization process
which makes all beats to have equal length. When
the signal is irregular in period, the large difference
between the reference beat and the signal beat occurs.
That is because signal beat is subtracted by the mismatching
value. On this research, the Dynamic Time
Warping (DTW) technique is applied instead of Period
Normalization process in preprocessing method.
This technique matches the signal beat to the most
correlate point of reference beat, so that all signal
beats are subtracted by the closest value. This makes
the residual value to be smaller. Therefore, the DTW
technique helps to reduce the residual value between
signal and reference beat [7]. In consequence, the
error reduces in reconstructed signal and compression
ratio increases. The Discrete wavelet Transform
(DWT) and threshold technique which based on EPE
are applied to increase efficiency in compression ratio.
In conventional threshold technique which based
20 Y. Chompusri and S. Yimman: Energy Packing Efficiency Based Threshold Level Selection for DTW ... (19-28)
on the maximum of wavelet coefficient’s magnitude
is not suitable. From our previous work, when the
threshold value is set to twice, the increasing of compression
ratio is not higher than 0.05 %. In this proposed
method, the threshold value is set based on the
level of reserved energy of wavelet coefficient in each
level of decomposition. This thresholding technique
increases the compression ratio up to 25 %.
2. THE ALGORITHM
Fig. 1 shows process of the proposed compression
method. First, R peak is detected in QRS complex
detection and then reference beat is determined from
searching every beat to find the most repeated period
and each value of reference beat. Afterward signal is
aligned to reference beat by Dynamic Time Warping
technique. The residual difference is computed from
subtracting signal and match point of reference beat.
After that the Discrete Wavelet Transform is used
to decompose signal into wavelet coefficients and the
thresholding process sets zero value to the insignificant
coefficients. Finally, signal is compressed by
encoding.
Fig.2: Block diagram of proposed reconstruction algorithm.
provides 99.7% accuracy of QRS Complex. The period
of ECG signal is defined between R-R interval
which is illustrated in Fig. 3.
Fig.3:
ECG Signal.
Block diagram of proposed compression algo-
Fig.1:
rithm.
In Fig. 2 is the reconstruction process of this algorithm.
The compressed residual and warp path
are decoded and then the Inverse Discrete Wavelet
Transform combines all coefficients into reconstructed
residual signal. In the mean time, the warp path is
calculated back to match the reconstructed residual
signal with the reference beat. Then the matching
point of reference beat is added back to the reconstructed
residual signal.
2.1 Preprocessing
The morphology of ECG is shown in Fig. 3.
Firstly, each beat is split from the whole signal. QRS
complex of each beat is detected using algorithm of
Hamilton and Tompkins [6] which process steps are
band pass filtering, differentiating, squaring, moving
averaging and the peak decision rule. This algorithm
After every beat is found, the reference beat is determined.
The period of reference beat is chosen from
the most frequent period to gain the smallest residual
difference. The most repeated period is searched
from all signal beats and the most frequent period
occurring is defined as the reference period. After
that, each point of reference beat is computed from
averaging each point of every beat which its period
equals to the reference period. In residual difference
method, the residual is calculated from the difference
between the reference beat and signal. Generally, signal
period is normalized to be equal to the period of
reference beat such as extending the original signal
by zero or mean value [8]. For the regular period,
the signal is perfectly reconstructed but not for irregular
period signal. The high residual difference occurs
when normalized period signal is subtracted by
the deviate point of reference beat. The Period normalization
process increases the residual value that
makes more errors in reconstruction. In this paper,
the Dynamic Time Warping technique is presented to
decrease the magnitude of residual value. When the
reference beat is determined from the most frequently
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 21
detected period of ECG signal, then DTW is used to
warp every point of each beat to the reference beat.
2.2 Dynamic Time Warping Algorithm
The Dynamic Time Warping (DTW) technique is
applied to reduce the residual value from residual differencing
process. This technique replaces of the Period
Normalization process. Dynamic Time Warping
is the non linear matching technique and usually used
in speech and pattern recognition. This technique is
the robust measurement and can compare the similarity
between two signals which are not equal in time
series ignoring both the local shift and the global shift
of signal [9]. DTW will match the most correlate
points between compressed or stretched beat to the
reference beat as illustrates in Fig. 4, the upper is the
shorter period beat, the middle is the reference beat
and the lower is the longer period beat, respectively.
This technique maps all signal beats to the most similarity
point of reference beat with no concerning to
the length of beat.
Fig.5:
matrix.
Calculation the optimal warp path from cost
A warp path, W, indicates the matching pair between
signal X and signal Y.
W = w 1 , w 2 , ..., w k , ..., w K ; max(I, J) ≤ K < I + J
(3)
Where K is the length of warp path and the element
of warp path, w k = (i, j), is the matching index.
There are some constraints of warp path, W, that
must begin and stop in diagonally opposite corner of
cost matrix and be monotonically increasing as (4)
and (5). Therefore every index of each time series
must be matched.
w 1 = (1, 1), w K = (I, J) (4)
Fig.4: Matching of similarity between processed beat
to reference beat.
The cost matrix and the optimal warp path are
illustrated in Fig. 5. The warp path is searched after
the adjacent lowest value cell of cost matrix. This
path matches similar points between two beats which
their period lengths are not equal.
The given two time series X and Y with lengths I
and J, respectively.
X = x 1 , x 2 , ..., x i , ..., x I (1)
Y = y 1 , y 2 , ..., y j , ..., y J (2)
w k = (i, j), w k+1 = (i ′ , j ′ ); i ≤ i ′ ≤ i+1, j ≤ j ′ ≤ j+1
(5)
The optimal warp path, W, is only the minimum
of distance warp path, D(w ki , w kj ) , as shown in (6).
[ K
]
∑
W = min D(w ki , w kj )
k=1
(6)
This path could be calculated using dynamic programming
to fill the cost matrix. The value in each
cell of the cost matrix is the minimum warp distance,
D(i, j) of two time series of point i and j as in (7).
D(i, j) = Dist(i, j)
+ min {D(i − 1, j − 1), D(i − 1, j), D(i, j − 1)}
(7)
Where Dist(i, j) is usually the Euclidean distance
between x i and y j .
22 Y. Chompusri and S. Yimman: Energy Packing Efficiency Based Threshold Level Selection for DTW ... (19-28)
On this proposed method, the modified warp path
is computed to any warp path which length, K, is
longer than the length of processed beat, J. In this
case, that warp path is shortened to the length of
the processed beat. If single point of the processed
beat maps to more than one point of the reference
beat, warp pair which gives the minimum Euclidean
distance between two signals is chosen as in (8).
EP E Ci = ĒCi
E Ci
× 100% (10)
Where ĒCi is the total energy of coefficient in level
i after thresholding and ĒCi is the total energy of
coefficient in level i before thresholding.
Table 1 shows the contributing energy in each subband
using three level of decomposition.
w j = min {..., w(i − 1, j), w(i, j), w(i + 1, j), ...} (8)
Then the length of total warp path is equal to the
length of overall signal.
Each element of warp path contains the index of
matching pair between all signal points and reference
beat points.
2.3 Residual Difference
The residual will be calculated from each warp pair
which contains the matching index between processed
beat and reference beat. The warp pair indicates
which point of reference beat will be subtracted from
processed beat to receive the residual difference. The
residual difference is acquired by (9).
Table 1:
Energy contribute in each sub-band level
Residual j = y j − x i ,w k = (i, j) (9)
In this case, the residual value will be small because
each signal point, y j , is subtracted by the closest
value of reference point, x i .
2.4 Discrete Wavelet Transform and Threshold
Residual signal composes of both significant and
redundant information. Discrete Wavelet Transform
(DWT) is used to decompose the residual signal into
sub-band wavelet coefficients. In this research, the
daubechies wavelet is chosen because the daubechies
wavelet shape is compact enough to receive the information
precisely in time [10]. The result of transformation
composes of approximation coefficient and
detail coefficient in each level of decomposition. The
value of coefficient in detail coefficient is relatively
low comparing to the approximation coefficient.
Threshold is applied to mitigate the insignificant
data in each level of decomposition. The coefficient
which its magnitude is lower than significant level is
set to zero value. This mitigation process causes nonsignificant
distortion in reconstructed signal because
of energy invariance property of orthonormal wavelet
transform.
The significant of coefficient is determined by
threshold value which based on Energy Packing Efficiency
(EPE) [11, 12]. The EPE is a percentage
quantity that presents a measure of the total preserved
energy of a certain sub-band after thresholding
with respect to the total energy in that sub-band
before thresholding and is defined as (10)
Most of energy contributes in lower sub-band especially
in approximation coefficient. Energy presenting
in approximation coefficient level, E CA , is varying
from 52% to 96% of total energy which depends
on point number and shape of signal. To prevent the
significant distortion, the approximation coefficient is
preserved by setting high value of EPE. The energy in
detail coefficient level, E CD , is less in higher sub-band
level and the difference of energy between adjacent
sub-band levels is relative small which is not greater
than 5.9% in regular case and 2.7% in irregular case
of the total energy in that sub-band. Therefore, the
same EPE value of each level of detail coefficient is
chosen. The magnitude of detail coefficient is lower
than the magnitude of approximation coefficient and
the number of detail coefficient is larger than that of
approximation coefficient. To achieve of elimination
of redundant information, the EPE of detail coefficient
can be selected in small value. The following
algorithm is defined to calculate the threshold value
[12].
1. Calculate the total Energy, E Ci in each level i of
the wavelet coefficient, X.
E Ci = ∑ X 2 (11)
2. Desire the retained energy in wavelet coefficient
after thresholding, Ē Ci such as ĒCi = 0.99E Ci .
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 23
3. Reorder the wavelet coefficient in descending order
of the magnitude, X s .
4. Use the following pseudo code to compute the
threshold value.
Set energy = 0
Set k = 0
While energy
24 Y. Chompusri and S. Yimman: Energy Packing Efficiency Based Threshold Level Selection for DTW ... (19-28)
Table 2: CR and PRD of proposed method with
100,103 and 105 (regular)
Table 4: CR and PRD of proposed method with
117,119 and 203 (irregular)
Table 3:
CR and PRD of previous method
Therefore, this research focuses on the compression
of irregular period signal such as signal record 117,
119 and 203.
From evaluation of energy in each level of subband,
the energy of approximation coefficient is more
than 82%. Especially in high varying signal, the energy
is about 95%. Therefore, degree of energy contributing
in approximation coefficient indicates the
regularity of signal.
From Table 4, the testing shows the results that
PRD of this proposed algorithm is low because of the
DTW algorithm improving the alignment between
each beat and reference beat to have equal length.
Consequently, error from Period Normalization process
will not occur. For compression ratio, the performance
is better than previous method but still low.
The increasing of CR1 value is small even though the
retained energy is decreased to half. This method
still does not achieve for CR1 because total compressed
data is added up by warp path which data
size is larger than data size of compressed residual
signal. Moreover, threshold setting has no effect on
warp path’s size. Last column of Table 4 shows CR2
value when the warp path is not included in calculation.
The compression ratio excluding warp path,
CR2, is higher than two times of the compression ratio
including warp path, CR1, especially in case of
small EP E DC setting.
The threshold’s effect on irregular period signal is
different from those on regular period signal. When
EP E AC value reduces, it makes small change in
higher compression ratio but large change in higher
PRD, which is more than three times. If EP E DC
value decreases, compression ratio has much increasing
while PRD has not much increasing.
Fig. 9, Fig. 10 and Fig. 11 are the test results of
signal 117, 119 and 203 respectively. The maximum
error is marked at sampling number 400.
There is another achievement of proposed method
that all peaks are always reconstructed at the same
instant as in original signal. This is the benefit from
Dynamic Time Warping technique which maps the
peak of original signal to the peak of reference beat.
Therefore, the peak will be always added back to
where it’s original position in reconstruction process.
From Fig. 9 shows that there is only small error
around peak point when applied this algorithm to
signal record 117. Even though, record 119 and 203
which are highly varying period as shown in Fig. 10
and Fig. 11, all peaks remain in reconstructed signal
and at the exact position.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 25
Fig.6: Record 100 at EP E AC =99% and EP E DC =50% threshold setting a.) Original signal b.) Reconstructed
signal c.) Error between (a) and (b).
Fig.7: Record 103 at EP E AC =99% and EP E DC =50% threshold setting a.) Original signal b.) Reconstructed
signal c.) Error between (a) and (b).
4. CONCLUSION
The aim of this proposed method is to improve the
algorithm of ECG compression especially on irregular
period ECG. Dynamic Time Warping technique
is applied in preprocessing process instead of Period
Normalization. This technique can reduce the residual
value. In consequence, the error between original
signal and the reconstructed signal is enormously
decreased. But there is some disadvantage of this
technique which is the addition warp path producing
from Dynamic Time Warping process. Since the size
of warp path more than two times the size of compressed
residual, that extremely reduces the compression
ratio. To improve compression ratio, the threshold
selection based on EPE is applied and the threshold
value depends on contributed energy in each subband.
The more varying period is the more energy
in low sub-band, particularly in approximation coefficient.
The high EPE value of approximation coefficient
will be selected to keep the essential information
and the low EPE value of detail coefficient will be
chosen to increase the compression ratio while only
small change of error occurs. Another advantage of
this proposed method is the improving of the reconstructed
signal especially around peak point which
is the importance clinical information. This correctness
arises from DTW process which pairs all peaks
to peak of reference beat. Thereby, all peaks will al-
26 Y. Chompusri and S. Yimman: Energy Packing Efficiency Based Threshold Level Selection for DTW ... (19-28)
Fig.8: Record 105 at EP E AC =99% and EP E DC =50% threshold setting a.) Original signal b.) Reconstructed
signal c.) Error between (a) and (b).
Fig.9: Record 117 at EP E AC =99% and EP E DC =50% threshold setting a.) Original signal b.) Reconstructed
signal c.) Error between (a) and (b).
ways be reconstructed from reconstruction process.
No matter to the regularity of signal period, the result
from proposed method shows less distortion than
other methods.
For further work, this research will keep on increasing
the compression ratio by developing of algorithm
which reducing the size of warp path to lower the
total size of compressed signal and by developing algorithm
for choosing the best threshold value based
on Energy Packing Efficiency to receive the optimal
value between compression ratio and error.
References
[1] H. J. L. M. Vullings, M. H. G. Verhaegen, and
H. B. Verbruggen, “Automated ECG segmentation
with dynamic time warping,” in Engineering
in Medicine and Biology Society Proceedings of
the 20th Annual International Conference of the
IEEE, 1998, pp. 163-166
[2] M. L. Hilton, “Wavelet and wavelet packet compression
of electrocardiograms,” Biomedical Engineering,
IEEE Transactions on, vol. 44, pp.
394-402, 1997.
[3] S. M. S. Jalaleddine, C. G. Hutchens, R.
D. Strattan, and W. A. Coberly, “ECG data
compression techniques-a unified approach,”
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 27
Fig.10: Record 119 at EP E AC =99% and EP E DC =50% threshold setting a.) Original signal b.) Reconstructed
signal c.) Error between (a) and (b).
Fig.11: Record 203 at EP E AC =99% and EP E DC =50% threshold setting a.) Original signal b.) Reconstructed
signal c.) Error between (a) and (b).
Biomedical Engineering, IEEE Transactions on,
vol. 37, pp. 329-343, 1990.
[4] L. Zhitao, K. Dong Youn, and W. A. Pearlman,
“Wavelet compression of ECG signals by the
set partitioning in hierarchical trees algorithm,”
Biomedical Engineering, IEEE Transactions on,
vol. 47, pp. 849-856, 2000.
[5] M. Shaou-Gang, Y. Heng-Lin, and L. Chih-
Lung, “Wavelet-based ECG compression using
dynamic vector quantization with tree codevectors
in single codebook,” Biomedical Engineering,
IEEE Transactions on, vol. 49, pp. 671-680,
2002.
[6] P. S. Hamilton and W. J. Tompkins, “Compression
of the ambulatory ECG by average beat subtraction
and residual differencing,” Biomedical
Engineering, IEEE Transactions on, vol. 38, pp.
253-259, 1991.
