Performance Evaluation of Neural Classifiers Through Confusion Matrices To Diagnose Skin Conditions

İhsan Ömür BUCAK
International Journal of Artificial Intelligence and Expert Systems (IJAE), Volume (5) : Issue (2) : 2014 15
Performance Evaluation of Neural Classifiers Through
Confusion Matrices To Diagnose Skin Conditions
İhsan Ömür BUCAK iobucak@meliksah.edu.tr
Faculty of Engineering and Architecture
Department of Electrical-Electronics Engineering
Melikşah University
Talas-Kayseri, 38280, Turkiye
Abstract
In this paper we have aimed to diagnose skin conditions using Artificial Intelligence (AI) based
classifier algorithms and do the performance analyses of those presented algorithms through
confusion matrices. These algorithms are being used in a large array of different areas including
medicine, and display very distinct characteristics in the sense that they are grouped under
different categories such as supervised, unsupervised, statistical, or optimization. The objective of
this study is to diagnose skin conditions using seven different well-known and popular as well as
emerging Artificial Intelligence based algorithms and to help general practitioners and/or
dermatologists develop a careful and supportive approach that leads to a probable diagnosis of
skin conditions or diseases. These algorithms we chose as neural classifiers include Back-
Propagation (BP), Random Forest (RF), Support Vector Machines (SVMs), Linear Vector
Quantization (LVQ), Self-Organizing Maps (SOMs), Naïve Bayes, and finally Bayesian Networks.
All of these algorithms have been tested and their results of diagnosing skin conditions/diseases
by using data set from Dermatology Database have been compared.
Keywords: Artificial Intelligence, Artificial Neural Networks, Medical Diagnosing, Neural
Classifiers, Skin Conditions/Diseases, Confusion Matrix, F-score.
1. INTRODUCTION
Diagnosing skin conditions and diseases at an early stage is an extremely important in controlling
the increasing global numbers of skin cancer cases each year at an alarming rate. The numbers
are devastating worldwide and have doubled in the last 15 years [1]. Globally, every third case
diagnosed as a cancer belongs to a skin cancer and, in the USA alone; every fifth American is
expected to catch the skin cancer during their life time [1, 2].
If, particularly, the symptoms are not treated at their early stages, they mostly become very
dangerous. Furthermore, many symptoms look alike even though they are caused by different
diseases. This may even bring about more challenges for dermatologists to make correct
diagnosis. Erythema, scaling, definite borders, and size or density of the skin changes are very
important clinical features. Modern medicine along with the new technologies comes to
dermatologists’ rescue in the correct diagnosis of skin conditions in two ways: either from its
symptoms or through diagnostic examination. The diagnostic examination will include a range of
tests such as skin biopsy, blood test, X-rays, MRI/CT scans, allergy testing, and swab from a
sensitive area as well as a physical examination to diagnose the skin condition. Most of these
tests are used as standard diagnostic tools to help determine the causes of any skin condition
and have been proven to be very effective at diagnosing the skin condition. Skin conditions may
also have a genetic connection as the cause of any particular skin condition. Therefore, it is
important to share any past history of skin problems, family history with dermatologist at the time
of examination as many skin conditions are inherited from family.

