The Journal of International Medical Research

2009; 37: 779 – 790

Infrared Thermography Based on

Artificial Intelligence as a Screening
Method for Carpal Tunnel Syndrome
Diagnosis
B JESENŠEK PAPEŽ1, M PALFY1, M MERTIK2 AND Z TURK1
1
Department of Physical Medicine and Rehabilitation, Medical Centre Maribor, Maribor,
Slovenia; 2Faculty of Information Studies, Novo Mesto, Slovenia

This study further evaluated a computer-
based infrared thermography (IRT) system,
which employs artificial neural networks
for the diagnosis of carpal tunnel syndrome
(CTS) using a large database of 502 thermal
images of the dorsal and palmar side of 132
healthy and 119 pathological hands. It
confirmed the hypothesis that the dorsal
side of the hand is of greater importance
than the palmar side when diagnosing CTS
thermographically. Using this method it
was possible correctly to classify 72.2% of
all hands (healthy and pathological) based
on dorsal images and > 80% of hands when
only severely affected and healthy hands
were considered. Compared with the gold
standard electromyographic diagnosis of
CTS, IRT cannot be recommended as an
adequate diagnostic tool when exact
severity level diagnosis is required, however
we conclude that IRT could be used as a
screening tool for severe cases in
populations with high ergonomic risk
factors of CTS.

KEY WORDS: CARPAL TUNNEL SYNDROME; DIAGNOSIS; INFRARED THERMOGRAPHY; SKIN TEMPERATURE;
ARTIFICIAL INTELLIGENCE

Introduction and degree of compression of the median

Carpal tunnel syndrome (CTS) is attributed
to compression of the median nerve inside
the carpal tunnel. Tendinitis of the flexor
tendons is thought to be one of the most
common causes of median nerve
compression due to increased pressure in the
carpal tunnel caused by oedema.1 CTS is the
most frequently encountered peripheral
entrapment neuropathy, and is associated
with major costs for medical treatment and
workers’ compensation.2 The clinical
symptoms of CTS depend on the duration
nerve. The first symptoms affect the sensory
nerve fibres but, as the compression persists,
motor nerve fibres undergo damage as well.3
Clinical signs and symptoms alone are not
sufficient to confirm the diagnosis and
electrodiagnostic tests, such as
electromyography (EMG), are needed, but
they are expensive and cause discomfort.
Infrared thermography (IRT) is an
excellent method of examination, and is
useful because if is safe, painless, non-
invasive, easy to reproduce and has low

779
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography
running costs.4 It measures heat emitted
from a surface, according to black body
radiation law. Herrick and Herrick5
maintained that thermal patterns of CTS
showed a decreased vascular heat emission
pattern over the median nerve distribution.
Tchou et al.6 demonstrated the high
sensitivity and specificity (ranging between
98% and 100%) of IRT in the diagnosis of
unilateral CTS. Despite many years of
searching for a practical use for IRT as a
diagnostic tool for peripheral nerve injuries,
including CTS, it has not become an
established tool and is considered to be fairly
unreliable.7
During our preparations for the study we
found that the manual analysis of thermal
images can be a tedious job, where accuracy
is of extreme importance as small errors in
defining the areas of interest can lead to
significant differences in the measured
temperatures. This led us to look for a better,
more accurate method which resulted in our
software module being developed, which
consists of two modules: the first module
performs image segmentation and extracts
temperature readings, while the second
module performs image analysis and
attempts to diagnose CTS.8 Diagnosis relies
on previously acquired knowledge from
other images stored in an artificial neural
network (ANN). The results appeared to be
very reliable: classification success rates were
usually > 80% when focusing on the dorsal
side of the hand,8 which suggests that this
artificial intelligence-based IRT technology
could be used to diagnose CTS. However,
only 112 images were available (with 56
dorsal views),8 which was not very many
bearing in mind the important role that the
learning process plays in ANN
development.9 Since then, the image
database has grown considerably and the
present study reports the most recent results.
780
Patients and methods
PATIENTS AND STUDY DESIGN
Patients were selected from a pool of patients
referred by general practitioners and various
specialists in Maribor (Slovenia) and the
surrounding areas to the Department of
Physical Medicine and Rehabilitation,
Medical Centre Maribor, from March 2007 to
October 2008. The exclusion criteria were:
previous operation for CTS, negative
electrodiagnostic test for CTS, and abnormal
hand anatomy (amputations, injuries, other
anomalies) that would prevent the
acquisition of standard images of the hands.
Volunteers who did not exhibit symptoms of
CTS and who were subsequently confirmed
by EMG as not having CST were used to
acquire images of healthy hands (dorsal and
palmar views).
All patients were informed about the trial.
The method used was non-invasive and
none of the patients had any objections to
their voluntary participation. The study
protocol was approved by The National
Medical Ethics Committee of the Republic of
Slovenia.
The diagnosis of CTS was based on
symptomatology (paraesthesias and pain
mainly at night) and clinical examination
(Phalen test, Tinel test, sensibility and motor
evaluation), and was confirmed by nerve
conduction studies (NCS).10 – 12 On the basis
of NCS carried out according to the criteria of
Stevens,13 the severity of CTS was divided
into three groups: (i) mild CTS, prolonged
median sensory latency and normal distal
motor latency; (ii) moderate CTS, prolonged
median sensory latency and prolonged distal
motor latency; and (iii) severe CTS,
prolonged median sensory latency and
prolonged distal motor latency with either
absent sensory nerve action potentials or low
amplitude or absent thenar compound
muscle action potentials.
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