[7] B. Huang and W. Kinsner, “ECG frame classification
using dynamic time warping,” in Electrical
and Computer Engineering, IEEE CCECE
2002, Canadian Conference on, 2002, pp. 1105-
1110 vol.2.
[8] C. Hsiao-Hsuan, C. Ying-Jui, S. Yu-Chien, and
K. Te-son, “An effective and efficient compression
algorithm for ECG signals with irregular
periods,” Biomedical Engineering, IEEE Transactions
on, vol. 53, pp. 1198-1205, 2006.
28 Y. Chompusri and S. Yimman: Energy Packing Efficiency Based Threshold Level Selection for DTW ... (19-28)
[9] D. Xiao-Li, G. Cheng-Kui, and W. Zheng-Ou,
“A Local Segmented Dynamic Time Warping
Distance Measure Algorithm for Time Series
Data Mining,” in Machine Learning and Cybernetics,
2006 International Conference on, 2006,
pp. 1247-1252.
[10] D. S. S. Lee, B. J. Lithgow, and R. E. Morrison,
“New fault diagnosis of circuit breakers,”
Power Delivery, IEEE Transactions on, vol. 18,
pp. 454-459, 2003.
[11] M. Abo-Zahhad and B. A. Rajoub, “An effective
coding technique for the compression of onedimensional
signals using wavelet transforms,”
Medical Engineering & Physics, vol. 24, pp. 185-
199, 2002.
[12] B. A. Rajoub, “An efficient coding algorithm
for the compression of ECG signals using
the wavelet transform,” Biomedical Engineering,
IEEE Transactions on, vol. 49, pp. 355-362,
2002.
[13] C. Hsiao-Hsuan, C. Ying-Jui, S. Yu-Chien, and
K. Te-Son, “A high performance compression
algorithm for ECG with irregular periods,” in
Biomedical Circuits and Systems, 2004 IEEE International
Workshop on, 2004, pp. S2/4-9-12.
Yotaka Chompusri was born in Thailand,
1976. She received B.Eng. degree
in Control Engineering from King
Mongkut’s Institute of Technology Ladkrabang,
Thailand in 1998 and received
M.Sc. degree in Electrical Engineering
from University of Southern
California, USA in 2002. She is
currently staff at Instrumentation and
Electronics Engineering Department of
King Mongkut’s University of Technology
North Bangkok, Thailand. Her research works in area of
control system, digital signal processing, data compression and
pattern recognition.
Surapun Yimman was born in Thailand,
1967. He received B.Ind.Tech. and
M.Eng. degree from King Mongkut’s
Institute of Technology North Bangkok,
Thailand in 1991 and in 1998, respectively
and received D.Eng. in Electrical
Engineering from King Mongkut’s Institute
of Technology Ladkrabang, Thailand
in 2007. He is currently staff at Industrial
Physics & Medical Instrumentation
Department of King Mongkut’s
University of Technology North Bangkok, Thailand. His research
interests in area of digital signal processing and application.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 29
No Visual Mismatch Negativity (MMN) for
Simultaneously Presented Audiovisual
Stimuli: Evidence from Human Brain
Processing
W. Sittiprapaporn, Member
ABSTRACT
The present study employed simultaneous audiovisual
stimuli in the oddball paradigm to re-examine
the effects of attention on audio, visual and audiovisual
perception. The study was designed to investigate
whether task-related processing of audio
and visual features was independent or task-related
processing in one modality might influence the processing
of the other. Electroencephalogram (EEG)
was recorded from 12 normal subjects. ANOVA
showed statistically significant of the interaction between
electrode site and modality. The difference
waves with 100-200 ms latency at the anterior sites
were markedly different to the posterior sites. The
emergence of posterior negativity in the audio-visual
modality might not be attributed to visual discrimination
process as it did not appear in the visual
modality. The findings reveal the processing of a feature,
hierarchically dependent on another feature in
the audiovisual perception.
Keywords: Event-related potentials, oddball
paradigm, Mismatch negativity, bisensory processing
1. INTRODUCTION
The human central auditory system has a remarkable
ability to establish memory traces for invariant
features in the acoustic environments such as music
and speech sounds in order to correct the interpretation
of natural acoustic sound heard. Even when no
conscious attention is paid to the surrounding sounds,
changes in their regularity can cause the listener to
redirect his or her attention toward the sound heard
[1]. Event-related potential (ERP) recordings have
bought new insight to the neuronal events behind
auditory change detection in audition. ERPs components
(see Fig. 1) reflect the conscious detection
of a physical, semantic, or syntactic deviation from
Manuscript received September 30, 2009; revised on December
28, 2009.
W. Sittiprapaporn is with College of Music, Mahasarakham
University, Khamriang, Kantarawichai, Mahasarakham 44150
Thailand.
E-mail addresses: drwichain@hotmail.com (W. Sittiprapaporn)
the expected sounds [1]. The ERP recordings thus
allow one to probe the neural processes preceding the
involvement of the attentional mechanisms. For instances,
ERPs have been recorded that reflect memory
traces representing sounds composed of several
simultaneous or successive tonal elements [2-4].
Fig.1: Event-Related Potential (ERP) Components.
Mismatch negativity (MMN) component of ERP is
theoretically elicited in the auditory cortex when incoming
sounds are detected as deviating from a neural
representation of acoustic regularities (see Fig. 2).
It is mainly generated in the auditory cortex [5] occurring
between 100 to 250 ms and thus long been
regarded as specific to the auditory modality [6-7].
The automatic change-detection system in the human
brain as reflected by the MMN requires the storage
of the previous state of the acoustic environment for
detecting an incoming deviating sound [6,8]. MMN
implies the existence of an auditory sensory memory
that stores a neural representation of a standard
against which any incoming auditory input is compared
[9]. In the auditory modality, MMN is an automatic
process which occurs even when the subject’s
attention is focused away from the evoking stimuli [6].
Its onset normally begins before the N2b-P3 complex
which occurs when attention is directed to the stimuli.
The duration of MMN varies with the nature of
30 W. Sittiprapaporn: No Visual Mismatch Negativity (MMN) for Simultaneously Presented Audiovisual Stimuli ... (29-35)
the stimulus deviance but it invariably overlaps N2b
when the latter is present [10].
Fig.2:
ERP.
Mismatch Negativity (MMN) Component of
Previous study [11] has stated that the automatic
detection of stimulus change plays a part in directing
attention to events of biological importance. If
this is the case, one would expect a similar mechanism
to operate in the visual modality. Although
it is clear that the MMN can be elicited in auditory
modality in the absence of attention, it remains somewhat
unclear whether there is an analogous automatic
deviant-related negativity (DRN) elicited outside the
auditory modality. Näätänen [12] has stated that
the automatic detection of stimulus change plays a
part in directing attention to events of biological importance.
If this is the case, one would expect a
similar mechanism to operate in the visual modality.
Even though MMN had not mentioned to be
appeared in the visual modality [12], several studies
have shown that visual stimuli deviating from repetitive
visual standards can also elicit a visual analogue
of the MMN in the same latency range. This visual
MMN seems to be mainly generated in occipital areas
[12,13] with possibly a more anterior positive component
[14,15]. In addition, Cammann’s study [16]
showed a widely distributed MMN change between
150 and 350 ms, with a parietal maximum suggesting
that this MMN may occur in the visual modality.
Recently, Pazo-Alvarez et al. [17] reviewed several
previous reports to provide convincing evidence for
the existence of this visual MMN. Moreover, crossmodal
attention studies clearly showed that deviant
visual stimuli elicited MMN, largest over the inferior
temporal cortex. This visual MMN increased in amplitude
with attention, but it was also evident during
inattention [18,19].
In the present study, simultaneous audio-visual
stimulus in the oddball paradigm was used to reexamine
the effects of attention on MMN in auditory,
visual and audiovisual dimensions. Attentional
ERP components were analyzed in a situation where
target stimuli were combinations of both auditory
and visual features. Interactive processing of stimulus
features would then be indicated by the absence,
reduction or early termination of the attentionrelated
components [20] as a function of processing of
the other feature. If visual-specific components are
evoked by visual deviances, then the present audiovisual
paradigm will help to separate them from the
effect of visual information on the auditory-specific
MMN process by facilitating the focus of attention on
auditory and visual MMNs elicited with bimodal features.
The audio-visual paradigm was also designed
to investigate whether task-related processing of visual
and auditory features was independent or taskrelated
processing in one modality might influence the
processing of the other.
2. MATERIALS AND METHODS
2.1 Subjects
Twelve right-handed normal subjects (6 males and
6 females) with a mean age of 24.83 (SD= 3.54) participated
in the experiment. All participants had normal
hearing and had been corrected to normal vision
(self reported). None of them had more than three
years of formal musical training and none had any
musical training within the past five years. All participants
had no history of neurological or psychiatric
history. After a complete description of the intended
study, written informed consent was obtained. The
subjects were paid for their participation.
2.2 Stimuli
Stimuli consisted of a set of four audio-visual stimuli
that were distinguished by frequencies (Hz) for audio
and features for visual appearing on the screen.
The duration of the stimuli were 300 ms. The stimulus
system (STIM, Neurosoft, Inc. Sterling, USA)
was employed for controlling the presentation of the
stimuli. An oddball paradigm [1] was chosen for presenting
randomized stimulus sequences consisting of
all four sets of equiprobable audio-visual stimuli (a simultaneous
combination of audio and visual stimuli):
the deviant was ’X’ with 1800 Hz tone (Visual Target
Audio Target; hereafter, VTAT) in 10% probability,
and the standard was ’Y’ with 800 Hz tone (Visual
Non-target Audio Non-target; hereafter, VNAN) in
70% probability were presented as preferred-deviant
to be able to check that participants were attending
the stimuli. Additionally, the deviant ’X’ with
800 Hz tone (Visual Target Audio Non-target; hereafter,
VTAN) and the deviant ’Y’ with 1800 Hz tone
(Visual Non-target Audio Target; hereafter, VNAT)
were used in 10% probabilities (see Fig. 3). While visual
stimuli were presented on the computer screen,
acoustic/audio stimuli were delivered binaurally to
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 31
the participants through plastic tubes and earpieces.
Sound density was adjusted to be 85 dB above the
participant’s hearing threshold.
30 Hz. The averaging epoch was 900 ms, and the 100
ms before the onset of the presenting stimuli served
as baseline. The artifact-free epochs were filtered at
0.1-15 Hz, baseline corrected and averaged. The artifacts
rejection was conducted in all channels with
threshold of ± 100 µV before averaging. Epochs with
EEG or EOG with a large (>100 µV) amplitude were
also automatically rejected. The limitation restricted
EEG recording time to 90 mins, thus minimizing the
risk of participants’ fatigue (see Fig. 4).
Fig.3: Schematic presentation of the stimuli in
bisensory conditions. Stimuli in different modalities
are presented simultaneously. The VTAT and VNAN
conditions appeared in 10% and 70% probabilities, respectively.
Additionally, the VTAN and VNAT conditions
were equally presented in 10% probabilities.
The stimulus is presented with 300 ms duration, and
inter-stimulus interval is 1,800 ms in every condition.
2.3 Procedures
The experiment consisted of 3 blocks and each
block had 300 trials. Every stimulus was presented
with 300 ms exposure duration and inter-stimulus interval
was 1,800 ms (from audio/visual stimuli onset
to onset) in every condition. Subjects sat in an electrically
shield and soundproofed room with the response
buttons under their hands. The subjects had
to press the button on the response pad when the deviant
(VTAT) was presented and ignored any other
types of stimuli. Prior to the experimental session, a
practice block was administrated to ensure that the
subjects understood the task.
2.4 Electroencephalogram recording
Electroencephalographic (EEG) data were collected
in an electrically and acoustically shielded
room. EEG was recorded from a Quick-Cap equipped
with 128 channels according to the international 10-
20 system using Scan system (Scan 4.2, Neurosoft,
Inc. Sterling, USA). Linked mastoids were used as
reference. Eye movements were monitored with two
EOG electrodes. Four electrodes monitored horizontal
and vertical eye movements for off-line artifact
rejection. Vertical and horizontal electro-oculogram
(EOG) was recorded by electrodes situated above and
below the left eye, and on the outer canthi of both
eyes, respectively. Impedance was maintained at 5kΩ
or less. During the experiment, EEG was amplified
with a bandpass of 0.05 - 100 Hz, sampled at 1,000 Hz
and stored on a hard disk for off-line analysis. ERPs
were averaged separately for each types of stimulus.
They were digitally filtered with a bandpass of 0.1 -
Fig.4:
Electroencephalogram (EEG) Recording.
2.5 EEG data analysis
After the data recordings, the EEG was segmented
into 1000 ms epochs, including the 100 ms prestimulus
period. The baseline was corrected separately
for each channel according to the mean amplitude
of the EEG over the 100 ms period that preceded
stimulus onset. The EEG epochs contained
amplitudes exceeding ±100 µV at any EEG channels
were automatically excluded from the averaging.
The epoch was separately averaged for the standard
and deviant stimulus. The average waveforms obtained
from the standard and deviant stimuli were
digitally filtered by a 0.1 - 15 Hz band-pass filter and
finally baseline-corrected. To analyze the deviantrelated
components, difference potentials were calculated
where responses elicited by the VNAN stimuli
were subtracted from responses to VTAN and VNAT
stimuli after stimulus onset referred to visual (Vi)
modality as in (1) and auditory (Au) modality as in
(2), respectively.
and
(V NAN) − (V T AN) = (V i) (1)
(V NAN) − (V NAT ) = (Au) (2)
In the audio-visual (AV) modality, VTAT minus
VNAN difference was also calculated as in (3).
32 W. Sittiprapaporn: No Visual Mismatch Negativity (MMN) for Simultaneously Presented Audiovisual Stimuli ... (29-35)
(V NAN) − (V T AT ) = (AV ) (3)
The amplitude of the difference waveform was expressed
in microvolt and its latency in milliseconds.
MMNs were statistically assessed by two-tailed t-tests
comparing the averaged amplitude of the deviant minus
standard difference waveform to zero in the 40
ms time-window around the latency of the peak in
the grand-average responses. To compare these components,
MMN amplitudes were further assessed via
two-way analyses of variance (ANOVA) with repeated
measurements. The factors were modality (three levels:
Vi, Au and AV), and electrode site (two levels:
anterior sites at F3, Fz, F4, C3, Cz, C4, and posterior
sites at P3, Pz, P4, O1, Oz, O2).
3. RESULTS
Reaction times and response accuracy (mean and
standard deviation: SD) are shown in Table 1. Figure
5 presents the grand-average deviant-related components
in the Au, Vi and AV modalities producing
deviant-related negativities (DRNs). DRNs were divided
into an early DRN1 around 100-200 ms and
a late DRN2 around 200-300 ms. According to the
previous studies showing that MMN appears between
100 to 250 ms [6] and the characteristics of DRN2
match with those of N2b component [23]. The present
study thus associated DRN1 mainly with MMN in
which we focus in this report, and DRN2 with a mixed
wave of MMN and N2b.
way repeated measures ANOVA shows that the
interaction between electrode site and modality of
MMN amplitudes at 100-200 ms of all modalities was
statistically significant [F(11,429) = 8.27, p ¡ 0.0001].
At 200-300 ms, significant levels were also reached in
the same interaction for N2b component [F(11,429)
= 6.50, p ¡ 0.0001]. As shown in Fig. 2, the difference
waves with 100-200 ms latency at the anterior
sites were markedly different to the posterior sites.
Additionally, there was no MMN elicitation for the
Vi modality at the posterior sites compared to the
Au and AV modalities. We thus compared MMN
mean amplitudes of all modalities. Two-way repeated
measures ANOVA shows that the interaction between
posterior electrode site and modality was statistically
significant [F(17,663) = 27.52, p ¡ 0.0001] and significant
level was also reached in the interactions between
anterior electrode site and modality [F(17,663)
= 52.37, p ¡ 0.0001]. We then compared the MMN
mean amplitude values of Au, Vi and AV difference
potentials at Fz site. The difference was statistically
significant [F(2,78) = 8.75, p ¡ 0.0001]. Like
the MMN, they showed similar significant effect on
the N2b amplitude at Oz site [F(2,78) = 6.50, p ¡
0.0001].