İhsan Ömür BUCAK
A systematic approach to diagnosing skin conditions has also been developed. Five point of
descriptive paradigm can be applied to both rashes and lesions. These are distribution,
demarcation, color, surface change, and morphological features.
The objective of this paper is to diagnose skin conditions by using Artificial Intelligence based
(ANN) seven algorithms and compare their results against training and testing performances.
These algorithms, also known as neural classifiers, should be able to learn from the skin data set
of very large dermatology database [3], and return the probability of existence of any diagnosed
disease by helping dermatologists to improve the diagnosis of dermatological disease.
2. METHODS AND ALGORITHMS
In this paper, neural classifiers are based on the field of machine learning. Machine learning is
one of the six disciplines within Artificial Intelligence (AI) and make up most of AI along with
natural language processing, knowledge representation, automated reasoning, computer vision,
and robotics [4]. Machine learning was first defined by A. Samuel in 1950 as a field that gives
computers a specific capability of learning with no explicit programming aid. Russell and Norvig
atributes an ability to machine learning for adapting itself to new conditions and for detecting and
extrapolating patterns [4]. The field of machine learning usually distinguishes three types of
feedback to determine the nature of the learning problem agent faces: supervised, unsupervised,
and reinforcement learning [4]. In this paper, most of the algorithms under consideration lie on the
first two categories of these three cases. One of the objectives in this study is to compare the
supervised and unsupervised learning algorithms.
2.1 Backpropagation Algorithm
Artiﬁcial Neural Networks (ANNs) are the most effective approaches for pattern recognition [5-9].
ANNs cannot succeed a medical doctor, but they can help the doctor in identification of problem
or diagnosis of the disease [10]. Backpropagation (a.k.a propagation of error) is a familiar method
used to teach Artiﬁcial Neural Networks upon how to accomplish a given task. Bryson and Ho
described it first in 1969 [11], but the work of Rumelhart and his colleagues led it to gain
recognition in 1986, and the algorithm started thereafter to evolve in the research area of artificial
neural networks [12].
Backpropagation algorithm trains the artificial neural network by presenting a set of exemplary
input/output pairs to the network. The algorithm adjusts the internal weights of the network
iteratively to minimize the mean squared error between the network’s predicted outcome and the
true result as the set of different training examples in the form input/output pairs are introduced to
the system. The weight adjustment will yield an average response since the error is acquired over
the entire training set. For the network to be able to accurately categorize any input set, it is
significant that the training set contain examples that represent the full range of possible inputs.
The new input pattern is expected to be accurately categorized by the network even if the new
input sets and any of the training patterns do not seem alike.
Backpropagation is a supervised learning method whose objective is to find a mapping function
that establishes the best mathematical relation between a set of inputs and its correct output.
Single layer neural networks also known as perceptrons can easily learn some input and output
patterns. However, these single layer perceptrons are incapable of learning some linearly
separable patterns. A multi-layered network can overcome its limitation of having only one single
layer by adding more layers (effectively more internal layers) to learn different abstract features in
each layer by finding patterns in the layer below it. Therefore, the ability to find patterns between
the current and prior layers successively will create internal representations. The motivation
behind the backpropagation lies in the methodology to train multi-layered neural networks in such
a way as to make it learn the convenient internal representations for its granted learning of any
random mapping between input and output. It needs a training set created by a dataset of the
desired output in consideration of many inputs. It is most beneficial for feed-forward type

İhsan Ömür BUCAK
networks. It also requires a differentiable activation function for its artificial neurons.
Fundamentally, the steps of the backpropagation algorithm can be summarized as follows [13]:
1. Introduce a training pattern to the network.
2. Compare the network’s output with the desired one and calculate the error for each output
neuron.
3. Calculate each neuron’s output with its local error.
4. Adjust each neuron’s weight to minimize the local error.
5. Assign “blame” to neurons for the local error at the former level.
6. Repeat the process starting with step 3 using each neuron’s “blame” as its error on the ones
at the former level.
2.2 Linear Vector Quantization (LVQ)
LVQ was developed by Kohonen in 1984 [14]. Its basic tenet is to represent an n-dimensional
vector by lower dimensional vectors. Learning in LVQ is to determine which output vector will
match up with an input vector. Vector set in the output layer is called reference vector. The
objective is to match up the input vector with the reference vectors. Generally, it is used in the
solution of classification problems. Only one of the output vectors takes the value of 1 and the
others take 0. The reference vector with the value of 1 determines the class of input vector. The
distance between the inputs and the outputs is found by Euclidian method. It runs according to
“winner takes all” principle, and not the all weights but the weights of the winning reference vector
only is updated in such a way that it gets closer to the input vector if both belong to the same
class, and is moved away from it if they belong to different classes [15, 16].
LVQ is a supervised classifier and is used to adjust the bounds between classes to maximize
classification. It has three layers like multi-layer nets. There is no information processing in the
input layer. However, data entry to LVQ net is performed through the input layer. Distance that
will provide the closest vector to the input vector is computed in hidden or Kohonen layer (a.k.a.
intermediate layer). Each weight vector between the input and the intermediate layer is called
reference vector. Assigning or classifying the output layer to one of the reference vectors is
fulfilled in output layer. The weights between the hidden and the output layers are constants and
their value are equal to 1. An LVQ net includes a single layer of nodes where each node shows a
class. The weight adjustment procedure on this single layer’s nodes is applied to adapt each
node to the input vectors occurring with the corresponding class. LVQ is trained as based on the
following steps:
1. Assign very small numbers between 0 and 1 to weight vectors initially.
2. Pick up a training vector.
3. Discover the class (node) of minimum distance with this training vector.
4. Update the node’s weights in accordance with the following expression:
If the target and the output class are the same,
)(
oldjoldjnewj wxww −+= α (1)
else
)(
oldjoldjnewj wxww −−= α . (2)
5. Reduce learning rate α.
6. Go to step 2 and repeat until the termination criterion is met.