IMAGE ACQUISITION AND capability to extract similarities from

SOFTWARE
As reported previously,8 all images were
taken with a Neo Thermo TVS-700 camera
(NEC Avio Infrared Technologies, Tokyo,
Japan) with a resolution of 320 × 240 pixels
from a standard distance of 80 cm under
similar ambient conditions (room
temperature 23 ± 2 °C; relative humidity 56 ±
8%). The thermography guidelines issued by
the International Academy of Clinical
Thermography were taken into
consideration.7 Patients were required to
position their hands as parallel as possible to
the camera lens and to spread their fingers
wide apart (Fig. 1). A software application
that we developed specifically for image
segmentation and extraction of temperature
readings from acquired images was used. By
utilizing an ANN, the extracted data were
then used in an attempt to diagnose CTS.
Although there are many data mining
techniques suitable for this kind of pattern
recognition, ANNs were chosen due to
previous positive experience and their
seemingly unrelated data.14 The software
application modules as well as the ANNs are
described in detail below.
IMAGE SEGMENTATION AND
EXTRACTION OF DATA
The Neo Thermo TVS-700 camera is capable
of recording pixel temperatures to an
accuracy of 0.04 – 0.08 °C, depending on the
image acquisition conditions such as distance
to target, its emissivity, ambient temperature,
relative humidity and closeness to other
potentially irradiating bodies. After each
pixel was read from a proprietary image file
format and its temperature calculated using
formulae provided with the camera, the
image was displayed on a screen and a
simple, yet efficient, edge-detection algorithm
was used to trace an outline of the depicted
hand. To ensure the accuracy of edge
detection, a sufficient contrast (temperature
difference) between the hand and its image
background needed to be created before the
image was taken. For this purpose, a cooled

FIGURE 1: Typical thermal image of the dorsal side of a patient’s hand

781
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

towel was placed underneath each hand, CLASSIFICATION USING AANs

making sure that the hand never came into
contact with it. The segmentation process
took place once the outline of the hand had
been extracted from the image background.
Since the goal was to measure differences in
the temperature of the different skin
segments and, taking into consideration the
conclusions of previous studies,5,15,16 the
hand was divided into 12 areas of interest:
fingers (five segments), metacarpus (five
segments) and carpus with wrist (two
segments) (Fig. 2). This was done by
calculating the hand’s centre of mass and
determining characteristic points at the
hand’s edge (e.g. at the fingertips and
between the fingers). The mean temperature
of each segment was calculated and, together
with the ambient temperature, stored in an
output file.
The increasing volume of data in modern
science calls for computer-based approaches
in order to extract useful information. Data,
numerical and otherwise, must be analysed
and processed to convert them into
information that informs, instructs, answers,
or otherwise aids understanding and
decision-making. Data mining techniques
provide a variety of different automated
sophisticated tools for data analysis and
there is now an explosive growth of methods
in the data mining field.17,18 ANNs are
amongst the best known and established
machine-learning tools with their origins
dating back to 1943, when McCulloch and
Pitts19 introduced a mathematical model of a
neural cell – an artificial neuron (Fig. 3).
Since then, and especially in the two decades
since Rumelhart et al.20 introduced a new