The additivity of the MMN was also examined by
adding together the Au and Vi MMNs and comparing
this ’modelled’ (AuVi) MMN with the AV MMN
in order to see the possible attention effects on the
additivity of MMN. If processing of Au and Vi is independent
of the others, the sum of the MMNs to
both modalities should be equal to the MMN elicited
by the AV modality. We found that the additivity of
Au and Vi MMNs existed in both anterior and posterior
locations. However, the additivity of Au and Vi
MMNs amplitude was slightly larger than that of the
corresponding AV modality, being maximum at P3
(mean amplitude; AuVi vs. AV: -1.56 (0.02) µV vs.
-1.12 (0.02) µV, t(39) = -21.89, p < 0.0001). Moreover,
the N2b component, following MMN, was larger
than that of the AV modality, being maximum at Fz
(mean amplitude: -4.51 (0.04) µV vs. -3.97 (0.05)
µV, t(39) = -45.99, p < 0.0001). The N2b was also
followed by a positive component identified as P3a [1]
(see Fig. 6).
4. DISCUSSION
The main finding of our study indicates that the
prominent response to the Au, Vi and AV modalities
produces deviant-related negativities. DRNs were divided
into early DRN1 (or MMN), and late DRN2
(or N2b). As shown in Fig. 5, the difference waves
with 100-200 ms latency at the anterior sites were
markedly different to the posterior sites. There was
no MMN elicitation for the visual modality at the
posterior electrode sites compared to the auditory
and audiovisual modalities. The MMN was significantly
larger only in the anterior sites, being maximum
at F3 (t(39) = -68.04, p ¡ 0.0001). This result
was consistent with a previous study showing no posterior
negativity elicitation in the difficult discrimination
task [22]. Moreover, the present result extends
previous findings [23,25] which showed that the
deviance related ERP effects in vision could be separated
from automatic processing of other stimulus
features. We hypothesize that the emergence of posterior
negativity (MMN) in the present study is not
to be attributed to visual discrimination process. Our
result supports the view proposed by Näätänen [12]
that “no MMN appears to occur in the visual modality”.
However, several studies have shown that visual
stimuli deviating from repetitive visual standards can
also elicit a visual analogue of the MMN in the same
latency range [12,24,14,15].
Furthermore, cross-modal attention studies showed
that deviant visual stimuli elicited MMN, largest
over the occipital and inferior temporal cortex [19].
This visual MMN was not affected by the processing
load during attention to the other modality and had
restricted, occipito-temporal distribution, consistent
with generation in modality-specific sensory cortex.
This early MMN-like portion of the visual deviancerelated
negativity was independent of attention. It
increased in amplitude with attention, but it was also
evident during inattention [19]. However, Alho et al.
[25] has argued that if a visual MMN exists, its elicita-
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 33
tion may have a higher threshold than auditory MMN
which evoked by any discriminable change. The effect
of target-specific negativity is thus a considerable
contamination factors in which the impact of simultaneous
memory traces in different modalities could
also be considered. In addition, the persistence of the
visual MMN may reflect the automatic detection of
physical change in sequences of visual stimuli [12].
As shown in Fig. 5, the identical N2b components
were elicited by Vi and AV modalities, whereas in case
of Au modality, latency of this component was longer
than that of the Vi and AV modalities. The N2b effect
suggests the attention-related rechecking of the
outcome of within-modality analyses. Such activity
would have been different upon the different discrimination
demand [20,22]. Therefore, the process underlying
N2b component thus performs independent
within-dimension selection [20]. The shorter N2b latency
to AV further suggests that this component is
a correlate of processes following the elementary discrimination
processes, instead of being an on-line correlate
of such processes [22].
The additivity of the MMN was also examined by
adding together the Au and Vi MMNs and comparing
this ’modelled’ (AuVi) MMN with the AV MMN in
order to see the possible attention effects on the additivity
of MMN. Assuming that the MMNs to these
features are generated by different, non-interacting
neural populations, each deviating feature in the bimodal
deviants should elicit its own MMN. The additivity
of Au and Vi deviants should thus elicit a
larger MMN than the AV deviant. As shown in Fig.
6, our findings show that the additivity of the Au
and Vi MMNs was larger than that of the AV modality.
This implies that there are complex interactions
between brains processes involved in analyzing several
simultaneous deviant features in the AV modality.
Our results are in the line of previous studies
revealing that at least partly different neural populations
are involved in processing deviance in different
auditory features and being suggest the independent
MMN generators for these features [26-29]. Moreover,
the underadditivity of AV MMN in the present study
suggests either that common neural populations are
involved in the controlled processing of changes in different
features [27] or that the populations are separate
but strongly interacting [29]. However, Paavilainen
et al [29] has argued that the additivity hypothesis
of MMN does not hold at least in its simplest
form which presupposes that the processing of various
features is completely independent of each other,
the contributions of the different simultaneous MMNs
just simply ’piling up’ in the ERPs. According to this
hypothesis, the AuVi MMN in the present study was
slightly larger than did the corresponding AV modality,
being maximum at P3 site. Like AuVi MMN, the
N2b component was also larger than that of the AV
modality, being maximum at Fz site.
, our findings are consistent with the previous
study showing that the deviant-related negativities
consist of two successive components, the earlier being
generated at the auditory cortex and the latter at
the frontal areas [30]. In addition, the degree of additivity
may be different from those two components.
That is, the existence of several partially overlapping
and interacting brain processes may complicate the
estimation of the additivity of MMN [29]. The negativity
associated with deviants in the present study
thus resembles auditory mismatch negativity inasmuch
as it occurs automatically while the focus of
attention is directed away from the evoking stimuli.
Finally, the morphology of ERPs to AuVi modality
differed to the AV modality. The N2b was followed
by a large positive component identified as P3a [6]
in the AuVi modality, but little P3a was evident in
the AV modality (see Fig. 6). Like the MMN/N2b,
the empirical AuVi P3a tended to be larger than the
AV P3a. Its amplitude increased as the number of
the deviant features was increased, with the additivity
of Au and Vi deviants eliciting a larger P3a
than did the corresponding AV deviants. This implies
that both Au and Vi modalities were demanding,
performance in the inattention remained high,
and the occurrence of P3a component depend on another
feature. The occurrence of AV P3a also implied
the complex interactions between brains processes
involved in analyzing several simultaneous deviant
features. This component possibly reflects involuntary
attention-switching mechanism to deviant
stimuli [23,31,32]. Consequently, our results support
the view that the processing of a feature, hierarchically
dependents on another feature [20,22].
5. CONCLUSION
The present study demonstrates the audiovisual
interaction following elementary within-modality discrimination
processes. MMN and N2b effects suggest
the attention-related rechecking of the outcome
of within-modality analyses. The task-related processing
of audio and visual features was independent
and one modality might influence the processing of
the other. This findings support the view that the
processing of a feature, hierarchically dependent on
another feature in the condition of audio-visual perception.
6. ACKNOWLEDGMENT
This research was conducted in cooperation with
Clinical Cognitive Neuroscience Center (CCNC) in
the Seoul National University, College of Medicine,
Seoul, Korea. The author gratefully thank Prof. Dr.
Jun Soo Kwon, Dr. Do-Hyung Kang, Dr. Kyung
Whun Kang and Dr. Bo Reom Lee for their guidance
and providing the equipment and data sources during
this work. In addition, this research was supported by
34 W. Sittiprapaporn: No Visual Mismatch Negativity (MMN) for Simultaneously Presented Audiovisual Stimuli ... (29-35)
the following organizations; the Brain Research Center
of the 21st Century Frontier Research Program
(Ministry of Science and Technology of Republic of
Korea), and the International Scholar Exchange Fellowship
(ISEF) Program, 2006-2007 (Korea Foundation
for Advanced Studies).
References
[1] M. Tervaniemi, M. Rytkönen, E. Schröger, R. Ilmoniemi
and R. Näätänen, “Superior formation of
cortical memory traces for melodic patterns in musicians.”
Learning & Memory 8, 295-300, 2001
[2] E. Schröger, M. Tervaniemi, C. Wolff and R.
Näätänen, “Preattentive periodicity detection in
auditory patterns as governed by time and intensity
information.” Brain Res. Cogn. Brain Res. 4,
145-148, 1996
[3] C. Alain, D.L. Woods and K.H. Ogawa, “Brain
indices of automatic pattern processing.” NeuroReport,
6, 140-144, 1994
[4] K. Alho, M. Tervaniemi, M. Huotilainen, J.
Lavikainen, H. Tiitinen, R.J. Ilmoniemi, J. Knuutila
and R. Näätänen, “Processing of complex
sounds in the human auditory cortex as revealed by
magnetic brain responses.” Psychophysiology, vol.
33, pp. 396-375, 1996
[5] M. Scherg, J. Vajsar and T.W. Picton, “A source
analysis of the late human auditory evoked potentials.”
J. Cogn. Neurosci., vol. 5, pp. 363-370, 1989
[6] R. Näätänen, “Attention and Brain Function.”
Erlbaum, Hillsdale, NJ.; 1992
[7] G. Nyman, K. Alho, P. Iaurinen, P. Paavilainen,
T. Radil and K. Rainikainen, “Mismatch negativity
(MMN) for sequences of auditory and visual stimuli:
evidence for a mechanism specific to the auditory
modality”, Electroenceph. Clin Neurophys.,
vol. 77, pp. 436-444, 1990
[8] E. Brattico, I. Winkler, R. Näätänen, P. Paavilainen
and M. Tervaniemi, “Simultaneous storage
of two complex temporal sound patterns in auditory
sensory memory.” NeuroReport, vol. 13, pp.
1747-1751, 2002
[9] W. Ritter, D. Deacon, H. Gomes, D.C. Javitt and
H.G. Jr. Vaughan, “The mismatch negativity of
event-related potentials as a probe of transient auditory
memory: a review.” Ear Hear., vol. 16, pp.
52-67, 1995
[10] A. Tales, P. Newton, T. Troscianko and S. Butler,
“Mismatch negativity in the visual modality”.
NeuroReport, vol. 10, pp.3363-3367, 1999
[11] R. Näätänen, “The role attention in auditory information
processing as revealed by event-related
potentials and other brain measures of cognitive
function.” Behav. Brain Sci., vol. 13, pp. 201-288,
1990
′
[12] I. Czigler, L. Bal azs ′ and L.G. PatO, “Visual
change detection: event-related potentials are dependent
on stimulus location in humans”. Neurosci.
Lett., vol. 364(3), pp. 149-153, 2004
[13] G. Yucel, C. Pettv, G. McCarthv and A. Belger,
“Graded visual attention modulates brain response
evoked by task-irrelevant auditory pitch changes”.
J. Cogn. Neurosci., vol. 17(2), pp. 1819-1828, 2005
[14] I. Czigler, L. Balazs and I. Winkler, “Memorybased
detection of task-irrelevant visual changes”.
Psychophysiology, vol. 39, pp. 869-873, 2002
[15] D.J. Heslenfeld, Visual mismatch negativity. In:
J. Polish (ed). Detection of change: event-related
potential and fMRI findings. Dordrecht: Kluwer
Academic Publishers; pp. 41-60, 2003.
[16] R. Cammann, “Is there a mismatch negativity
(MMN) in the visual modality?” Behv. Brain Sci.,
vol. 13, pp. 234-235, 1999
[17] P. Pazo-Alvarez, F. Cadaveira and E. Amenedo,
“MMN in the visual modality: a review”. Biol.
Psychology, vol. 63, pp.1999-236, 2003
[18] D.J. Woods, K. Alho and A. Algazi, “Intermodal
selective attention. I. Effects on event-related potentials
to lateralized auditory and visual stimuli”.
Electroenceph. Clin. Neurophys., vol. 82, pp.341-
355, 1992
[19] K. Alho, D.L. Woods, A. Algazi and R.
Näätänen, “Intermodal selective attention. II. Effects
of attentional load on processing of auditory
and visual stimuli in central space”. Electroenceph.
Clin. Neurophys., vol. 82, pp.356-368, 1992
[20] H.G.O.M. Smid, A. Jakob and H.J. Heinze, “An
event-related brain potential study of visual selective
attention to conjunctions of color and shape”.
Psychophysiology, vol. 36, pp. 264-279, 1999
[21] K. Alho, P. Paavilainen, K. Reinikainen, M.
Sams and R. Näätänen, “Separability of different
negative components of the event-related potential
associated with auditory stimulus processing”.
Psychophysiology, vol. 23, pp. 613-623, 1986
[22] I. Czigler and L. Balazs, “Event-related potentials
and audiovisual stimuli: multimodal interactions”.
NeuroReport, vol. 12, pp. 223-236, 2001
[23] S. Berti and E. Schröger, “A comparison of auditory
and visual distraction effects: behavioral and
event-related indices”. Brain Res. Cog. Brain Res.,
vol. 10, pp. 265-273, 2001
[24] S. Berti and E. Schroger, “Distraction effects in
vision: behavioral and event-related potential indices”.
NeuroReport, vol. 15, pp.665-669, 2004
[25] K. Alho, D.L. Woods and A. Algazi, “Processing
of auditory stimuli during auditory and visual
attention as revealed by event-related potentials”.
Psychophysiology, vol. 31, pp. 469-479, 1994
[26] S. Levänen, R. Hari, L. McEvoy and M. Sams,
“Responses of the human auditory cortex to
changes in one vs. two stimulus features”. Exp.
Brain Res., vol. 97, pp. 177-183, 1993
[27] E. Schröger, “Processing of auditory deviants
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 35
with changes in one versus two stimulus dimensions”.
Psychophysiology, vol. 32, pp. 55-56, 1995
[28] R. Takegata, P. Paavilainen, R. Näätänen and
I. Winkler, “Independent processing of changes in
auditory single features and feature conjunctions
in human as indexed by the mismatch negativity”.
Neurosci. Lett., vol. 26, pp. 109-112, 1999
[29] P. Paavilainen, S. Valppu and R. Näätänen,
“The Additivity of the auditory feature analysis in
the human brain as indexed by the mismatch negativity:
1+1=2 but 1+1+1
36 A. Leelasantitham and S. Kiattisin: A Diagnosis of Tonsillitis using Image Processing and Neural Network (36-42)
A Diagnosis of Tonsillitis using Image
Processing and Neural Network
A. Leelasantitham and S. Kiattisin, Members
ABSTRACT
Tonsillitis is a disease occurring mostly in child and
adults as this disease may take to the other effects.
Nowadays, a detection of tonsil grand exploits medical
doctor’s diagnosis to check on oral cavity. Therefore,
this paper presents a diagnosis of tonsillitis using
image processing and neural network (NN). There are
three steps described as follows. The first step is localization
of tonsil grand (TG) using Ellipses Hough
Transform. The second step is feature extraction using
three important factors which can be indicated in
swelling by images of TG in terms of a) a dimensional
ratio of TG, b) an average intensity of TG and c) a
purulent surface of TG (yes/no) using power spectrum
of two dimensional Fast Fourier Transform (2D
FFT). The final step is verification using NN. The
three factors are inputted into NN, and TG samples
of 50 images are used for training into the NN. They
are divided by tonsillitis patience 25 images and usual
TG of 25 images. This experiment uses 100 images
of TG for testing NN. Therefore, the overall accuracy
is at approximately 90% in terms of comparing with
the results from the medical doctor.