İhsan Ömür BUCAK
LVQ usually performs efficiently if the initial weight vectors are close to their final values. That is
why some heuristic procedures are used for weight initialization, like Kohonen's SOM or any
other vector quantization procedure, like k-means.
2.3 Self Organizing Map (SOM)
SOM is an unsupervised learning type of artificial neural network that transforms its input space
of training samples with arbitrary dimension into a low-dimensional discrete map by preserving
topological properties during its training session [17, 18]. Self-organizing maps, not the other
artificial neural networks, use only a neighborhood function in order not to lose the topological
properties of the input space [18]. It was first introduced by Teuvo Kohonen, and therefore is also
known as Kohonen map [19]. It accomplishes a competitive learning. Euclidian distances of all
weight vectors to the training samples are computed for a given training sample to the network.
Best identical neuron with the input is called the best matching unit (BMU). The BMU weights and
the neurons in its instantaneous neighborhood are adjusted in connection with the input vector.
The amount of the change is expected to become smaller over time and with distance from the
BMU. A neuron with weight vector )(tWv is updated by the following equation:
)),()()((),()()1( tWtDttvtWtW vvv −Θ+=+ α (3)
where α(t) is a monotonically decreasing learning parameter, Θ(v, t) is the neighborhood function
that depends on the distance between the BMU and neuron v, and D(t) is the input vector. The
neighborhood function generally utilizes a Gaussian function. Nevertheless, it can simply be used
as one for all neurons in the BMU neighborhood and zero otherwise. The neighborhood function
concurs with time without regard to the functional form [20]. When the neighborhood is large
initially, the self-organizing takes place in the overall scope. When the neighborhood is shrunken
to just a couple of neurons the weights converge to their local estimates.
The neuron whose weight vector stands closest to the input vector at the end of mapping is called
winning neuron. This is clearly resolved by computing the Euclidean distance between input and
weight vectors.
2.4. Support Vector Machines (SVMs)
Another methodology used for classification purposes is known to be Support Vector Machines
[21, 22]. In this method, the classification is aimed to be done through the aid of a linear or a
nonlinear function. This method is based on the estimation of the most suitable function to
separate the data from each other. A hyperplane is a concept in geometry. It is a generalization of
the plane into a different number of dimensions. A H0 hyperplane in terms of the points on itself
can be expressed as
0:0 =+bXWH T
. (4)
If the space between the two sides of H0 hyperplane, say, H1 and H2, is required to be maximized,
the following objective function needs to solved:
Minimize
2
wwT
(5)
constrained by )...2,1(1)( nibxwy i
T
i =≥+ where W represents the weight vector, X
represents the training set, y represents the classes, b is a constant number, and finally, n the
number of attributes. The problem of the nonseperability of the data can be resolved through the
addition of nonnegative and error-indicative ξi loose variables into the optimization model. The
cases where the data can not be separated in linear or linear classifiers are not complex enough
sometimes, then the nonlinear classifiers can be used instead of the linear ones to map the data