FIGURE 2: Typical thermal image of the dorsal side of a patient’s hand after
segmentation into 12 areas of interest: fingers (five segments), metacarpus (five
segments) and carpus with wrist (two segments)

782
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

A0 = 1

W0

A1

W1
I
N
P
W2 Sum Output
U
T
A2
Σ
S Input Activation
function function

Wn

An

FIGURE 3: Mathematical model of an artificial neuron (A0, A1, etc. represent attributes
[inputs]; W0, W1, etc. represent the weight of each input, which changes with the
learning process)19

weight-adaptation algorithm for multilayer interconnections. When such a network is

neural networks in 1986, ANNs have been
increasingly used in problem domains
involving classification and decision-
making; with multilayer perceptrons (also
known as ‘feed-forward back-propagation’
neural networks) being the most widely used
topological variation.21
We decided to implement our own version
of a multilayer perceptron, designed
especially for the purpose of classifying CTS
cases and coded in the C# programming
language, giving us full control over the
machine-learning process. A detailed
description of multilayer perceptrons is
beyond the scope of this article and is well
documented in the literature.17,21,22 In short,
an ANN is an abstract computational model
of the human brain. Like the brain, an ANN
is composed of (artificial) neurons and
viewed in an illustrative format, neurons can
be represented as nodes and
interconnections as edges (Fig. 4). For a
better understanding, it should be pointed
out that an ANN can be trained to adapt its
parameters (weights) through a machine-
learning process (artificial intelligence) and
it can, consequently, learn to classify input
objects correctly. In the present study, the
input object was a hand, described by its
numerical features (i.e. mean temperature of
different segments) and the classification
result was a simple decision of whether or
not the hand was affected by CTS. Another
variation of an ANN was also used, where
the classification result was more complex,
giving the CTS severity level according to
NCS (ranging from 0 [normal] to 3 [severe]).
To test the reliability of an ANN, the data

783
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

Inputs Hidden layer 1 Hidden layer 2 Output layer

In1

In2 Out1

Out2

Inn

FIGURE 4: Illustrative view of an example of an artificial neural network (multilayer

perceptron)

to be classified are usually divided into two
sets. The first set consists of cases used for the
learning process through which the ANN
adapts itself. The second set consists of cases
previously unknown to the trained ANN.
These cases are then classified and the results
compared to actual class values. In theory, if
the training set is large and diverse enough,
an ANN can accumulate enough knowledge
to classify unknown objects reliably. In the
present study, the training set was
constructed by computer random selection
and represented 80% of the whole set, with
the remaining 20% used for testing the
classification reliability. Before training
started, the temperature of each individual
segment was normalized according to the
mean temperature of the whole hand. This
ensured that differences in the temperatures
of hands between different people did not
influence the classification outcome. Such
temperature differences between individuals
are common and usually unrelated to any
pathology. The remaining cases were then
classified. Different classification attempts
were carried out by altering the inclusion
criteria for the input set (data from all
images, only from dorsal images, only from
palmar images, and data from dorsal
images of healthy and severe cases only).
Each time a fully connected multilayer
perceptron with two hidden layers (nine
neurons in the first hidden layer, five
neurons in the second hidden layer) was
used and 100 training repetitions were
carried out (each case from the training set
was used 100 times). ANN topology and
training process parameters were chosen
according to the results of our previous
work.8,14 Each classification was repeated five
times, each time using a randomly selected
training set of 80% of all hands (200 or 148
hands when mild and moderate CTS cases
were omitted), and the percentage of
successful classifications was calculated from
the mean of the five attempts.
STATISTICAL ANALYSIS
All statistical analyses were performed using
the R Project for Statistical Computing, a
software language and environment for
statistical computing, which is available as
free software under the terms of the GNU
General Public License.23 Where appropriate,