Keywords: tonsillitis; diagnosis; neural network;
image processing; 2D FFT,Ellipses Hough Transform
1. INTRODUCTION
Tonsil grand (TG) is the group of tissue type as
lymphatic gland which is placed inside of buccal cavity
called ‘Palatine Tonsil’. The inflamed disease of
tonsil is a swell of tonsil grand (like a little of tubercules)
because of the mostly causes from infectiousness
such as virus and bacteria. Tonsillitis is not a
serious disease but the disadvantage of the inflamed
TG will show its area in an oral cavity. If it has
occurred both of inflammable and growing TG to a
big grand, then a respiratory pathway of human will
be clogged in their breath or else this infection may
severely become a ulcer in throat. Infectious inci-
Manuscript received on September 30, 2009; revised on December
28, 2009
Adisorn Leelasantitham and Supaporn Kiattisin are from
Computer and Multimedia Engineering, School of Engineering,
University of the Thai Chamber of Commerce 126/1,
Vibhavadee-Rangsit Road, Dindaeng, Bangkok 10400, Thailand
E-mail addresses: adisorn lee@utcc.ac.th (A. Leelasantitham)
supaporn kai@utcc.ac.th (S. Kiattisin)
dence is occurred in the blood and it may affect, for
example, the heated level of middle ear, the sinusitis,
the bronchitis or the pneumonia. Such diseases are
important problems in the tonsillitis of child in the
case of the inflammation or the severe heart occurred
from the infectious streptococcus. If it takes time to
be chronic, then it may cause a valve disease with
disabilities and heart failure [1].
Diagnosis of doctor will check the oral cavity to see
any features of TG which is exposed for questioning
symptoms of history from the patience. For diagnosis
in oral cavity, the doctors will see directly the oral
cavity to observe the features of TG, the disease can
be intercommunication, and the patience has drawn
a breath odor. An alternative method to help the
medical doctor for diagnosis of the tonsil is the use of
images and computer because the medical doctor will
check the oral cavity to see closely the tonsil grand of
the patience. As a result, he sometimes may receive
the disease from the patience. To support the practical
doctor, it has taken pictures in the oral cavity to
see a feature of tonsil grand for instead of checking
directly in the oral cavity. In open literature, only
three related papers have involved with finding out
the tonsil using images [2], [3]. Such two papers have
checked two factors as follows: 1) a size of TG using
the sized ratio of a usual TG (from the previous
diagnosis) and a new TG (from the present diagnosis);
as well as, 2) an average intensity of the area
of TG. These two factors will be sent to fuzzy logic
to make a decision. However the information of such
two papers is not sufficient to analyze for checking
tonsillitis because of only two factors (i.e. the size
and intensity).
In this paper, a diagnosis of tonsillitis using image
processing and neural network (NN). There are three
steps described as follows. The first step is localization
of tonsil grand (TG) using the Ellipses Hough
Transform. The second step is feature extraction using
three important factors which can be indicated in
swelling by images of TG in terms of a) a dimensional
ratio of TG, b) an average intensity of TG and c) a
purulent surface of TG (yes/no) using power spectrum
of two dimensional Fast Fourier Transform (2D
FFT). The final step is verification using NN. The
three factors are inputted into NN, and TG samples
of 50 images are used for training into the NN. They
are divided by tonsillitis patience 25 images and usual
TG of 25 images. This experiment uses 100 images
of TG for testing NN. Therefore, the overall accuracy
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 37
is at approximately 90% in terms of comparing with
the results from the medical doctor.
2. FEATURES OF TONSIL GRANDS
Figure 1 shows tonsil grand (TG) called “Palatine
Tonsil” to be two lymph nodes located in the oral
cavity on the side of neck with nearly tongue baseness
where we can see it when we open a mount.
Figure 2 shows the tonsillitis which the oral cavity
can be found in the grown dimension and the timidly
inflammable red. If it reaches severe inflammation,
then it will be seen in translucent whites covering
TG called as “ulcer” [4], [5].
A system of the detection of tonsillitis im-
Fig.3:
ages.
Fig.1: The usual tonsil grand [5].
Fig.4: Flowchart for detection of tonsillitis images
based on medical knowledge and neural network.
3.1 Localization
Fig.2: Tonsillitis called as “ulcer” [5].
3. PROPOSED METHODOLOGY
Figure 3 shows a system of the detection of tonsillitis
images which the system consists of a tonsil
grand (TG), a camera and a computer. The process
is to start the camera capturing the TG and sending
it to the computer for processing three factors i.e. 1)
the dimensional ratio of TG, 2) the average intensity
of TG and 3) the purulent surface of TG.
Figure 4 shows a flowchart of the detection of tonsillitis
images based on medical knowledge and neural
network. There are three main steps of the proposed
methodology i.e. the localization, the feature extraction
and the verification. The details of the three
main steps can be described as follows.
This paper uses the Ellipse Hough Transform
(EHT) to find the edge of TG because the shape of
TG is nearly related in ellipse shape. Therefore, the
equation of the EHT can be defined as [6]
(u·cos θ+v·sin θ−x) 2
a 2
+
(−u·sin θ+v·cos θ−y)2
b 2 = 1
where u and v are image point,
H(x,y,a,b,θ) is a hough space,
x, y, a, b and θ are parameters for the EHT.
(1)
Figure 5 shows an example for localizing the edge
of TG. The process is started from a TG image converted
to a binary image. As a result of the binary
image of TG, the position of TG is located by the
EHT. As shown in Figure 5, the dotted line of TG
edge is generated by equation (1) until it finds out
(0 or 1) all possible parameters (x,y,a,b,θ) that could
pass (u,v) vote for parameters (x,y,a,b,θ), then the
solid line can cover the final edge of TG.
38 A. Leelasantitham and S. Kiattisin: A Diagnosis of Tonsillitis using Image Processing and Neural Network (36-42)
Fig.6: An example for finding the dimensional ratio
of TG.
3.2.2 The average intensity of TG
Fig.5:
An example for localizing the edge of TG.
3.2 Feature Extraction
3.2.1 The dimensional ratio of TG
The first factor is a result of the TG dimensional
ratio which can be found by measuring a size of TG
(ellipse shape) which is the width (W) and height (H).
The position of TG shape is obtained from the EHT.
Therefore, the equation is equal to
R = W H
(2)
which R is the dimensional ratio of width and
height,
W is the width of TG (pixels),
H is the height of TG (pixels).
Figure 6 shows an example for finding the dimensional
ratio of TG. The values (pixels) of W and H
are obtained from the binary image which the shape
of TG is related to the ellipse shape shown in Figure
6. Therefore, the result of the dimensional ratio of
TG is calculated by equation (2). Note that the W
and H are the minor and major axes, respectively, of
ellipse shape.
The second factor is a result of the TG average
intensity resulted from three colors (red, green and
blue [R, G, B]) which are not the same values. R
values are normally higher than G and B values in the
TG image. However, there are some images which G
or B values are nearly to R values. All R, G and B
are important to use this method for verifying the TG
image. For another reason, the TG generally is a high
redness then it may be the tonsillitis, nevertheless it
sometimes may not be the tonsillitis. Moreover, there
are two factors i.e. the first and third factors being
necessary to consider in this case. Such two factors
will also be utilized for classification of the TG image
(the normal TG and tonsillitis) using neural network
(NN) which will be described further in Section 3.3.
In addition, there are usually many methods for
classification of the normal TG and tonsillitis such
as histogram intersection or Euclidean distance etc.
However, there are limitations of such methods for
classifying the TG image because they will classify
only the data group in a linear model. In this work,
the scattering plot is not clear for a separation of
two data groups; however, some data also have overlapped
between the normal and tonsillitis, as will be
seen later in Section 3.3. As a result, it is difficult
to classify both groups for a correction of the results.
Therefore, this work proposes to use neural network
for classification of the TG image i.e. the normal TG
and tonsillitis.
The average intensity of TG exploiting the intensity
of R, G and B in the TG area can be described
as the formula
A =
n∑ ∑
r + n ∑
g + n b
i=1
i=1
3n
i=1
(3)
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 39
where A is the average intensity of TG area,
r is the intensity of red,
g is the intensity of green,
b is the intensity of blue,
n is the total pixels of the TG area.
Then, the average intensity (A) of TG area is normalized
by 255 for a suitable value as following equation
N =
A
255
where N is the normalized value of A (0 to 1).
(4)
Figure 7 shows an example for displaying the values
of intensity (red, green and blue) on the area of
TG. For example, it can be seen from Figure 7 that
the values of a pixel (in case of n =1) selected from
the TG area consist of r = 198, g = 118 and b = 69.
Therefore, the average intensity (A) using equation
(3) is equal to (198+118+69)/3 = 128.33. In case of
n 1, the average intensity (A) is calculated from the
TG area for n 1 and, the summations of r, g, and b
are obtained from n 1. Finally, the average intensity
(A) will be normalized by 255 for the suitable value,
as shown in equation (4).
be an inflammation of the TG. Therefore, the 2D FFT
and power spectrum of 2D for calculating the result
of the surface on the TG will be defined as follows [7]
F (u, v) = 1
MN
M−1
∑
x=0
N−1
∑
y=0
f(x, y)e −j2π(ux/M+vy/N) (5)
where F(u,v) is the 2D FFT result of picture,
N,M are width and height of picture,
f(x,y) is brightness value at (x,y).
The result from the equation (5) is taken to calculate
the power spectrum of 2D as the equation
P (u, v) = |F (u, v)| 2 = R 2 (u, v) + I 2 (u, v) (6)
where P is the power spectrum of 2D,
R is real part,
I is imaginary part.
Figure 8 shows an example of the resulting image
using the 2D FFT on the area of TG. It can be seen
from Figure 8 that the binary image of the TG area
is applied for equation (5). Therefore, the 2D FFT
image of the TG area is shown in Figure 8. Figure 9
shows a resulted graph derived from Figure 8 around
the TG area using the 2D FFT. It can be seen from
Figure 9 that the calculated value of power spectrum
of 2D FFT image is at approximately 9.05 × 10 4 i.e.
it will be the purulent surface on the TG (or tonsillitis).
Note that the power spectrum (P) of 2D FFT
is derived from the resulting values in equation (5)
substituted in equation (6).
Fig.7: An example for displaying the values of intensity
(red, green and blue) on the area of TG.
3.2.3 The purulent surface on TG
The third factor is to calculate a result of a surface
on TG using the power spectrum of 2 dimensional
Fast Fourier Transform (2D FFT). For example, if the
calculated result reveals smooth surface on the TG,
then it may be a normal TG whilst if the calculated
result shows purulent surface on the TG, then it may
Fig.8: An example of the resulting image using the
2D FFT on the area of TG.
40 A. Leelasantitham and S. Kiattisin: A Diagnosis of Tonsillitis using Image Processing and Neural Network (36-42)
Fig.9: An example of a graph for the power spectrum
of the 2D FFT image around the TG area.
Fig.10:
factors.
The scattering plot of the three important
3.3 Verification
3.3.1 Data Analysing
This section describes an analysis of the data consisting
of three important factors based on medical
knowledge. Such three important factors are as follows:
a) the dimensional ratio of TG,
b) the average intensity of TG and
c) the purulent surface of TG.
The values of these three factors are plotted on
three axes consisting of the width/height (W/H), the
average intensity and the purulent surface for the
first, second and third axes, respectively. The plotted
values of the W/H, the average intensity and the purulent
surface are derived from the equations (2), (4)
and (6), respectively. Therefore, the scattering plot
of the three important factors will be shown in Figure
10. As shown in Figure 10, symbols of ‘•’ and ‘×’ are
represented to the usual TG and tonsillitis, respectively.
It can be seen from Figure 10 that the data
groups of ‘•’ and ‘×’ cannot be classified clearly in
terms of a simple method of linear analysis because
there is some data overlapped between two groups.
Therefore, neural network is an alternative method
for solving this classification which will be described
further in Section 3.3.2.
3.3.2 Neural Network
This section designs neural network (NN) using a
WEKA package [8] which there are 3 layers i.e. input,
hidden and output layers, as shown in Figure
11. Numbers of nodes consist of 3, 2 and 2 nodes
for the input, hidden and ouput layers, respectively.
It can be seen from Figure 11 that the three factors
(the W/H, the average intensity and the purulent surface)
are inputted to the input layer whilst there are 2
nodes for the output layer consisting of the usual TG
(value = 0) and tonsillitis (value = 1). The equation
for calculation of backpropagation neural network can
be defined as [9]
∆W ij = ηδ i x j + α∆W ij (k − 1) (7)
which W ij is weight to connect between output
node j and input node i
η is step size
α is coefficiency of momentum
x j is input signal from node j
δ i is the quantity of difference of error , δ i can be
shown as below
for output unit
otherwise
δ i = {O ′ i(t i − O i )
O ′ i is derivative of O i
T i is a required signal to node i
O i the actual output at node i.
Fig.11:
O ′ i
∑
Wmi δ m (8)
Designing three layers of neural network.
4. EXPERIMENTAL RESULTS AND DISS-
CUSIONS
This section will verify the proposed method for
the detection of tonsillitis images based on medical
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 41
knowledge and neural network (NN). This paper uses
a MATLAB program for implementation of the first
part of image processing and the second part of NN
using a WEKA package [8]. Details of preparing the
neural network can be described as follows:
a) To demonstrate data inputs for training of neural
network, there are 25 images of tonsillitis whilst
there are 25 images of the usual TG.
b) To demonstrate data inputs for testing the detection
of tonsillitis, there are 100 images of TG. They
are taken into NN.
Table 1 shows the testing results of verifying TG
images which the output results are the usual TG or
tonsillitis. It can be seen in Table 1 that the 100
images of TG are tested by both of the medical doctor
and the proposed method. As shown in Table
I, the results from the medical doctor can diagnosis
the following outputs i.e. 40 and 60 images for the
usual TG and tonsillitis, respectively; whilst the results
from the proposed method can classify the following
outputs i.e. 35 and 55 images for usual TG
and tonsillitis, respectively. There is an error image
of them for the usual TG which their accuracy is at
approximately 87.5%, whilst; there is also an error
image of them for the tonsillitis which their accuracy
is at approximately 91.67%. Therefore, the overall
accuracy is at approximately 90%.
Table 2 shows summary results of the true or false
positive, and the true or false negative. It can be seen
in Table 2 that there are 35 images for the true positive
being the usual TG whilst there are five images
for the false positive being the tonsillitis. As shown
in Table 2, there are 55 images for the true negative
being the tonsillitis whilst there are five images for
the false negative being the usual TG. From the experimental
results, the correct results depend on clear
images, and NN is considered in terms of numbers of
training set. If there are more samples of training set,
the new results will chance an increment from the old
results.
Table 1: The testing results of verifying TG images.
Output
Results
Correct
Results
Correct
Results
Accuracy
(%)
of TG From From The
Images Medical
Doctor
Proposed
Method
Usual TG 40 35 87.5
Tonsillitis 60 55 91.67
Table 2: Summary results of the true or false positive,
and the true or false negative.
True False
Positive 35 5
Negative 55 5
5. CONCLUSION
This paper has presented the diagnosis of tonsillitis
using image processing and neural network (NN).
There are three steps described as follows. The first
step is localization of tonsil grand (TG) using the Ellipses
Hough Transform. The second step is feature
extraction using three important factors which can
be indicated in swelling by images of TG in terms of
a) a dimensional ratio of TG, b) an average intensity
of TG and c) a purulent surface of TG (yes/no) using
the power spectrum of two dimensional Fast Fourier
Transform (2D FFT). The final step is verification using
NN. The three factors are inputted into NN, and
TG samples of 50 images are used for training into
the NN. They are divided by tonsillitis patience 25
images and usual TG of 25 images. This experiment
uses 100 images of TG for testing NN. Therefore, the
overall accuracy is at approximately 90% in terms of
comparing with the results from the medical doctor.
6. ACKNOWLEDGMENT
The authors are grateful to Mr. Kritchanon Jirawanitcharoen
for his useful help in this work.
References
[1] Medical Doctor Worawut Chareonsiri:
http://www.bangkokhealth.com
[2] P. Phensadsaeng, P. Kumhom, and K. Chamnongthai,
2005, “A Method of Vision-Based Tonsillitis
Detection”, Sixth International Conference
on Intelligent Technologies (InTech’05), 14-
16 December 2005, Phuket, pp. 789-792.
[3] P. Phensadsaeng, P. Kumhom, and K. Chamnongthai,
2006, “A Computer-aided-Diagnosis of
Tonsillitis Using Tonsil size and color”, International
Symposium on Circuits and System (IS-
CAS2006), Greece, May 21-24, 2006, pp. 5563-
5566.