İhsan Ömür BUCAK
into a richer feature space including nonlinear ones by constructing a hyperplane in that space. In
these cases, φ functions can be utilized to analyze by carrying the data to the spaces with larger
dimensions. φ functions are used for choosing a good mapping which turns out to be the
improved results for the problem. Encoding prior knowledge and getting the right complexity of
function class will serve the understanding of choosing the good mapping [22].
2.5. Random Forests (RFs)
Random forest is a collection of CART-like decision trees that follows certain rules. There is no
pruning in RF as opposed to classical decision tree since the ensemble and bootstrapping
schemes help random forest overcome overﬁtting issues [23-25]. Each tree is constructed using
the following algorithm:
Step 1. Form a training set by N sample cases for this tree in which the number of training sample
cases is chosen at random but with replacement from original data (i.e., take a bootstrap sample).
This sample will be the training set for growing the tree.
Step 2. Specify a number m of the M input variables to determine the decision at a node of the
tree. The number m should be less than the number M such that at each node, m variables are
selected at random out of the M. Calculate the best split based on these m variables in the
training set to split the node. The value of m is held constant during the forest growing.
Step 3. Each tree is fully grown to the largest extent possible and not pruned.
Forest error rate is dependent on two things: correlation and strength. Correlation is directly
proportional to the forest error rate. Increasing the correlation between any two trees in the forest
will increase the forest error rate. Strength is attributed to each individual tree in the classifier.
Strength of the individual trees is inversely proportional to the forest error rate. That is, increasing
the strength of the individual trees will decrease the forest error rate
The value of m affects both the correlation and the strength. Therefore, reducing it reduces both
and increasing it increases both. m lies on somewhere in between as its optimal range. A value of
m in the range can quickly be found by the oob (out-of-bag) error estimation rate. Random forests
is somewhat sensitive to this adjustable parameter only.
2.6 Naïve Bayes Classifier
A Naïve Bayes classifier is a simple probabilistic feature model based on an application of Bayes'
theorem along with an assumption of strong independence on a feature of a class. Here, the
more explanatory term for the probabilistic model is “independent feature model” and the prefix
naïve is driven by the word “strong”. Independence means that all the features of a class
contribute independently to the probability of a particular object that is classified. Simply, the
presence of one feature of the class has no relationship, on no account whatsoever, with the
presence of any other feature, given the class variable. Although it is simplified considerably by
its assumptions and naïve design, these types of classifiers have been proven to perform
reasonably better in many complicated real-world environments [26-28].
Naïve Bayes models use the maximum likelihood method in many practical applications for
estimating the particular values of the mean and variance as the model parameters to make the
observed results the most likely, given the model. Generally, the model parameters are picked
out by the maximum likelihood method to maximize the likelihood function between the model
picked and the data observed [27, 28].
Naïve Bayes has the advantage of requiring only a small percent of training data to estimate the
parameters when it is used as classifier, and is inclined to train very efficiently for some types of
probability models in a supervised learning paradigm. Additionally, each class requires only the