784
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

data were presented as mean ± SD or as beneficial influence of palmar segments on

percentages. Comparisons between patient
and volunteer groups regarding segment
temperatures were analysed by the non-
parametric Mann–Whitney–Wilcoxon test
since it can be used to compare two
independent groups of sampled data and,
unlike the parametric t-test, it makes no
assumptions about the distribution of the
data (e.g. normality). A P-value of < 0.05 was
considered to be statistically significant.
Results
A total of 71 patients and 57 volunteers
participated in the study. In total, 502 images
of 251 hands (dorsal and palmar side of each
hand) were acquired; of these, 132 were
healthy and 119 were pathological. Out of 71
patients there were 52 females and 19 males
(mean age 56.8 years, range 23 – 90 years),
who contributed 274 images of 137 hands
(the exclusion criteria meant that only one
hand was suitable in five patients). According
to CTS severity based on NCS, there were 33
mild, 33 moderate and 53 severe cases, while
18 hands were unaffected. The volunteers (35
females, 22 males; mean age 47.6 years,
range 25 – 74 years) contributed 228 images
of 114 healthy hands.
As our preliminary results8 showed no
the outcome of classification, this study
focused on dorsal segments. The mean
classification success rate of five
classification runs, when only dorsal
segments were included and no distinction
was made between patients with different
NCS severities, was 72.2% (Table 1). When
all segments were included (dorsal and
palmar), however, the mean success rate
dropped considerably to 62.0% (Table 2),
and even more so when only palmar
segments were considered (58.0%; Table 3).
These results would suggest that palmar
segments are not as important as dorsal
segments and they may even interfere with
successful classification. This presumption
was supported by statistical analysis. A non-
parametric Mann–Whitney–Wilcoxon test
was performed, which confirmed the
importance of dorsal segments when
compared with palmar segments for the
diagnosis of CTS using IRT (Table 4). With
one exception, the temperature of all dorsal
segments of the patients’ hands was
significantly different compared with the
healthy hands of volunteers (P < 0.05).
Furthermore, when the analysis was focused
only on severe cases (according to NCS) and
omitted the mild and moderate cases from

TABLE 1:
Mean classification success rate of five classification runs for diagnosis when only dorsal
segments were included and no distinction was made between patients with different
severities of CTS (n = 51; 20% of all hands)
CTS correctly CTS incorrectly
Classification run classified, n classified, n Success rate (%)
1 37 14 72.5
4 38 13 74.5
3 37 14 72.5
4 36 15 70.6
5 36 15 70.6
Mean ± SD 36.8 ± 0.84 14.2 ± 0.84 72.2 ± 1.64
CTS, carpal tunnel syndrome.

785
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

TABLE 2:
Mean classification success rate of five classification runs for the diagnosis of CTS when
dorsal and palmar segments were included and no distinction was made between patients
with different severities of CTS (n = 51; 20% of all hands)
CTS correctly CTS incorrectly
Classification run classified, n classified, n Success rate (%)
1 33 18 64.7
2 31 20 60.8
3 32 19 62.7
4 30 21 58.8
5 32 19 62.7
Mean ± SD 31.6 ± 1.14 19.4 ± 1.14 62.0 ± 2.23
CTS, carpal tunnel syndrome.

TABLE 3:
Mean classification success rate of five classification runs for the diagnosis of CTS when
only palmar segments were included and no distinction was made between patients with
different severities of CTS (n = 51; 20% of all hands)
CTS correctly CTS incorrectly
Classification run classified, n classified, n Success rate (%)
1 29 22 56.9
2 31 20 60.8
3 29 22 56.9
4 30 21 58.8
5 29 22 56.9
Mean ± SD 29.6 ± 0.89 21.4 ± 0.89 58.0 ± 1.75
CTS, carpal tunnel syndrome.

TABLE 4:
Statistical significance of the differences between segment temperatures in patients’ and
volunteers’ hands according to the non-parametric Mann–Whitney–Wilcoxon test
Statistical significance
Segment Dorsal side Palmar side
Whole hand 0.007162 0.1621 (NS)
First finger (thumb) 0.034990 0.1676 (NS)
Second (index) finger 0.029850 0.2174 (NS)
Third (middle) finger 0.017270 0.1169 (NS)
Fourth (ring) finger 0.017840 0.1701 (NS)
Fifth (little) finger 0.053060 (NS) 0.2399 (NS)
Below first finger 0.008782 0.0939 (NS)
Below second finger 0.004250 0.1610 (NS)
Below third finger 0.003147 0.1722 (NS)
Below fourth finger 0.003551 0.2700 (NS)
Below fifth finger 0.011470 0.4620 (NS)
NS, not statistically significant (P > 0.05).