[4] Medical Doctor Panuwit Pumhirun:
http://www.vichaiyut.co.th
[5] K. Jirawanitcharoen, S. Kiattisin, A. Leelasantitham
and P. Chaiprapa, “A Method of Detecting
Tonsillitis Images Based on Medical Knowledge
and Neural Network”, The 2009 International
Conference on Artificial Intelligence
and Neural Networks (ICAINN 2009), Beijing,
China, August 8-11, 2009.
[6] X. Yu, H.W. Leong, C. Xu and Q. Tian, “A robust
and accumulator-free ellipse hough transform”,
Proceedings of the 12th ACM International
Conference on Multimedia (ACM Multimedia
2004), New York, USA, October 10-16,
2004, pp. 256-259.
[7] C. Gonzalez, Rafael and E. Woods, Ricard, Digital
Image Processing 2nd , Prentice Hall, 2002.
[8] I. H. Witten and E. Frank, Data mining: Practi-
42 A. Leelasantitham and S. Kiattisin: A Diagnosis of Tonsillitis using Image Processing and Neural Network (36-42)
cal machine learning tools and techniques (2 ed.),
Morgan Kaufmann, San Francisco, CA, 2005.
[9] Dan W. Patterson, Artificial Neural Networks
Theory and Application, Prentice Hall, 1996.
Adisorn Leelasantitham received the
B.Eng. degree in Electronics and
Telecommunications and the M.Eng.
degree in Electrical Engineering from
King Mongkut’s University of Technology
Thonburi (KMUTT), Thailand, in
1997 and 1999, respectively. He received
his Ph.D. degree in Electrical Engineering
from Sirindhorn International Institute
of Technology (SIIT), Thammasat
University, Thailand, in 2005. He is currently
the Assistant Professor in Department of Computer and
Multimedia Engineering, School of Engineering, UTCC, Thailand.
His research interests include analog circuits, image processing,
medical images, computer graphics, AI, neural networks,
embedded systems and robotics.
Supaporn Kiattisin received B.Eng.
in Computer Engineering from Chiangmai
University in 1995, M.Eng. in
Electrical Engineering and Ph.D. in
Electrical and Computer Engineering
from King Mongkuts University of Technology
Thonburi (KMUTT), Bangkok,
Thailand. She currently works at
Computer Engineering and Multimedia,
School of Engineering, University of the
Thai Chamber of Commerce (UTCC).
Her research interests include medical imaging, computer vision
and modeling. She is member of TESA, ThaiBME, IEICE
and IEEE .
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 43
Treatment Planning Technique Applying the
Combined CT Imaging and Computational
Fluid Dynamics (CFD) Analysis
K. Hemtiwakorn ∗ , N. Phoocharoen, Guest members,
V. Mahasittiwat, and M. Sangworasil, Members
ABSTRACT
Combined medical imaging with computational
fluid dynamics (CFD) analysis was used to demonstrate
and evaluate airflow in human nasal cavity.
The processes of this study are segmentation, meshing,
solving, and post-processing. Firstly, Mimic
10.01 software was used to segment the nasal cavity.
Secondly, Pro-STAR/amm software was employed to
generate trimmed mesh of 338,496 elements. Finally,
computational grid model was imported to STAR-CD
version 3.24 for numerical computation, and visualize
the solution by post-processing. The results show
the nasal airflow simulation in both normal and abnormal
human noses. It can be seen that for the
normal nose breathing, the velocity magnitude of airflow
was greatest in a nasal valve area which is the
narrowest area of the human nose. Also, the air flow
commonly passes through the main nasal passage and
middle meatus areas, and velocities in both sides are
not different. For the abnormal nose of nasal septum
deviation, the result showed that the velocity
magnitude inside the narrowed left side is very high
and greater than the right side about 50-90 percent.
Therefore, combined CT data and CFD analysis is
very useful for doctor to evaluate the physiology and
phenomena of airflow inside the patient’s nose. Rhinosurgion
could be employed those technique for the
preparing the rhinosurgical planning. Nevertheless,
the experimental validation is not reported in this paper
but will be published in the further publication.
In conclusion, this study is the important beginning
of an applied CFD analysis with medical imaging.
CT imaging combines with CFD analysis could be
useful for rhinologist as a nose function evaluation
technique. CFD based on CT imaging could be applied
as a tool for the treatment planning.
Keywords: CFD, Nasal airflow, Treatment planning
* Corresponding author.
Manuscript received on December 25, 2009.,
K. Hemtiwakorn is with the Biomedical Engineering, Faculty
of Engineering, King Mongkut’s Institute of Technology
Ladkrabang, Bangkok, Thailand.
E-mail addresses: hemtiwakorn@gmail.com (K. Hemtiwakorn)
1. INTRODUCTION
Nose is an upper part of the respiratory tract which
is the opening of a human airway. Three main functions
of nose [1] are filtering, smelling, and optimization
temperature and humidity to the ambient air
before going through the lung. Since, nasal cavity of
human is very narrow and complicate,visualization
of nasal cavity is very difficult. The major symptoms
such as difficult breathing, loss of smell, or sinus infection
can increase the resistance to flow [2], and also
may be occur from the obstruction or nasal septum
deviation. The methods of rhinomanometry, acoustic
rhinometry, or medical imaging, such as CT or
MRI, can help the doctor to diagnose the pathology
in human nose, and treatment planning. However,
previous methods can not demonstrate the function
of nose, such as airflow, in normal or abnormal case.
CFD based on CT imaging and simulation software
is the useful technique to evaluate the airflow direction,
velocity, and pressure inside the nasal cavity for
pre and post operations. The combination of medical
imaging technique, such as CT or MRI, and computational
fluid dynamics (CFD) analysis could be applied
for the diagnosis, treatment planning, and evaluate
the physiology of the nose after operation. Thus, this
study proposed the combined CT imaging and CFD
analysis as a tool help doctor for planning the nose
operation.
2. NUMERICAL MODELING BASED ON
CT IMAGING
2.1 GEOMETRIC MODEL CONSTRUC-
TION
The first process of this research is construction
the geometric 3-dimensional nose model by using the
computed tomographic (CT) data obtained from volunteer
and segmentation software. DICOM data of
CT images were imported into MIMICS 10.01 software
(Materialize, USA) for segmenting the nasal
airway excluded paranasal sinus. A threshold level
range of about -1024 to -188 was applied to the area
consisting of air. After automatic segmentation with
threshold level of brightness, the manual segmentation
should be performed not only in axial images but
also in coronal and saggital images as shown in Fig.1.
44 K. Hemtiwakorn et al: Treatment Planning Technique Applying the Combined CT Imaging ... (43-49)
Finally, 3-dimensional nose model was reconstructed
as demonstrated in Fig.1.
2.2 COMPUTATION MESH GENERATING
The mesh generating software called Pro-STAR/amm
software (CD-Adapco, UK) was used for this purpose.
Unstructured mesh was generated in this study.
Also, it is necessary to create at least two cell layers
immediately next to the inner boundary of the
3-dimension model to obtain a stable and convergent
solution. Therefore, two layers were located near to
the wall. Fig.2 shows the trimmed-cell mesh with two
layers cell structure near the model surface.
2.3 FLUID DYNAMICS CONDITION/ GOV-
ERNING EQUATION
Airflow inside a human nasal cavity is in a dynamic
state. It decelerates and accelerates from a resting
state during normal breathing. The air is considered
as incompressible and Newtonian with constant fluid
properties. A numerical solution of the mean flow requires
resolving the mass and momentum conservation
equations (the “Navier-Stokes equations”) solved
by STAR-CD are in Cartesian tensor notation [1] as
follows:
∂
∂x j
(ρ · u j ) = 0 (1)
∂
(ρ · u j · u i − τ ij ) = − ∂p (2)
∂x j ∂x j
where x i is the Cartesian coordinate (i = 1,2,3)
u i = absolute fluid velocity component in direction
x i ,
p = piezometric pressure = p s − ρ 0 g m x m where ps
is static pressure, ρ 0 is reference density, the g m are
gravitational acceleration components and the x m are
coordinates relative to a datum where ρ 0 is defined
ρ = density
τ ij = stress tensor components
Standard low Reynolds number two-equations k−ε
model is the turbulence model applied in this study.
2.4 BOUNDARY CONDITION AND NU-
MERICAL CONTROL
The computational grid nose model was imported
into the commercial CFD software package, STAR-
CD 3.24 of Adapco, UK, which solve the finite volume.
Discretization of the governing equations was
conducted using a Mono Tone Advection and Reconstruction
Scheme (MARS), which is a secondorder
accurate differencing scheme. Of all differencing
schemes available in STAR-CD, MARS was chosen
for its optimal sensitivity to the solution accuracy
and skewness of the mesh structure [2]. The
Semi-Implicit Method for Pressure-Linked Equations
(SIMPLE) algorithm [3] was used to manage the
pressure-velocity coupling. To stabilize the solution,
under-relaxation factors were used for all the basic
variables (0.5 for momentum and turbulent equations
and 0.1 for pressure). The sets of linearized and discretized
equations for all variables were solved using
the algebraic multigrid method for improving the calculation
performance of trimmed meshes [2]. We assumed
that the uniform velocity profile and the axial
component of velocity were perpendicular to the flow
inlet faces. The boundary conditions at the nostrils
and the nasopharynx were set to the static pressure
= 0 and outlet boundary face is fixed mass flow rate
(kg/s). Turbulence intensity was set to 10% and the
dissipation rate at the inlet boundaries was set with
a dissipation length scale of 1 cm. Convergence was
judged by monitoring the magnitude of the absolute
residual sources of mass and momentum, normalized
by the respective inlet fluxes. The iteration is continued
until all above residuals fall below 0.001
3. INSPIRED NASAL AIRFLOW SIMULA-
TION RESULTS
In this study, the normal breathing in human was
studied in the first investigation. Next, the abnormal
breathing during the pathology in the patient’s
nose was studied. Comparison between normal and
abnormal breathing are useful for doctor to evaluate
the patient status and treatment planning. The results
of the airflow simulation in both normal and
abnormal noses are demonstrated in the following.
3.1 NORMAL BREATHING IN HUMAN
NOSE
Computer modeling was constructed using CT
data obtained from a female patient. CT scanning
was performed on a General Electric LightSpeed VCT
(Wisconsin, USA) scanning station with the following
parameters: 120 kVp, 200.0 mA, rotation time of
0.9 s, thickness 1.25 mm. To construct an anatomical
3-dimensional model of nasal cavity, 48 axial sections
from a computed tomographic image of a single subject
were used for the numerical modeling. The computations
were performed using a personal computer
with an Intel Core 2 Quad 2.4 GHz CPU and 6 GB
of memory, which typically took 2 hours per run to
complete.
The velocity magnitude of airflow in right nasal
cavity is visualized in saggital view as shown in Fig 3.
From this figure, it can be seen that the greatest velocity
magnitude occurs in nasal valve area, because
it is the narrowest area of the nose. Also, high values
and variation of velocity magnitude appears in the
anterior part of nasal airway in front of a turbinate
region. The flow was slower when passed through
the main nasal passage and turbinate regions because
there were wider than the anterior part of the nasal
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 45
Fig.1: CT images in 3 planes of coronal, axial, and saggital were used for segmentation. Magenta color
areas demonstrate a nasal airway (exclude paranasal sinus) after automatic and manual segmentations. 3D
reconstructed nose model is shown in this figure.
Fig.2: Trimmed cell mesh model was performed after constructing the 3D model obtained from CT images.
Cross-sectional grid models in coronal and saggital planes are shown in this figure.
cavity. For the airflow in turbinate meatus region,
almost flow pass through the middle meatus which is
wider than another meatus. Thus, slow flow passes
through the inferior and superior meatus.
46 K. Hemtiwakorn et al: Treatment Planning Technique Applying the Combined CT Imaging ... (43-49)
Fig.3: Velocity magnitude of the right nasal cavity is shown in a saggital plane. Landmark line of cutting
plane location is demonstrated.
Fig.4:
Coronal cutting planes of nasal model.
Moreover, cutting model in coronal plane to visualized
velocity magnitude in specific regions was
presented. The velocity magnitudes of airflow were
demonstrated in six planes as shown in Fig.4. Plane
A is a nostril. Plane B represents the nasal valve
area. Plane C is the head of inferior turbinate region.
Plane D is the beginning of middle turbinate region.
Plane E represents the middle part of nasal cavity,
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 47
Fig.5: The velocity magnitude (m/s) of nasal airflow in each coronal plane (A-F) obtained from normal
nose model.
and plane F is the beginning of superior turbinate
region.
The velocity magnitudes in each plane are presented
in Fig. 5. The results show that velocities increase
immediately in the nasal valve area. The highest
velocity magnitude of 1.88 m/s occurs in plane B
which is the nasal valve area. Moreover, it can be
seen that airflow commonly passes through the main
nasal passage and the middle meatus. However, slow
flow of 0.13-0.26 m/s appears in the inferior and superior
meatus. From the Fig.5, it can be seen that
flow commonly passed near the nasal septum more
than diffused to the nasal wall.
3.2 ABNORMAL BREATHING IN DEVI-
ATED NASAL SEPTUM
Computer modeling was constructed using CT
data obtained from a male patient with the nose obstruction
due to the nasal septum deviation. To construct
an anatomical 3-dimensional model of nasal
cavity, 64 axial sections from CT images of a single
subject were used for the numerical modeling.
From the CT data, the images show that nasal septum
is deviated to the narrowed left side of patient’s
nose as shown in Fig 6. To confirm the diagnosis
results of nasal septum deviation and narrowed left
side, rhinoscan was performed in the same patient.
The rhinoscan result shows that the left side is very
narrow as compare with the right side as shown in
Fig 7. Also, the inspired volume flow rate in a left
side is lesser than a right side as 34.6 cm2/s and 219.3
cm2/s for left and right sides, respectively.
Fig.6: CT image obtained from a male patient of
nasal septum deviation.
The first airflow simulation result in abnormal nose
model is demonstrated in Fig 8. The figure shows the
velocity magnitudes of airflow inside a nasal cavity
with a deviated nasal septum and very narrow in the
left side of nasal cavity. It can be seen that the velocity
magnitudes of airflow inside the left nasal cavity
are greater than the right side, particularly for the
obstructive area approximately 1.5-2.0 m/s for a left
side and 1.2 m/s for a right side, respectively. As
comparing with the normal nose model shown in Fig.
5, it can be seen that the velocity magnitude of airflow
48 K. Hemtiwakorn et al: Treatment Planning Technique Applying the Combined CT Imaging ... (43-49)
inside the both nasal cavities is not much difference.
This study presents nasal airflow simulation in
normal and abnormal adult noses by combining CT
imaging with computational fluid dynamics (CFD)
analysis. It enables us to study the physiology of
nasal breathing. The advantage of this method is the
airflow simulation on a realistic model derived from
CT imaging. Moreover, combined CT data and CFD
analysis is useful for doctor to diagnosis, evaluate the
phenomena of airflow inside the patient nose, treatment
planning, and pre-post operative airflow simulation.
6. ACKNOWLEDGMENT
The authors wish to acknowledge the facilities of
computer and software provided by the Computer
Service Center of King Mongkut’s Institute of Technology
Ladkrabang (KMITL), Bangkok, Thailand.
Moreover, authors would like to thank Mahidol University
and SWU, Bangkok, Thailand, for their supporting
the CT images. We would like to thank Assoc.
Prof. Manas Sangworasil for his kindly suggestion,
support, and encouragement throughout this research.
Finally, the author wish to thank the Office
of the Higher Education Commission, Bangkok, Thailand
for the financial support during this research.
Fig.7: Rhinoscan result showed that a left side of
nose is narrow.
4. APPLICATION OF CFD FOR A TREAT-
MENT PLANNING
To correct the abnormal nasal cavity of the patient,
treatment technique of surgery is an invasive
technique but necessary particularly for the case used
in this research. Therefore, preparation the surgical
planning is very important. Combined CT data and
CFD analysis is very useful for doctor to evaluate the
physiology and phenomena of airflow inside the patient’s
nose. Doctor can understand the affect of abnormal
nose geometry to the direction, velocity, and
pressure of airflow inside the nose. Moreover, rhinosurgion
could be applied those technique for the
rhinosurgical planning, and also visualize the nasal
airflow obtained from the process of pre-operation
preparation. Nevertheless, nasal airflow obtained
from pre-post operation can be calculated by a CFD
technique but does not show in the present. Also,
the experimental validation of nasal airflow simulation
obtained from this study will be reported in the
future publication.