İhsan Ömür BUCAK
calculation of the variances of the variables, not the entire covariance matrix since the variables
are assumed independent [27, 28].
2.7 Bayes Nets
Bayes Nets or Bayesian networks are graphical representation for probabilistic relationships
among a set of random variables. In contrast to undirected graphical models, which provide a
simple definition of independence, Bayesian networks, also called directed graphical models,
presents a more complicated concept of independence. However, their advantages are evident
and undeniable ones. Bayes nets can be used for causal, or "top down", reasoning. For example,
an arc from A to B can be treated causal as it is taking into consideration the directionality of the
arcs by indicating that A “causes” B. This can ease the way on how the graph is built [29].
Furthermore, Bayes nets can also be used to encode the deterministic relationships, and are
easier to learn as the model fits the data well [30, 31].
3. EXPERIMENTAL ANALYSIS
The differential diagnosis, which is a medical term defined as a systematic method of diagnosing
a disorder that lacks unique symptoms or signs [32], creates a substantial problem in
dermatology in diagnosing skin disease (erythemato-squamous). In this type of disease, the
clinical features of erythema and scaling are generally overlapped. The severity of erythema and
scaling, the presence of itching and koebner phenomenon, the formation of papules matter all
critically on every account. A biopsy is usually applied as the last option to explicitly diagnose the
disease of the clinical features detailed in [33]. Even these types of diseases can harbor many
histopathological features. Indicating the features of another disease initially, having the
characteristic features thereafter is a common cause that complicates the diagnosis [33].
Additionally, some skin samples may not show the typical histopathological features of the
disease [33]. In this work, the same dataset for classes and features (both clinical and
histopathological) used in [33] have also been employed.
The networks have been trained by using the training data set with 33 features. Furthermore, 40
inputs have been used for training and 20 inputs for testing for each class. 66.66 percent of the
entire data of 360 has been split up for training and 33.33 percent for testing. The input layer has
been composed of 33 nodes; each was a representative of 33 features from the data entry form.
The output layer has been composed of 6 nodes; each was a representative different class of
diagnostic diseases such as C1: Psoriasis, C2: Seborrheic dermatitis, C3: Lichen planus, C4:
Pityriasis rosea, C5: Chronic dermatitis, and C6: Pityriasis rubra pilaris [33].
4. PERFORMANCE EVALUATION BY CONFUSION MATRICES
In this paper, the objective was initially driven to compare supervised algorithms with
unsupervised ones for the given dataset above. Therefore, BP and LVQ were chosen randomly to
represent the supervised algorithms and SOM the unsupervised one. The error rates shown in
Figure 1 were discussed in detail in [34]. In summary, SOM yielded the highest error rate and
failed to separate one or two classes for some lower iteration numbers. Furthermore, LVQ
learned the earliest due possibly to the most appropriate weight vector initialization; hence it
achieved the maximum success rate the earliest, too. Although LVQ achieved a relatively fast
learning, BP generated the minimum error rate. The results have been obtained by using
MATLAB [35].

İhsan Ömür BUCAK
FIGURE 1: Error rates for BP, LVQ, and SOM algorithms [34].
Figure 2 shows the time performance of the supervised and unsupervised algorithms [34].
FIGURE 2: Time performance of BP, LVQ, and SOM algorithms [34].
However, a more comprehensive performance evaluation which also covers the emerging
algorithms as presented in this paper has been a prime motive to extend the work beyond these
three. Therefore, the need to extend the performance evaluation has been satisfied greatly by the
aid of confusion matrices. A confusion matrix is a specific table used to evaluate the performance
of a supervised or an unsupervised algorithm. Each column and row of the matrix represents the
instances in a predicted and an actual class, respectively. Since it is easy to see if the system is
confusing two classes, most possibly, in the form of mislabeling one as another, therefore the
name is given and used as the confusion matrix. The confusion matrix is often called the
contingency table or the error matrix outside the artificial intelligence community. The confusion
matrices below have been obtained for the algorithms, Naïve Bayes, Backpropagation, Random
forest, Support vector machines, Linear vector quantization, Self-organizing maps, and finally
Bayes theorem to diagnose the skin diseases.

İhsan Ömür BUCAK
NaïveBayes
1 2 3 4 5 6
20 0 0 0 0 0 1
0 18 0 0 0 2 2
0 0 20 0 0 0 3
0 0 0 20 0 0 4
0 0 0 0 20 0 5
0 0 0 0 0 20 6
BP
1 2 3 4 5 6
20 0 0 0 0 0 1
0 18 0 2 0 0 2
0 0 20 0 0 0 3
0 0 0 20 0 0 4
0 0 0 1 19 0 5
0 0 0 0 0 20 6
RF
1 2 3 4 5 6
20 0 0 0 0 0 1
1 17 0 2 0 0 2
0 0 20 0 0 0 3
0 0 0 20 0 0 4
0 0 0 0 20 0 5
0 0 0 0 0 20 6
SVM
1 2 3 4 5 6
20 0 0 0 0 0 1
0 18 0 2 0 0 2
0 0 20 0 0 0 3
0 0 0 20 0 0 4
0 0 0 0 20 0 5
0 0 0 0 0 20 6