786
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

the database (n = 185), the classification
results improved (83.2%; Table 5). The
decision to focus on severe cases only
originated from results of the statistical
analysis that compared dorsal segments
across different NCS severity levels (Table 6).
When the dorsal segments of all CTS cases
were compared with the corresponding
segments of healthy hands, there were
statistically significant differences between
all of them with only one exception (fifth
[little] finger). The differences between the
segments of hands of severe CTS cases and
segments of healthy hands were even more
significant (Fig. 5).
Discussion
In recent years, artificial intelligence-based
systems have proved to be useful and
successful tools that have often been used for
decision support, data mining and
knowledge discovery in medicine. Despite its
practical advantages over EMG, IRT has not
established itself in the field of CTS

TABLE 5:
Mean classification success rate of five classification runs for the diagnosis of CTS when only
dorsal segments of severe cases and healthy hands were used (n = 37; 20% of all hands)
CTS correctly CTS incorrectly
Classification run classified, n classified, n Success rate (%)
1 31 6 83.8
2 32 5 86.5
3 29 8 78.4
4 32 5 86.5
5 30 7 81.1
Mean ± SD 30.8 ± 1.30 6.2 ± 1.30 83.24 ± 3.52
CTS, carpal tunnel syndrome.

TABLE 6:
Statistical significances of the differences between dorsal segment temperatures for
different NCS severity levels of CTS according to the non-parametric Mann–Whitney–
Wilcoxon test
Severe– Moderate– Mild– All CTS–
Segment Normal normal normal normal
Whole hand 1.79 × 10-6 0.285800 (NS) 0.237400 (NS) 0.007162
First finger (thumb) 5.07 × 10-5 0.509300 (NS) 0.244800 (NS) 0.034990
Second (index) finger 6.19 × 10-5 0.449800 (NS) 0.295100 (NS) 0.029850
Third (middle) finger 4.90 × 10-5 0.397900 (NS) 0.470900 (NS) 0.017270
Fourth (ring) finger 3.49 × 10-5 0.557500 (NS) 0.550600 (NS) 0.017840
Fifth (little) finger 4.03 × 10-4 0.619200 (NS) 0.473400 (NS) 0.053060
Below first finger 2.32 × 10-6 0.165400 (NS) 0.111200 (NS) 0.008782
Below second finger 3.36 × 10-7 0.218600 (NS) 0.166000 (NS) 0.004250
Below third finger 1.42 × 10-7 0.228700 (NS) 0.175600 (NS) 0.003147
Below fourth finger 1.24 × 10-7 0.279400 (NS) 0.176200 (NS) 0.003551
Below fifth finger 1.69 × 10-7 0.318200 (NS) 0.132300 (NS) 0.011470
NS, not statistically significant (P > 0.05).
CTS, carpal tunnel syndrome; NCS, nerve conduction studies.

787
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

Severe CTS vs normal All CTS vs normal

1.000000
Log scale of statistical significance (P-value)

0.100000 P = 0.053060 (NS)

0.010000

0.001000

0.000100

0.000010

0.000001

0.000000

an
d b) ge
r
ge
r r
ge inge inge
r r er er er er
h h um ) fin ) fin ) fin f f f ing f ing fing fing
ole (t x le ng little
)
fir
st
on
d ird urth fifth
W
h er de dd (ri ec th fo
f ing d (in (mi rth h
(
l o w s o w low
i r st o n i r d ou F ift Be l ow B el e low Be
F c Th F B e B
Se