5. CONCLUSION
References
[1] Z.V.A. Warsi, Conservation form of the Navier-
Stokes Equations in General Nonsteady Coordinates,
AIAA Journal, 19, pp.240-242, 1981.
[2] STAR-CD 3.2 Users’ Guide. Computational Fluid
Dynamics Ltd., 2004.
[3] S.V. Patankar, D.B. Spalding. A Calculation Procedure
for Heat, Mass and Momentum Transfer in
Three-Dimensional Parabolic Flows. Int. J. Heat
Mass Tranfer, 15, pp. 1778-1806, 1972.
K. Hemtiwakorn obtained her B.Sc
and M.Sc in radiological technology in
2005 and 2008, respectively from the
Mahidol University, Bangkok, Thailand.
She is currently pursuing her Ph.D in
biomedical engineering, Faculty of Engineering,
Bangkok,Thailand. She also
the research assistant in Ramathipbodi
Hospital in several research projects
about radiological technology and applied
biomedical engineering. Her research
interests include computational fluid dynamics (CFD)
in human body, biomechanics, and apply the CFD for evaluate
and treatment planning in the ENT patient.
N. Phoocharoen obtained his B.Eng
in Mechanical Engineering from King
Mongkut’s Institute of Technology Ladkrabang
(KMITL), Bangkok, Thailand.
He is currently pursuing his M.Eng
in Mechanical Engineering, KMITL,
Bangkok, Thailand. He has more than
10 years experience in CFD research in
both mechanical and biomedical engineering.
He is now the research assistant
in the Computer Service Center,
KMITL, Bangkok, Thailand.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 49
Fig.8: The velocity magnitude (m/s) of airflow inside the abnormal nose model of deviated nasal septum in
each coronal plane (A-H).
V. Mahasithiwat obtained his MD
from Prince of Songkla University
(PSU). He is currently the ENT expert
in Faculty of Medicine, Srinakharinwirot
University. His research interests
include ENT technology and development
of earring aids, nose operation
technique, and eye tracking.
M. Sangworasil was born in Bangkok,
Thailand in 1951. He received the Bachelor
of engineering and Master o Engineering
from King Mongkut’s Institute
of Technology at Ladkrabang, Bangkok,
Thailand in 1973 and 1977 respectively,
and the D. Eng (Electronics) from Tokai
University, Japan, in 1990. Following
his graduate studies, he worked almost
28 years at Electronic De- partment,
Faculty of Engineering, King Mongkut’s
Institute of Technology at Ladkrabang, Bangkok where he is
currently an associate professor. His research interest are in the
area of image process with emphasis on Imag Reconstruction,
3D modeling, Image Classication and Image Filtering.
50 K. Jaruwongrungsee et al: Analysis of Quartz Crystal Microbalance Sensor Array with ... (50-54)
Analysis of Quartz Crystal Microbalance
Sensor Array with Circular Flow Chamber
K. Jaruwongrungsee 1,2,∗ , T. Maturos 1 , P. Sritongkum 1 ,
A. Wisitsora-at 1 , Guest members, M. Sangworasil 2 , and A. Tuantranont 1 , Members
ABSTRACT
In this paper, quartz crystal microbalance
(QCM) sensor array integrated with circular
poly(dimethylsiloxane) (PDMS) chambers is developed
for flow injection based bio-sensing. An
array of QCM sensors was fabricated on a single
quartz substrate by Cr/Au sputtering through
shadow masks and integrated with PDMS chambers,
made by PDMS based microfluidic fabrication technology.
Gold electrodes of QCM sensors were functionalized
with carboxylic group and flow injection
analysis was conducted for protein G binding. The
resonance frequencies for four sensors were continuously
monitored during protein G injection in a constant
flow of PBS buffer solution. It was found that
there was a significant variation in resonance frequency
shift responses of identical QCM sensors in
the circular QCM chamber. The cause of such undesired
variation was then analyzed by fluid dynamic
simulation. The simulation results reveal that the
flow in a circular-shaped QCM chamber is primarily
turbulent. In addition, the degree of turbulence is
increased with flow rate. Thus, sensors at various locations
see different sample dispersions causing their
sensing behaviors to be significantly different.
Keywords: Quartz crystal microbalance; QCM
1. INTRODUCTION
Quartz crystal microbalance (QCM), an ultrasensitive
mass sensor based on piezoelectric effect, is
one of the most powerful methods for chemical and
biological sensing. QCM has a wide range of applications
in many fields such as thin-film measurement,
chemical analysis [1], gas sensor [2], humidity sensor
[3-4], and biosensor [5]. When the electric power
is applied to a pair of electrodes, sandwiched between
quartz crystal, mechanical force is generated
* Corresponding author.
Manuscript received on January 29, 2009.,
1 1Nanoelectronics and MEMS Laboratory, National Electronics
and Computer Technology Center, Thailand
telephone: +662-564-6900 ext.2108, Fax: +662-564-6756
2 Department of Electronics, Faculty of Engineering, King
Mongkut’s Institute of Technology Ladkrabang, Thailand
E-mail addresses: kata.jaruwongrungsee@nectec.or.th
(K.Jaruwongrungsee) adisorn.tuantranont@nectec.or.th
(A. Tuantranont)
via piezoelectric effect. With this effect, quartz crystal
will resonate at its natural frequency by positive
feedback through oscillator circuit. The oscillation
frequency depends on many factors including quartz
thickness, quartz density, type of cut, ambient conditions
(temperature, pressure, humidity, etc) and most
importantly deposited mass.
For the reason that oscillated frequency can be
changed as a function of mass, quartz crystal has
been used as a kind of mass sensor. AT-cut, the most
popular cut of quartz crystal, was chosen on account
of its frequency stability around room temperature.
In 1959, Sauerbrey derived the equation of frequency
shift of the quartz resonator in gas phase [6]:
∆f = −
2f o
2 ∆M
√
ρq µ Q A
(1)
where: ∆f is the frequency shift of the resonator,
f0 is the fundamental frequency, ρ q is the density
of quartz (2.648 g/cm 3 ), µ q is the shear modulus of
quartz (2.947×1011 g/cm×s 2 ), ∆M is the mass deposited
on the surface of electrode and A is piezoelectrically
active area (2/(ρ q µ q ) 1/2 can be expressed
as a constant, k, which is equal to 2.26×10 −7 ). This
equation can describe only the added mass rigidly
deposited on the electrode surface. For the attached
liquid molecules, the frequency shift can be described
by equation of Kanazawa and Gorgon [7], which use
to find the frequency shift based on the liquid properties
of the liquid on electrode surface:
∆f = −f 3 2 o
√
ρL η L
πρ q µ q
(2)
where: ρ L and η L are the density and absolute viscosity
of the liquid, respectively.
In the last few years, many researchers interested
in QCM based sensor-array for advanced sensing
applications, including electronic nose, electronic
tongue and biosensor array. However, majority of
reported QCM sensor arrays are still based on simple
combination of single QCM sensors from different
substrates [8-10]. The main problem of
this scheme is the mismatch characteristics among
QCM devices due to different piezoelectric properties,
quartz-thickness, temperature, pressure, mass
and material properties of electrode layers. To min-
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 51
(a)
(b)
Fig.1:
Schematic of QCM sensor array (a) electrode layout and (b) electrodes superimposed by PDMS
imize the matching problem, QCM sensors in an array
should be fabricated on the same quartz substrate
under the same process parameters and conditions
[11-12]. In addition, the effect of environmental
conditions including temperature and humidity
can be greatly reduced by employing one
QCM in the matched array as the reference sensor
so that the change in QCM characteristics due
to ambient parameter can be effectively canceled.
In this work, a QCM based biosensor array is integrated
with poly(dimethylsiloxane) (PDMS) microfluidic
platform for flow-injection based detection.
2. SENSOR FABRICATION
2.1 QCM Sensor Array Fabrication
An array of QCM sensors was fabricated on a single
disc of 1-inch 5 MHz AT-cut quartz crystal substrate.
The number of QCM sensors on a quartz substrate
was designed as 4 and the patterns of electrode
pair were laid out symmetrically within 1-inch
circle as illustrated in Figure 1 (a). Since QCM
electrodes were located on both sides of quartz disc,
it is difficult to make electrical connection for testing.
To circumvent this problem, the electrode layouts
for both sides were made different. The back
side patterns contain contact pads for both electrodes
while the front side patterns have no contact
pad but have connection line that run to the pad
onto the other side of quartz disc through the circumference.
The Cr/Au electrode was deposited on
quartz substrate by sputtering through electroplated
microshadow-mask. Microshadow-mask was fabricated
by Ni electroplating on photoresist-patterned
stainless-steel plate. The standard dry-film photolithography
process was used to produce photoresist
pattern of electrode designs on stainless steel
plate. Uncovered area on stainless-steel plate was
then coated by 40 m thick Ni film by electroplating
process. Ni electroplating was conducted in Nickel
sulphate plating solution for 4 hours. Nickel shadow
masks were obtained by removing the photoresist by
sodium hydroxide solution and detached the nickel
film from stainless steel plate.
Before QCM electrode deposition, blank quartz
disc substrates were cleaned in piranha solution (1:4
mixture of 50% H2O2 and 97% H2SO4) at 120 oC
for 10 min. The shadow mask was then aligned and
attached on a blank quartz disc by a permanent magnet.
The chromium (Cr)/ gold (Au) layers were deposited
on one side of quartz substrates by sputtering
through the first set of electroplated microshadowmasks.
The Cr and Au layers were successively sputtering
under argon dc plasma. Before each sputtering
run, the substrate was cleaned by 75W RF plasma for
5 min to improve adhesion to underlying material.
The sputtering pressure, sputtering current and time
for chromium were 3x10-3 mbar, 0.2 A and 2 min, respectively.
Next, gold was sputtered under the same
current and pressure for 10 min. The 50 nm-thick Cr
and 300 nm-thick Au layers were obtained. The other
Cr/Au layers were then sputtered on the other side of
quartz disc through the second set of microshadowmasks.
The second shadow masks were aligned to the
first pattern via a cross marker at the center of the
disc.
2.2 PDMS Micro Chamber Fabrication
Silicon wafer were cleaned in piranha solution at
120 oc for 10 min, carefully rinsed several time in
deionized water and dried with gentle stream of air.
After that silicon wafer were dehydrated at 150-200
oC for 10 min. SU-8 photoresist was spin-coated on
silicon wafer using a spin coater (Laurell technologies
corp. model WS-400A-6NPP), then soft baked
to remove all the solvent in the layer. The photoresist
coated wafers were exposed using MJB4 mask
aligner (SUSS microtec) then post-baked in order to
selectively cross-link the exposed portions of the film.
The sample was left in the desiccators to cool down
slowly at room temperature for more than 13 hours.
52 K. Jaruwongrungsee et al: Analysis of Quartz Crystal Microbalance Sensor Array with ... (50-54)
Finally sample was developed, cleaned with deionized
water and isopropyl alcohol and gently dried with air.
Spin speed, exposure time, baked time and developing
time were optimized to achieve a smooth surface
on mold. The SU-8 mold thickness was investigated
by interferrometer (Polytec MSA400), the measured
result show that the mold thickness is 214.9 µm.
Sylgard 184 Silicone Elastomer kit (Dow Corning),
consisting of PDMS was prepared by mixing the precursors
sylgard with a curing agent at a ratio of 10:1
by volume. The prepolymer mixture was degassed at
20-50 mTorr in ambient temperature desiccator with
a mechanical vacuum pump for 10 min to remove any
air bubbles in the mixture. PDMS mixtures were
gradually poured onto the SU8 master mold to the
height over the depth of designed chamber. Next,
PDMS slab was cured at 80 oc for 3 hour. Finally, it
was peeled-off from the mold. Figure 2 shows the photograph
of fabricated QCM sensor array with PDMS
chamber.
with circular design. The fluid dynamic simulation
was carried out by imposing a constant sample flow
rate condition throughout the chamber.
Table 1: Measured frequency shift from the experiment
Sensor Number Measured Frequency Shift (Hz)
1 52
2 44
3 163
4 73
Fig.3:
Meshed circular design of sensor chamber
Fig.2:
QCM sensor array with PDMS chamber
3. EXPERIMENTAL RESULTS AND DIS-
CUSSIONS
The sensor array was prepared for protein G binding.
The surface of gold electrode was functionalized
with carboxylated polyvinyl chloride (PVC-COOH)
by self-assembly monolayer (SAM) technique. Flow
injection measurement was conducted for protein G
binding. The resonance frequencies for four sensors
were continuously monitored during protein G injection
in a constant flow of PBS buffer solution. It was
found that there was a significant variation in resonance
frequency shift responses of identical QCM
sensors, as shown in table 1. In order to analyze this
problem, computational fluid dynamics (CFD) simulation
is used to design and simulate sample transport
in flow injection QCM sensing system.
The 3D-simulation of fluid flow through QCM sensors
was performed by ANSYS program. Figure 3 illustrates
the finite element model of QCM chamber
Figure 4 shows sample dispersion trajectories at
various flow rates including 50 µl/min and 100
µl/min. Simulation results show that the flow in
this circular-shaped QCM chamber design is primarily
turbulent. In addition, the degree of turbulence
is increased with flow rate. The sample dispersion
especially on both far sides of chamber where sensor
electrodes are located is nonlinear. The sensors
at various locations see different sample dispersions
causing their sensing behaviors to be significantly different.
Thus, the simulation results can explain the
observed signal variation among four QCM sensors
in the circular PDMS chamber. In order to solve this
problem, QCM chamber will be redesigned with different
geometries and the results will be presented
elsewhere.
4. CONCLUSIONS
In conclusion, quartz crystal microbalance (QCM)
sensor array integrated with poly(dimethylsiloxane)
(PDMS) chambers has been designed and fabricated
for flow injection based bio-sensing. It was found
that there was a significant variation in resonance
frequency shift responses of identical QCM sensors
in the circular QCM chamber. The cause of such
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 53
(a)
(b)
Fig.4: The geometries of turbulent flow effect inside
circular chamber in 3D simulations and velocity variation
over the sensor surface at different flow rate (a)
50 µl/min, (b) 100 µl/min
undesired variation was then analyzed by fluid dynamic
simulation. Fluid dynamic simulation of QCM
chamber shows that the sample flow in circular QCM
chamber is turbulent. The results can explain the
observed significant variation in sensing responses of
identical QCM sensors. In the future, QCM chamber
will be redesigned with different geometries to
obtain laminar flow through all sensor electrodes so
that every sensor experiences the same sample dispersion
under a constant flow condition.
5. ACKNOWLEDGMENT
This project is funded by NECTEC’s Sensor Platform
Technology under QCM sensor for food safety
project. The authors would like to acknowledge with
gratitude to Dr. Chamras Promptmas at Mahidol
University for information supporting. Moreover, we
would like to thanks Asia Institute of Technology
(AIT) for the use of simulation software and hardware.
References
[1] J. Rabe, S. Buttgenbach, J. Schroder, and P.
Hauptmann, “Monolithic miniaturized quartz
microbalance array and its application to chemical
sensor systems for liquids,” IEEE Sensors
Journal, Vol. 3, pp. 361 - 368, August 2003.
[2] Masashi Kikuchi, Katsuya Omori, and Seimei
Shiratori, “Quartz Crystal Microbalance (QCM)
Sensor for Ammonia Gas using Clay/ Polyelectrolyte
Layer-by-Layer Self-Assembly Film,”
IEEE Sensors Proceedings 2004, vol. 2, pp. 718-
721, October 2004.
[3] K. Jaruwongrungsee, A. Wisitsoraat, T. Lomas
and A. Tuantranont, “Humidity sensor utilizing
multiwalled carbon nanotubes coated quartz
crystal microbalance,” The 2nd IEEE Nanoelectronics
Conference, pp. 961-964, March 2008.
[4] H. Ito, S. Kakuma, R. Ohba, and K. Noda,
“Development of a humidity sensor using quartz
crystal microbalance,” SICE Annual Conference
2003, pp. 1175-1178, August 2003.