İhsan Ömür BUCAK
LVQ
1 2 3 4 5 6
20 0 0 0 0 0 1
0 17 0 3 0 0 2
0 0 20 0 0 0 3
0 1 0 19 0 0 4
0 0 0 0 20 0 5
0 0 0 0 0 20 6
SOM
1 2 3 4 5 6
0 0 0 0 0 20 1
0 8 0 12 0 0 2
20 0 0 0 0 0 3
0 19 0 1 0 0 4
0 0 20 0 0 0 5
0 0 0 0 20 0 6
BayesNET
1 2 3 4 5 6
20 0 0 0 0 0 1
0 19 0 1 0 0 2
0 0 20 0 0 0 3
0 0 0 20 0 0 4
0 0 0 0 20 0 5
0 0 0 0 0 20 6
FIGURE 3: Confusion Matrices of the Neural Classifiers.
The basis of our problem underlies the classification of skin diseases which shows the close
symptoms to each other. Once the confusion matrices are studied in detail, it can be observed by
the employed methods that the diseases the 1st
, 3rd
, 5th
, and the 6th
can disassociate themselves
successfully. On the other hand, misclassification over the entire classes has occurred in the 2nd
and the 4th
classes, i.e., in the Seborrheic dermatitis, a chronic skin disease characterized by the
inflammation of the scalp or the face along with the appearance of scaly and flaky substance over
the surface of the skin, and Pityriasis rosea, a harmless, common skin condition that causes a
rash, although the methods differ greatly from each other. This can particularly be seen
prominently in the LVQ and SOM algorithms. Hence, the errors arising can be interpreted as the
result of these two diseases that gives more close symptoms as compared to the others, rather
than the methods used.
5. COMPARATIVE EVALUATION
As mentioned previously, the objective of this study aims to classify the skin conditions and
diseases which, in particular, shows the close symptoms to each other. Furthermore, we have
deliberated on diagnosing skin conditions and diseases that needed to be treated at their early
stages before getting out of control towards skin cancer.

İhsan Ömür BUCAK
Any comparative evaluation to the similar cases can only be meaningful. However, it is rather
difficult to observe any case study similar to the ones used in our research. In general, a huge
variety of applied methods and differing skin conditions and diseases as the application area can
complicate further the success of such comparative evaluation. We are confident that, in terms of
Artificial Intelligence and Machine Learning perspective, our algorithms not only cover the
conventional ones proven to be successful and reliable when applied to several medical
diagnostic tasks, but also include some recently developed methods that seemed promising for
overall medical diagnosis and treatment [36-39]. For example, we have obtained the best result
when we have applied the Bayesian network classifier. In addition, naïve Bayes has yielded very
good result, too. Our conclusion has also been drawn by some other experiments and studies in
the literature [36-39]. Therefore, it shows that physicians prefer explanations as based on the
Bayesian classifiers.
6. CONCLUSION
ANNs have been proven to be very efficient tools for pattern recognition and they can be
successfully used in dermatological applications. They can increase the detectability of
dangerous diseases and lower the mortality rate of patients. Both supervised and unsupervised
based neural network algorithms can readily be used for implementing an application similar to
the considered one here, and utilized in the field of medicine to determine the skin diseases.
Medical doctor has to decide if there is a need to analyze some part of the skin. Therefore, it is up
to the medical doctor what to do with the results after running the program. The outcome obtained
from the program can be very helpful. Although computers cannot replace the dermatologist, they
can make their work easier and more effective. Proposed study might be also very useful for
general practitioners, who do not have wide knowledge about dermatology.
FIGURE 4: Error Percentages of the Neural Classifiers.