FIGURE 5: The statistical significances of the differences between the dorsal segment
temperatures in the group with ‘severe’ carpal tunnel syndrome (CTS) and the
‘normal’ healthy group (bottom line, data from Table 6) were even more pronounced
than the statistical significances of the differences between the dorsal segment
temperatures in all CTS severity groups and the ‘normal’ healthy group (top line, data
from Table 4) (NS, not statistically significant [P > 0.05])

diagnosis.7 Results from our previous hands. Consistent with other studies, the
research showed how an intelligent system present study clearly demonstrated the
based on an ANN can be used to classify validity of this hypothesis (Table 4).5,6,25 In
thermal images depicting healthy and CTS addition, it was ascertained that palmar
hands without the need for a cold stress test,8 segments are of no use when using IRT to
which is considered to be the standard diagnose CTS.
provocation procedure when diagnosing One of the initial goals, after achieving
nerve entrapments.24 High classification promising results on the smaller image
success rates (> 80% on several occasions) database, was to extend the classification of
were achieved on a small set of images (n = CTS cases to different severity levels
112).8 Due to the possibility of misleading according to NCS. For this purpose, the ANN
results from such a small initial image set, a was redesigned to be able to classify input
much larger database (n = 502) was then data into four different classes (normal,
acquired. The present study tested a mild, moderate and severe); however the
hypothesis, already discussed in the classification results were poor. Another
literature5 and suggested by the results of our established machine-learning tool,26 a
previous study,8 that dorsal segments are of decision tree, was also tested but did not
greater importance than palmar segments produce useful results. Thus, the idea of
when analysing thermal images of CTS trying to develop a direct association

788
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography
between ANN outputs and NCS severity
levels was abandoned. Based on statistical
analysis that showed the most significant
differences were between severe CTS cases
and healthy hands, it was decided to include
only severe cases. After training and testing
the ANN on this reduced database (dorsal
images of severe CTS level and healthy
hands only, n = 185), the best results were
achieved with a mean classification success
rate that exceeded 83% (Table 5). These
results might suggest that IRT could be used
during NCS diagnostic procedures, bearing
in mind that it is non-invasive and patient-
friendly when compared with EMG.
Nevertheless, IRT cannot be recommended as
an adequate diagnostic tool when an exact
diagnosis of CTS severity level is required. It
could, however, be used as a screening
method when dealing with populations with
high ergonomic risk factors for developing
CTS, such as repetitive hand and wrist use,
work with vibrating tools, hand use in the
cold, prolonged wrist extension and flexion,
• Received for publication 22 December 2008 • Accepted subject to revision 10 January 2009
• Revised accepted 11 May 2009
Copyright © 2009 Field House Publishing LLP
References
1 Smutz WP, Miller SC, Eaton CJ, et al:
Investigation of low-force high-frequency
activities on the development of carpal-tunnel
syndrome. Clin Biomech 1994; 9: 15 – 20.
2 Stevens JC, Sun S, Beard CM, et al: Carpal
tunnel syndrome in Rochester, Minnesota 1961
to 1980. Neurology 1988; 38: 134 – 138.
3 von Schroeder HP, Botte JM: Carpal tunnel
syndrome. Hand Clinics 1996; 12: 643 – 655.
4 Hackett MEJ: The place of thermography in
medicine. Acta Thermographica 1976; 1: 176 –
180.
5 Herrick RT, Herrick SK: Thermography in
detection of carpal tunnel syndrome and other
compressive neuropathies. J Hand Surg (Am)
1987; 12: 943 – 949.
6 Tchou S, Costich JF, Burgess RC, et al:
Thermographic observation in unilateral
carpal tunnel syndrome: report of 61 cases. J
Hand Surg (Am) 1992; 17: 631 – 637.
7 American College of Neurology Therapeutics
789
and sustained wrist or palmar pressure. As
CTS has reached epidemic proportions in
many occupations and is a major cause of
loss of work and workers’ compensation,1
many researchers in both medical and non-
medical fields are looking for a CTS
diagnostic method that could be also used in
the area of occupational health and safety.
According to the findings of the present
analysis of severe cases of CTS,
thermography-based methods could be
recommended for screening for severe nerve
entrapments since these methods would help
reduce long waiting lists for EMG
investigations. The method developed and
described in the present study, with its clear
identification of severe CTS cases, could be
used to prioritize patients on a waiting list
for EMG diagnosis as EMG is still the only
reliable gold standard.
Conflicts of interest
The authors had no conflicts of interest to
declare in relation to this article.
and Technology Assessment Subcommittee:
Assessment: thermography in neurologic
practice. Neurology 1990; 40: 523 – 525.
8 Jesenšek Papež B, Palfy M, Turk Z: Infrared
thermography based on artificial intelligence
for carpal tunnel syndrome diagnosis. J Int Med
Res 2008; 36: 1363 – 1370.
9 Rojas R: Perceptron learning. In: Neural
Networks: a Systematic Introduction. New York:
Springer, 1996; pp 77 – 99.
10 Rosenbaum RB, Ochoa JL: Carpal Tunnel
Syndrome and Other Disorders of the Median
Nerve: Clinical Presentation, 2nd edn. Boston:
Butterworth–Heinemann, 2002; pp 31 – 54.
11 Bland JDP: Carpal tunnel syndrome. BMJ 2007;
335: 343 – 346.
12 Viera AJ: Managment of carpal tunnel
syndrome. Am Fam Physician 2003; 68: 265 –
272, 279 – 280.
13 Stevens JC: AAEE minimonograph #26: the
electrodiagnosing of carpal tunnel syndrome.
Muscle Nerve 1987; 10: 99 – 113.
 B Jesenšek Papež, M Palfy, M Mertik et al.
Artificial intelligence-based infrared thermography