[5] Robert B. Towery, Newton C. Fawcett, and Jeffrey
A. Evans, “Determination of chloroplast
DNA in a cultured soybean line using a QCM
biosensor,” IEEE Sensors Journal, Vol. 4, pp.
489 - 493, August 2004.
[6] G. Sauerbrey, “Use of vibrating quartz for thin
weighing and microweighing,” Z. Phys., vol.155,
p.260, 1959.
[7] K.K. Kanazawa and J.G. Gordon, “The oscillation
frequency of a quartz resonator in contact
with a liquid,” Anal. Chim. Acta., vol. 175, p.99,
1985.
[8] A.Yuwono, and P. Schulze Lammers, “Performance
Test of a Sensor Array - Based Odor
Detection Instrument,” Agricultural Engineering
International: the CIGR Journal of Scientific
Research and Development, May, 2004.
[9] Pengchao Si, John Mortensen, Alexei Komolov,
Jens Denborg, and Preben Juul Moller, “Polymer
coated quartz crystal microbalance sensors
for detection of volatile organic compounds in
gas mixtures,” Analytica Chimica Acta, Vol. 597,
pp. 223-230, 2007.
[10] I.A. Koshets, Z.I. Kazantseva, and Yu.M.
Shirshov, “Polymer films as sensitive coatings
for quartz crystal microbalance sensors array,”
Semiconductor Physics, Quantum Electronics &
Optoelectronics, Vol. 6, pp. 505-507, 2003.
54 K. Jaruwongrungsee et al: Analysis of Quartz Crystal Microbalance Sensor Array with ... (50-54)
[11] A. Palaniappan, Xiaodi Su, Francis and E.H.
Tay, “Quartz sensor array using mesoporous silica
hybrids for gas sensing applications,” Proc.
2005 IEEE Sensors, pp. 1312-1315, 2005.
[12] K. Jaruwongrungsee, T. Maturos, A. Wisitsoraat,
C. Karuwan, T. Lomas, A. Tuantranont,
“Fabrication and Characterization of QCM Gas
Sensor Array on Single Quartz Disc,” The 4th
Asia Pacific Conference on Transducers and Micro/Nano
Technologies (APCOT’08 ), pp. 168,
June 2008.
Kata Jaruwongrungsee received the
B.Eng. and M.Eng. degree in electronics
engineering from King Mongkut’s
Institute of Technology Ladkrabang
(KMITL), Thailand, in 2003 and 2005
respectively. Since 2006, he has
been with the Nanoelectronics and
MEMS Laboratory, National Electronics
and Computer Technology Center
(NECTEC), Thailand, as an assistant
researcher. Presently, he is also doing
his Ph.D. on electronics engineering at KMITL. His research
is mainly focused on QCM based gas sensors and biosensors.
Anurat Wisitsora-at received his
Ph.D., M.S. degrees from Vanderbilt
University, TN, U.S.A., and B. Eng degree
in electrical engineering from Chulalongkorn
University, Bangkok, Thailand
in 2002, 1997, and 1993, respectively.
His research interests include microelectronic
fabrication, semiconductor
devices, electronic and optical thin film
coating, sensors, and micro electromechanical
systems (MEMS).
Manas Sangworasil was born in
Bangkok, Thailand in 1951. He received
the bachelor of engineering and Master
o Engineering from King Mongkut’s
Institute of Technology at Ladkrabang,
Bangkok, Thailand in 1973 and 1977 respectively,
and the D. Eng (Electronics)
from Tokai University, Japan, in
1990. Following his graduate studies, he
worked almost 28 years at Electronic Department,
Faculty of Engineering, King
Mongkut’s Institute of Technology at Ladkrabang, Bangkok
where he is currently an associate professor. His research interest
are in the area of image process with emphasis on Image
Reconstruction, 3D modeling, Image Classification and Image
Filtering.
Thitima Maturos received the B.S.
degree in physics from Thammasat University,
Pathumthani, Thailand in 2002
and the M.S. degree in Physics from
Mahidol University, Bangkok, Thailand
in 2006. Since 2006 she has been
with Nanoelectronics and MEMS Laboratory,
National Electronics and Computer
Technology Center (NECTEC),
Thailand, where she is now assistant researcher.
Her research interests include
microfabrication, cell manipulation and Lab-on-a-chip.
Pornpimol Sritongkum received her
Ph.D. in electroanalytical chemistry
from Cranfield University (United Kingdom)
in 2002. Her current research
is focused on the surface modification
of QCM- and electrochemical biosensors
and on the development of development
of novel diagnostic devices.
Adisorn Tuantranont received the
B.S. degree in electrical engineering
from King Mongkut’s Institute of Technology
Ladkrabang, Thailand, in 1995,
and the M.S. and Ph.D. degrees in
electrical engineering (laser and optics)
from the University of Colorado at
Boulder in 2001. Since 2001, he has
been the director of the Nanoelectronics
and MEMS Laboratory, National Electronic
and Computer Technology Center
(NECTEC), Pathumthani, Thailand. He also serves on
the founding committee of the National Nanotechnology Center
(NANOTEC), the first nanotechnology center in Thailand,
under the Ministry of Sciences and Technology. His research
interests are in the area of microelectro-mechanical systems
(MEMS), optical communication, laser physics, microfabrication,
electro-optics, optoelectronics packaging, nanoelectronics,
and lab-on-a-chip technology. He has authored more than 60
international papers and journals and holds five patents. Dr.
Tuantranont received the Young Technologist Award in 2004
from the Foundation for the Promotion of Science and Technology
under the patronage of H. M. the King.
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 55
Muscular-Contraction Classification :
Comparison Study Between Independent
Component Analysis and Artificial Neural
Network
D. Sueaseenak 1∗ , T. Chanwimalueang 2 ,
W. Iampa 1 , Guest members, and M. Sangworasil 1 , Member
ABSTRACT
We developed a multi-channel electromyogram acquisition
system using PSOC microcontroller to acquire
multi-channel EMG signals. An array of 4 x 4
surface electrodes was used to record the EMG signal.
The obtained signals were classified by a backpropagation-type
artificial neural network. B-spline
interpolation technique has been utilized to map the
EMG signal on the muscle surface. The topological
mapping of the EMG is then analyzed to classify
the pattern of muscle contraction using independent
component analysis. The proposed system was successfully
demonstrated to record EMG data and its
surface mapping. The comparison study of muscular
contraction classification using independent component
analysis and artificial neural network demonstrates
shows that performance of ANN classification
is as comparable as that of the ICA. The computational
time of ANN is also less than that of the ICA.
Keywords: Electromyography, Principal Component
Analysis, Independent Component Analysis, Artificial
Neural Network
1. INTRODUCTION
Electromyography (EMG) is the study of muscle
electrical signals. EMG is sometimes referred to as
myoelectric activity. Many muscular abnormalities
such as muscular dystrophy, inflammation of muscle,
peripheral nerve damages could results in an
abnormal electromyogram [1-6]. Figure 1 shows a
schematic representation of muscle units and its components.
EMG can be recorded by two types of electrodes;
invasive electrode the so-called wire or needle
electrodes and non-invasive electrode the so-called
* Corresponding author.
Manuscript received on December 25, 2009.,
1 D. Sueaseenak, W. Iampa and M. Sangworasil, Department
of Electronics, Faculty of Engineering, King Mongkut’s
Institute of Technology Ladkrabang, Thailand
2 T. Chanwimalueang, Biomedical Medical Engineering programme,
Faculty of Engineering, Srinakharinwirot University,
Thailand
E-mail addresses: emg7849@gmail.com (D. Sueaseenak)
Fig.1: Muscle unit [7]
surface electrode. Wire or needle electrodes records
individual muscle fiber action potentials which is an
ideal choice to evaluate the muscle activity [10]. However,
fine wire intramuscular electrodes require a needle
for insertion into the muscle and may cause a significant
pain. The choice of surface electrode is then
preferable. However, when EMG is acquired from
surface electrodes mounted directly on the skin, the
signal is a composite of all the muscle fiber action potentials
occurring in the muscles underlying the skin.
Estimating this force in general is a hard problem
due to difficulties in activating a single muscle in isolation,
isolating the signal generated by a muscle from
that of its neighbors, and other associated problems
[8-9]. The clinical application of EMG can be classified
into two mains categories. (i) Standard EMG [9]
is recorded from discrete sites on a muscle and thus
provides only a limited picture of the actual muscular
electrical activity in the vicinity of the recording electrode.
(ii) Array EMG recorded by an array electrode
which facilitates the clinical interpretation of electrical
activities through the mapping of these signals on
the muscle surface [10]. In this paper, a PSOC-based
multi-channel surface electrode array data acquisition
system is developed to acquire EMG data. The
EMG signals are then mapped using B-spline interpolation
technique. The EMG topological Mapping
is then used for classification of muscular contraction.
There exist a number of 2D-pattern classifications
56 D. Sueaseenak et al: Muscular-Contraction Classification : Comparison Study ... (55-62)
Fig.2:
Muscular contraction classification system
[10-13]. In this research, we compared the EMGcontraction
classification technique between independent
component analysis (ICA) and artifial neural
network (ANN) [15]. The results show that the ANN
classification yields as good performance as the ICA
in the faster computational time.
The paper is organized as follows: Section 2 is devoted
to the designed concept of multi-channel electromyogram
system. Section 3 briefly reviews fourierbased
features, Section 4 describes feature extraction
and topological mapping process. Section 5 briefly
introduces Independent Component Analysis. Section
6 briefly introduces Artificial Neural Network.
The experiment and results are shown in Section 7.
Discussion and Conclusion are provided in Section 8.
Fig.3:
system
Multi-channel electromyogram acquisition
2. DESIGN AND CONSTRUCTION OF
MULTI-CHANNEL EMG
EMG measurement is accomplished by the instrument
called electromyograph. The system, in general,
consists of instrumentation amplifier, notch filter, offset
adjustment, isolator, main amplification, and the
CRT display. The instrument amplifier is a front-end,
high CMRR differential amplifier which functions to
pick-up a low amplitude signal submersed in the highfrequency
noise. The notch filter gets rid of the 50Hz
noise while keeping the EMG signal intact. The offset
adjustment maintains the baseline level especially
during the subject’s motion. The function of isolator
is to separate the front-end section from the rear-end
section to protect the possible electrical shock to the
patient. The main amplification conditions the EMG
prior to be display with CRT. The complexity of the
electronic circuit becomes realized with the necessity
to monitor the multi-channel of EMG. Such complicate
designs, however, are made possible by the creation
of entirely reconfiguration and programmable
components the so-called Programmable System on
Chip Microcontroller (PSOCMicrocontroller).
The designed EMG system is capable of monitoring
16 channels of EMG simultaneously. Each channel
consists of 2 main parts; (i) EMG signal process-
Fig.4:
Raw EMG signal on laptop computer
ing unit and PSOC microcontroller. Figure 2 shows
the muscular contraction classification system Figure
3 shows Multi-channel electromyogram acquisition
system.
The EMG signal processing units consists of 3 subunits
(i) Instrumentation Amplifier. This subunit uses
the INA 2128 BUR-BROWN Integrated Circuit. The
IC can achieve a CMRR up to 120 dB and gain up to
1000.
(ii) Noise filter. The function of the filter is to
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 57
get rid of the 10-20 Hz noise which is classified as a 4. EMG FEATURE EXTRACTION AND
motion artifact.
MAPPING
(iii) Amplifier and Offset Adjustment. The objective
of this sub-unit is to Amplifier EMG signal and
maintains the appropriate offset voltage prior to interface
with the PSOC.
The PSOC microcontroller consists of 4 subunits
(i) PGA (Programmable Gain Amplification) This
subunit acts as the buffer and the main amplification
of EMG.
(ii) Low pass filter. The function of the filter is
to remove of the high frequency noise. The cut-off
frequency is at 500 Hz.
Fig.6: Feature extraction and mapping
(iii) DELTA-SIGMA. This subunit functions as a
8-bit analog to digital converter.
Figure 6 shows the feature extraction and mapping
process. Each of the 16 EMG channels will be
(iv) UART. This subunit functions to perform RS-
232 interfacing unit with PC
converted to frequency domain by taking the fourier
transform. The energy content of the EMG signal
3. CLASSICAL FOURIER-BASED FEATURESis then evaluated by computing area under the magnitude
squared of the fourier transform. The energy
Fourier transform is a useful technique that represents
a stationary signal in terms of a function of content on the 4x4 grid corresponding to the 4x4 electrode
shown in figure 7 is used for artificial neural
frequency by determining the frequency component.
The equation is represented as:
network classification. The 4x4 grid data was interpolated
to derive the 49x49 topological maps which
N−1
∑
are later applied to ICA for muscular contraction classification.
Figure 8 shows the topographical mapping
X (f) = x (n) e (−j2πfn/N) (1)
of various muscular contractions.
n=0
Where x(n) is a sequence of EMG input and X (f)
is complex vector provided by the sinusoidal coefficients.
The classical Fourier transform was applied
to the EMG signal. Then, the power spectral density
of the energy at various frequencies was computed
from square magnitude of the Fourier transform as
represented in Equation 2. Figure 5 shows the energy
content of EMG signal in form of the power spectrum.
P S (f) = |X (f)| 2 (2)
Fig.7:
16 Channel electrode placements
Fig.5:
EMG Power spectrum
5. FROM PRINCIPAL COMPONENT ANAL-
YSIS TO INDEPENDENT COMPONENT
ANALYSIS
Principal Component Analysis (PCA) is a statistical
technique which used to describe a large dimensional
space with a relative small set of vectors. It
is a popular technique for finding patterns in data of
high dimension, and is used commonly in both face
recognition and image compression. [13] Application
of PCA to face recognition is known as Eigen face.
The Eigen face technique is a powerful yet simple solution
to the face recognition dilemma. It uses much
more information by classifying faces based on general
facial patterns. Here we focus on the application
of PCA for muscular-contraction classification
The procedure for using PCA is divided into 2
steps. (i) Training step and (ii) Classification step.
58 D. Sueaseenak et al: Muscular-Contraction Classification : Comparison Study ... (55-62)
C = 1 M ΦΦT (5)
(iv) Compute the Eigenvalue and Eigenvector of C
represented as
Cν i = µ i ν i (6)
where µ i is the corresponding Eigenvalue of Eigen
vector ν i
(v) Project each training set on the Eigenspace
using the operation
Ω = V · Φ (7)
Where V is the Eigen matrix where each row is
the Eigenvector ν i can be written as
Ω = [ω 1 , ω 2 , . . . , ω M ] (8)
Where ω i is the coefficient of the training map i th
The Classification step is as follow:
Project vector form of the tested topological mapping
matrix T p to the Eigenspace using equation (5)
to derive ω s as
ω s = V · [T n − ψ]
The tested topological mapping matrix is classified
to class k which minimized
ε 2 k = ‖ω s − ω k ‖ 2 (9)
with 1 ≤ k ≤ M
The goal of independent component analysis (ICA)
is to minimize the statistical dependence between the
basic vectors. Mathematically, we can write
W X = U (10)
Fig.8:
Topological mapping
The Training step is as follow:
(i) Convert each cropped topological mapping matrix
into a vector T i of length N (N= map width*map
height). For the M data set, we let the training set
represented by, {T 1 , T 2 , T 3 , ..., T M } where M is the
vector of N 2
(ii) Compute the mean vector Ψ and the set of
deviation from the mean vector Φ = [Φ i , Φ 2 , ...Φ M ]
which is defined as
Ψ = 1 M
M∑
T i (3)
i=1
Φ i = T i − Ψ (4)
(iii) Compute the covariance matrix C which is
defined as
ICA searches for a linear transformation W that
minimizes the statistical dependence between each
row of U. There exists a number of iterative algorithm
to solve for W [16,17]. Most of them are optimized
for the dependence criteria including Kurtosis, Negentropy,
etc.[18]. In this paper, we applied the well
known ICA algorithm the so-called InfoMax purposed
by Bell and Sejnowski [19]. The idea of InfoMax has
been applied to Eigenvector of PCA by Barlett et. al.