İhsan Ömür BUCAK
FIGURE 5: F-scores of the Neural Classifiers.
The results have been summarized in the form of error percentages and F-scores as can be seen
from Figures 4 and 5. All the methodologies considered in this paper have produced satisfactory
results except for the SOM algorithm that has generated a relatively high percentage of error
(Please see Figures 4 and 5). Despite its lower success rate in this study, SOM is successful to
make or draw a distinction amongst classes, i.e., it is capable of classifying a difficult dataset as
an unsupervised learning method. Other experiments and/or studies in the literature tend to
encourage us to use this method for a wide range of analyses, even from such domain as
medical data. The best result has been obtained from Bayesian network classifier since it makes
its classification with an assumption of dependency among the features of its own data. One can
easily conclude that there exists correlations among the symptoms of a disease. Therefore, the
results obtained seem to be consistent if one considers the area of application. However, naive
Bayes has surprisingly produced a considerably good result in this application although, ironically,
it assumes no dependency among the features. Generally, ANNs which shows a lower
performance rate produces a higher accuracy rate once compared to SVMs. However, in this
study, SVM has produced a better result although the difference between them has remained
small; RF has produced very close result to BP is the another observation.
7. REFERENCES
[1] M. Antkowiak, Artiﬁcial Neural Networks vs. Support Vector Machines for Skin Diseases
Recognition, Msc. Thesis, Umea University, Sweden, 2006.
[2] http://www.skincancer.org/Skin-Cancer-Facts/, accessed May 27, 2010.
[3] A. Frank and A. Asuncion, UCI Machine Learning Repository [http://archive.ics.uci.edu/ml].
Irvine, CA: University of California, School of Information and Computer Science, 2010.
[4] S. Russell and P. Norvig, “Artificial Intelligence: A Modern Approach”, 3rd edition, Prentice
Hall, p.578, 2009.
[5] C.M. Bishop, “Neural networks and their applications”, Review of Scientiﬁc Instruments, 65(6),
pp. 1803–1832, 1994.
[6] C.M. Bishop, “Neural Networks for Pattern Recognition”, Clarendon Press, Oxford, UK, 1995.
[7] B.D. Ripley, “Pattern Recognition and Neural Networks”, Cambridge University Press,
Cambridge, UK, 1996.

İhsan Ömür BUCAK
[8] C. Stergiou and D. Siganos, “Neural networks”, SURPRISE 96 Journal, 4, 1996.
[9] R. Tadeusiewicz, “Sieci neuronowe, (Neural Networks)”, Akademicka Oﬁcyna Wydawnicza,
Warszawa, 1993.
[10] M. Antkowiak, “Recognition of skin diseases using artiﬁcial neural networks”, In Proceeding
of USCCS’05, pp. 313–325, 2005.
[11] A.E. Bryson and Y. Ho, “Applied optimal control: optimization, estimation, and control”,
Blaisdell Publishing Company or Xerox College Publishing, pp. 481, 1969.
[12] D.E. Rumelhart, G.E. Hinton, and R.I. Williams, “Learning representations by back-
propagating errors”, Nature, 323, pp. 533-536, 1986.
[13] E. Sivasankar, R.S. Rajesh, and S.R.Venkateswaran, “Diagnosing Appendicitis Using
Backpropagation Neural Network and Bayesian Based Classifier”, International Journal of
Computer Theory and Engineering, Vol. 1, No. 4, pp. 1793-8201, 2009.
[14] T. Kohonen, “Self-Organizing Maps”, Springer, Berlin, 1997.
[15] T. Kohonen, G. Barna, and R. Chrisley, “Statistical pattern recognition with neural networks:
Benchmarking studies”, In Proceedings of the International Conference on Neural Networks
(ICNN), Vol. I, pp. 61-68, 1988.
[16] T. Kohonen, Learning vector quantization. In: M.A. Arbib, editor, The Handbook of Brain
Theory and Neural Networks, pages 537–540. MIT Press, Cambridge, MA, 1995.
[17] http://www.nd.com/models/sofm.htm, Self-Organizing Feature Maps, accessed April 5, 2011.
[18] http://en.wikipedia.org/wiki/Self-organizing_map, Self-organizing map, accessed April 5,
2011.
[19] T. Kohonen and T. Honkela, “Kohonen network”, Scholarpedia 2007,
http://www.scholarpedia.org/article/Kohonen_network, accessed May 25, 2010.
[20] S. Haykin, “9. Self-organizing maps”, Neural networks - A comprehensive foundation (2nd
ed.), Prentice-Hall, 1999.
[21] N. Cristianini, “An Introduction to Support Vector Machines and Other Kernel-based Learning
Methods”, Cambridge University Press, 2000.
[22] B. Schlkopf and A.J. Smola, “Learning with Kernels: Support Vector Machines,
Regularization, Optimization, and Beyond”, MIT Press, 2001.
[23] L. Breiman, "Random Forests". Machine Learning 45 (1): 5–32, 2001.
[24] T.K. Ho, "Random Decision Forest". Proceedings of the 3rd International Conference on
Document Analysis and Recognition, Montreal, QC, 14–16 August 1995. pp. 278–282, 1995.
[25] T.K. Ho, "The Random Subspace Method for Constructing Decision Forests". IEEE
Transactions on Pattern Analysis and Machine Intelligence 20 (8): 832–844, 1998.
[26] M. Kantardzic. Data Mining - Concepts, Models, Methods, and Algorithms, IEEE Press,
Wiley-Interscience, 2003.
[27] S. Theodoridis and K. Koutroumbas. Pattern Recognition. San Diego: Academic Press,
2006.