14 Palfy M, Zorman M: Automatic identification Learning internal representations by error

of histological features in histological images
using artificial neural networks. Proceedings of
the 14th International Electrotechnical and
Computer Science Conference ERK 2005. 26 – 28
September, 2005, Portorož, Slovenia. Vol B: pp
330 – 333.
15 Ammer K: Diagnosis of nerve entrapment
syndromes by thermal imaging. Proceedings of
the First Joint BMES/EMBS Conference 1999. 13 –
16 October 1999, Atlanta, GA, USA. Vol 2: p 1117,
DOI: 10.1109/IEMBS.1999.804279 (available
from: http://ieeexplore.ieee.org/Xplore/login.jsp?
url=/stamp/stamp.jsp?arnumber=804279&isnu
mber=17403).
16 Schartelmüller T, Ammer K: Infrared
thermography for the diagnosis of thoracic
outlet syndrome. Thermologie Österreich 1996; 6:
130 – 134.
17 Larose DT: Discovering Knowledge in Data: an
Introduction to Data Mining. Hoboken (NJ): John
Wiley, 2005; pp 128 – 145.
18 Hand DJ, Mannila H, Smyth P: Principles of Data
Mining. Cambridge (MA): MIT Press, 2001; pp
391 – 397.
19 McCulloch WS, Pitts W: A logical calculus of the
ideas immanent in nervous activity. Bull Math
Biophys 1943; 5: 115 – 133.
20 Rumelhart DE, Hinton GE, Williams RJ:
propagation. In: Parallel Distributed Processing:
Explorations in the Microstructure of Cognition, Vol
1. Cambridge (MA): MIT Press, 1986; pp 318 –
362.
21 Kasabov NK: Foundations of Neural Networks,
Fuzzy Systems, and Knowledge Engineering.
Cambridge (MA): MIT Press, 1996; pp 251 –
322.
22 Kantardzic M: Data Mining: Concepts, Models,
Methods, and Algorithms. New York: John Wiley,
2003; pp 213 – 223.
23 The R Foundation for Statistical Computing:
The R Project for Statistical Computing, 2008.
Vienna: The R Foundation for Statistical
Computing (available from: http://www.r-
project.org/).
24 Ring EFJ: Cold stress testing of the hand. In: The
Thermal Image in Medicine and Biology (Ammer
K, Ring EFJ, eds). Vienna: Uhlen Verlag, 1995;
pp 237 – 240.
25 Ammer K: Nerve Entrapment and Skin
Temperature of the Human Hand. A Casebook of
Infrared Imaging in Clinical Medicine. Warsaw:
Medpress, 2003; pp 94 – 96.
26 Podgorelec V, Kokol P, Stiglic B, et al: Decision
trees: an overview and their use in medicine. J
Med Syst 2002; 26: 445 – 463.

Author’s address for correspondence

Dr B Jesenšek Papež
Department of Physical Medicine and Rehabilitation, Medical Centre Maribor,
Ljubljanska 5, SI-2000 Maribor, Slovenia.
E-mail: breda.jesensek@ukc-mb.si

790