[20] by minimize the statistical dependence between
each row of U in
W V = U (11)
where V as an Eigen Basis matrix where each row
is the Eigen vector ν i defined in (4). The new basis
W −1 U is then used in place of V. The Projection
of each training set on the new basis -space is hence
defined as
ω s = ( W −1 U ) · [T n − ψ] (12)
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 59
6. ARTIFICIAL NEURAL NETWORK
Two-layered artificial neural network using backpropagation
training protocol was used as a classifier.
The 4x4 EMG grid data served as the input of the
neural network. The eight outputs of neural network
that correspond to the eight classes of the muscular
contraction. Structure of artificial neural network is
shown in figure 9.
Table 1: ACCURACY OF ICA AND ANN CLAS-
SIFICATION
on the muscle surface. The mapping for various pattern
of muscular contraction were then recorded and
later analyzed with independent component analysis
to classify the pattern of muscular contraction pattern.
The comparison study of classification result
demonstrates that ANN provides the comparable performance
as the ICA. Yet the ANN computational
time is noticeably less than that of the PCA.
Artificial neural network used in the classifi-
Fig.9:
cation
9. ACKNOWLEDGMENT
The authors wish to thanks ASEAN University
Network/Southeast Asia Engineering Education Development
Network (AUN/SEED-Net) to support
scholarship and The DEMAMEDICAL CO., LTD to
support the ECG/EMG surface electrode and lead
wire to measure EMG signal in this research.
7. EXPERIMENT AND RESULTS
The 15 patterns of topographical mapping of eight
muscular contraction of forearm (120 maps) were
used in the training process of ICA and used 4x4
grid (120 data) for training process of ANN. The topographical
mapping of the 15 unknown contractions
was then used as the ICA tested set. The 15 unknown
4x4 grid data was then used as ANN testd set. Figure
8 shows the ICA training sets, the derived ICA basis
and the result of ICA classification. Table 1 shows
the accuracy of ICA and ANN classification.
8. DISCUSSION AND CONCLUSION
A multi-channel electromyogram acquisition system
using PSOC microcontroller was designed and
constructed to aquire multi-channel EMG signals.The
16 EMG channels was 4x4 grid data for classification
by using artificial neural network. The 4x4 grid data
was performed to a topological map of EMG signal
References
[1] G. A. Bekey, C Chang, J Perry, M.M. Hoffer,“Pattern
recognition of multiple EMG signals
applied to the description of human gait,” Proceedings
of IEEE, vol. 65, pp. 674-689, 1977.
[2] S. Boisset, F Goubel, “Integrated electromygraphy
activity and muscle work,” J Applied Physiol,
vol 35, pp. 695-702, 1972.
[3] C.J. DeLuca, “Use of the surface EMG signal for
performance evaluation of back muscle,” Muscle
& Nerve, vol. 16, pp. 210-216, 1993.
[4] R. Plonsey, “The active fiber in a volume conductor,”
IEEE Trans Biomed Eng, vol. 21, pp.
371-381, 1974.
[5] D.A. Winter, “Pathologic gait diagnosis with
computer averaged electromyographic profiles,”
Arch Phys Med Rehabil, vol. 65, pp. 393-398,
1984.
[6] K. Lyons, J Perry, J.K. Gronley, L Bbarnes, D.
Antonelli, “Timing and relative intensity of hip
extensor and abductor muscle action during level
60 D. Sueaseenak et al: Muscular-Contraction Classification : Comparison Study ... (55-62)
(a) Hand close
(b) Hand open
(a)
(c) Wrist extension
(d) Wrist flexion
(e) Wrist extension
(f) Wrist flexion
(g) Radial flexion
(h) Ulnar flexion
(b)
Fig.11:
Classification results
Fig.10: (a) Training Topological Mapping Input of
ICA; (b) ICA Basis
and stair ambulation: An EMG study,” Physical
Therapy, vol. 63, pp. 1597-1605, 1983.
[7] Elaine N.Marieb., Human anatomy and physiology,
Sixth Edition, The Pearson Education; pp.
296,2004
[8] K.S. Turker, T.S Miles, “Cross talk from other
muscles can contaminate electromyographic signals
in reflex studies of the human leg,” Neuroscience
Letters, vol. 111, pp. 164-169, 1990.
[9] J.W. Morrenhof, H.J. Abbink, “Crosscorrelation
and cross talk in surface electromyography,”
Electromyography, vol. 25, pp.
73-79, 1985.
[10] Basmajian JV, de Luca CJ. Muscles Alive - The
Functions Revealed by Electromyography. The
Williams & Wilkins Company; Baltimore, 1985
[11] J. Cartinhour, “ A Bayes classifier when the class
distributions come from a common multivariate
normal distribution ” ,IEEE Transactions on Reliability,
vol. 41, Issue 2, pp. 124 - 126, 1992.
[12] N.B. Karayiannis, M.M. Randolph-Gips, “ Soft
learning vector quantization and clustering algorithms
based on non-Euclidean norms: singlenorm
algorithms ” ,IEEE Transactions on Neural
Networks, vol. 16, Issue 2, pp. 423 - 435, 2005.
[13] E. Alpaydin, M.I. Jordan, “ Local linear perceptrons
for classification ” ,IEEE Transactions on
Neural Networks, vol. 7, Issue 3, pp. 788 - 794,
1996.
[14] Zhujie, Y.L. Yu, “ Face recognition with Eigen
faces ” ,Proceedings of the IEEE International
INTERNATIONAL JOURNAL OF APPLIED BIOMEDICAL ENGINEERING VOL.2, NO.2 2009 61
Conference on Industrial Technology, pp. 434 -
438, 1994.
[15] Comon, P., Independent component analysis; A
new concept? Signal Processing, vol 36, no. 3,
pp. 287-314,1994
[16] Cardoso, J.-F., Infomax and Maximum Likelihood
for Source Separation. IEEE Letters on
Signal Processing, vol. 4. pp. 112-114, 1997
[17] Hyvärinen, A., The Fixed-point Algorithm and
Maximum Likelihood Estimation for Independent
Component Analysis. Neural Processing
Letters, vol. 10: pp. 1-5, 1999.
[18] Hyvärinen, A. and E. Oja, Independent Component
Analysis: Algorithms and Applications.
Neural Networks, 2000. vol. 13, no. 4-5, pp. 411-
430, 2000.
[19] Bell, A.J. and T.J. Sejnowski, An informationmaximization
Approach to Blind Separation and
Blind Deconvolution. Neural Computation, vol.
7, no. 6, pp. 1129-1159, 1995
[20] Bartlett, M.S., H.M. Lades, and T.J. Sejnowski.
Independent component representations for face
recognition. in SPIE Symposium on Electronic
Imaging: Science and Technology; Conference
on Human Vision and Electronic Imaging III,
San Jose, CA, 1998
W. Iampa received the B.Sc. (Radiological
Technology) and M.Sc. (Radiological
Technology) from Mahidol
University, Thailand in 2005 and
2008, respectively. She is currently
a Ph.D. student in Department of
Electronics, Faculty of Engineering,
KMITL.
M. Sangworasil was born in Bangkok,
Thailand in 1951. He received the Bachelor
of engineering and Master o Engineering
from King Mongkut’s Institute
of Technology at Ladkrabang, Bangkok,
Thailand in 1973 and 1977 respectively,
and the D. Eng (Electronics) from Tokai
University, Japan, in 1990. Following
his graduate studies, he worked almost
28 years at Electronic De- partment,
Faculty of Engineering, King Mongkut’s
Institute of Technology at Ladkrabang, Bangkok where he is
currently an associate professor. His research interest are in the
area of image process with emphasis on Imag Reconstruction,
3D modeling, Image Classication and Image Filtering.
D. Sueaseenak received the B.Eng.
(Electrical Engineering) from Srinakharinwirot
University, Thailand, in 2005,
and the M.Eng. (Biomedical Electronics)
from King Mongkut’s Institute of
Technology Ladkrabang, Thailand, in
2007. He is currently a Ph.D. student
in Department of Electronics, Faculty of
Engineering, KMITL.
T. Chanwimalueang received the
B.Eng. (Electrical Engineering) from
Khon Kaen University, Thailand, in
2000, and the M.Eng. (Biomedical
Electronics) from King Mongkut’s Institute
of Technology Ladkrabang, Thailand,
in 2007. He is currently a Lecturer
in Biomedical Engineering programme,
Faculty of Engineering, Srinakharinwirot
University.
IJABME Manuscript Submission Guide
General
The International Journal on Applied Biomedical Engineering (IJABME) is published bi-annually by ThaiBME
Society. Contributed papers must be original that advance the state-of-the art and applications of Biomedical
Engineering. Both theoretical contributions (including new techniques, concepts, and analyses) and practical
contributions (including system experiments and prototypes, and new applications) are encouraged. The submitted
manuscript must have been copyrighted, published, or submitted or accepted for publication elsewhere, except in a
conference proceedings. The manuscript text should not contain any commercial references, such as company
names, university names, trademarks, commercial acronyms, or part numbers. Not accepted material will not be
returned.
Submission Due
Authors should document their work in relation to the open literature. The following limits on length (A4 size) should
be observed.
- End of March, for January-June issue.
- End of September, for July-December issue.
Length
Authors should document their work in relation to the open literature. The following limits on length (A4 size) should
be observed.
1) Regular papers, should be at least 6 pages and up to 12 pages in length, including figures and illustrations.
2) Correspondences, less than 3 pages in length, including figures and illustrations.
Styles for Manuscript
1) The manuscript must be in English.
2) Provide a carefully worded abstract of 150 to 300 words for regular papers and less than 150 words for
correspondences. If the manuscript is not submitted via the website, Name, address, and telephone number
of author(s) should be organised on a separate cover page.
3) The style for organization of a paper can be found in the manuscript template file.
Form and Submissions
1) The manuscripts itself
2) Cover Page: should include
- Paper title
- Paper category (regular paper or correspondence)
- Full name of all authors
- Name and address of the key author whom proofs and other correspondence can be sent
- Keywords (duplicate from the paper)
- Abstract (duplicate from the paper)
- Date of submission
Notes:
1. Authors are encouraged to submit their paper in electronic form via the IJABME website. The first sent
manuscripts must be in the form of pdf files. In any case of difficulty, the manuscript could be also attached to
standard emails and sent directly to the editor.
2. Authors who are not convenient to submit paper in electronic form, hard copy formatted is also appreciated.
Authors are to send four copies of the manuscript.
3. The accepted and/or revised manuscript must be resubmitted for IJABME publishing. Text files are
appreciated. The manuscript will be reformatted upon IJABME style.
4. Original illustrations (could be also good resolution image files) must be ready for immediately submission upon
acceptance of the manuscript. In the case of regular papers, also be prepared to provide a brief technical
biography and photograph of each author.
Where to Submit
Electronic version of the manuscript must be submitted online via IJABME website; http://www.ijabme.org.
Review Process
The review process usually takes 3 weeks to 3 months. The author is then notified the decision of the editor based
on the reviewers' recommendations. Authors may be asked to revise the manuscript if it is not accepted or rejected
in its original form. The un-accepted manuscript will not be returned to authors.
Page Charges
After the manuscript has been reviewed and accepted for its publication, the author's company or institution will be
asked to pay a charge of 1,000 THB (or 33 USD) per printed page to cover part of the cost of publication. ThaiBME
members will receive full reduction on page charge. Printing of colour pages are specially requested and fully paid
by the authors. Page charges can be exempted upon request.
Copyright
It is the policy of the ThaiBME to own the copyright to the published contributions on behalf of the interests of
ThaiBME, its authors, and their employers, and to facilitate the appropriate reuse of this material by others. To
comply with the Copyright Law, authors are required to sign a copyright transfer form before publication. This form,
a copy of which appears in this journal, returns to authors and their employers full rights to reuse their material for
their own purposes.
International Steering Committee
Chusak Limsakul (Chair), Thailand
Adhi Susanto, Indonesia
Chuchart Pintavirooj, Thailand
Ferdinand F.S. Cohen, USA
Ian Thomas, Thailand
James Koh, Singapore
Kazuhiko Hamamoto, Japan
Kosin Chamnognthai, Thailand
Ratko Magijarevic, Croatia
Eung Je Woo, Korea
Supaporn Kiattisin, Thailand
Tru Cao, Vietnam
Tsuyoshi Shiina, Japan
Somsak Choomchuay, (Sec.)
Thailand
Organizing Committee
General Chair
Tsuyoshi Shiina
Kyoto University, Japan
General Co-Chairs
Tohru Yagi
Tokyo Institute of Technology, Japan
Keiji Iramina
Kyushu University, Japan
Technical Program Chairs
Yasuhiko Jimbo
University of Tokyo, Japan
Kohji Masuda
Tokyo Univ. of Agri. and Tech., Japan
Publicity Chairs
Suparerk Janjarasjitt
Ubonrathathani Univ., Thailand
Adisorn Leelasantithum
UTCC, Thailand
Special Session Chair
Supan Tungjitkusolmun
KMITL, Thailand
Information Chair
Surapan Airphaiboon
KMITL, Thailand
Local Arrangement Chair
Makoto Yamakawa
Kyoto University, Japan
General Secretary
Kazuhiko Hamamoto
Tokai University, Japan
Supported by
• International Federation for Medical
and Biological Engineering (IFMBE)
• Technical Committee of Medical and
Biological Engineering, Society of
Electronics, Information and
Systems, The Institute of Electrical
Engineers of Japan (IEEJ)
• Thai Biomedical Engineering Society
(ThaiBME)
• International Journal of Applied
Biomedical Engineering (IJABME)
Call for Paper
The 3 rd BMEiCON is still intended to provide an international forum where
researchers, practitioners, and professionals interested in the advances in, and
applications of, biomedical engineering can exchange the latest research, results,
and ideas in these areas through presentation and discussion.
The BMEiCON2010 will be held in Kyoto, Japan, during August 27-28, 2010. The
organizing committee is pleased to invite all engineers, physicians, scientists,
technicians, and technologists to attend and help to shape the future of biomedical
technology. The topics for regular sessions include, but are not limited to, the
followings:
• Bioinformatics, Medical Informatics
• Biomechanics and Biomaterials
• Computer-Integrated and Assisted Surgery
• Medical Imaging, Image and Signal Processing
• Medical Simulation, Systems and Control
• Medical Expert Systems
• Molecular Bioengineering
• Rehabilitation and Clinical Engineering
• Tele-Health
• Biomedical Engineering Education
• Cell and Tissue Engineering
• Health Care Technology
• Medical Measurement and Instrumentation
• Medical Robotics and Automation
• Medical Navigations
• Physiological System Modeling
• Virtual and Augmented Reality in Medicine
• Other Related Topics
Submission:
• Authors are invited to submit a full paper according to the posted guidelines.
• Authors are expected to present their papers at the conference upon acceptance
and presenting authors are required to register for the conference.
• Papers should neither have been published elsewhere nor currently under review
by another conference or journal.
• All papers for BMEiCON2010 must be submitted electronically via the conference
website only.
http://www.bmeicon.org/bmeicon2010
Author’s Schedule:
Special session proposal due: April 15, 2010
Submission of Full Paper: April 30, 2010
Notification of Acceptance: June 15, 2010
Submission of Camera Ready Paper: July 16, 2010
Special session proposal; Submit to the special session chair at
ktsupan@kmitl.ac.th
General Secretary could be contacted by:
FAX: +81-463-50-2412 or Email: bmeicon2010@dm.u-tokai.ac.jp
Check the conference website for up-to-date information:
http://www.bmeicon.org/bmeicon2010
![[PDF mobi ePub] Photoshop Down & Dirty Tricks for Designers, Volume 2 Full PDF](https://documents.yumpu.com/000/064/341/729/2f539548af79705bca130c01eca1498779cd3e24/456e49706c31474658426942414d394f366c6b7571513d3d/39577a43364c3074334f4e4a51517458384f586f5a413d3d.jpg?AWSAccessKeyId=AKIAICNEWSPSEKTJ5M3Q&Expires=1611223200&Signature=G4X4oxouqt4Z6KpNSnrReJ0Y2uA%3D)
Change language