İhsan Ömür BUCAK
[28] T.M. Mitchell. Machine Learning. McGraw-Hill, 1997.
[29] J. Pearl, Causality: Models, Reasoning, and Inference, Cambridge University Press, 2000.
[30] R.E. Neapolitan. Learning Bayesian networks. Prentice Hall, 2004.
[31] N. Friedman, M. Linial, I. Nachman, and D. Pe'er, "Using Bayesian Networks to Analyze
Expression Data", Journal of Computational Biology (Larchmont, New York: Mary Ann
Liebert, Inc.) 7 (3/4): 601–620, 2000.
[32] http://www.thefreedictionary.com/differential+diagnosis, accessed October 7, 2013.
[33] G. Demiröz, H. A. Güvenir, and N. İlter, “Learning differential diagnosis of eryhemato-
squamous diseases using voting feature intervals”, Artificial Intelligence in Medicine, 13, pp.
147-165, 1998.
[34] İ. Ö. Bucak, Y. Sahin, and A. Serbetci, “Supervised vs. Unsupervised neural classifiers for
diagnosing skin diseases”, The Proceedings of Second International Conference on
Informatics (ICI’2011), Çanakkale-Turkiye, pp. 73-80, 27-29 April 2011.
[35] MATLAB R2008b, Natick, Massachusetts: The MathWorks Inc., 2008.
[36] K.K Manjusha, K. Sankaranarayanan, and P. Seena, “Prediction of Different Dermatological
Conditions Using naïve Bayesian Classification”, International Journal of Advanced Research
in Computer Science and Software Engineering, Volume 4, Issue 1, pp. 864-868, 2014.
[37] R. Blanco, I. Inza, M. Merino, J. Quiroga, and P. Larranaga, “Feature selection in Bayesian
classifiers for the prognosis of survival of cirrhotic patients treated with TIPS”, Journal of
Biomedical Informatics, Volume 38, Issue 5, pp. 376-388, 2005.
[38] S. Shirali-Shahreza, M.E. Mousavi, “A New Bayesian Classifier for Skin Detection”, 3rd
International Conference on Innovation Computing and Control, pp. 172-175, 2008.
[39] A. Masood, A.A. Al-Jumaily, “Computer Aided Diagnostic Support System for Skin Cancer: A
Review of Techniques and Algorithms”, International Journal of Biomedical Imaging, Volume
2013, 22 pages, 2013.

Performance Evaluation of Neural Classifiers Through Confusion Matrices To Diagnose Skin Conditions

More Related Content

What's hot (17)

Viewers also liked (10)

Similar to Performance Evaluation of Neural Classifiers Through Confusion Matrices To Diagnose Skin Conditions (20)

More from Waqas Tariq (20)

Recently uploaded (20)

Performance Evaluation of Neural Classifiers Through Confusion Matrices To Diagnose Skin Conditions