Introduction
Prostate cancer is the most common malignancy in men and majority leading cause of cancer deaths in the Eastern world. The genetic predisposition to prostate cancer is well established, as genomic instability is a common feature of many human cancers. Epidemiological studies have suggested that several risk factors (Li, et al., 1997; Steck et al., 1997; Magnusson et al., 1998; Marshall, 1991). Knowing about the genetic markers of prostate cancer in men with prostate cancer diagnosis could help. Inactivation or deletion a large number of tumor-related genes, which otherwise regulate normal cellular growth and suppression of abnormal cell proliferation, is recognized to be one of the major mechanisms of tumorigenesis (Goddard and Solomon, 1993; Waite and Protean, 2002). The successful treatment of prostate cancer relies on detection of the disease at its earliest stages. Although prostate-specific antigen (PSA)-based screening has been a significant advance in the early diagnosis of prostate cancer, identifying specific genetic alterations in a given family or patient will allow more appropriate screening for early disease. Mapping and identification of specific prostate cancer susceptibility genes is slowly becoming a reality.

Immune System (IS) can be regarded as a defence mechanism of the body. It seeks the condition of body and explores if there is any dangerous situation. If so, the related units are put into effect and necessary processes are held. Artificial Immune System (AIS) is an artificial intelligence method whose roots lie to simple mathematical models developed for the understanding of natural immune system [(L. N. de Castro, and Timmis, 2002; L.N. de Castro, and Von Zuben, 1999; Dasgupta, 1998). Later, with properties like learning, memory, distributed and organized working, etc. in these models, researchers had begun to scrutinize the natural immune system as an inspiration source of AIS. Since that time, many methods modelling or inspiring from some metaphors in natural immune system have been developed and applied to various problems (Chaudhuri et al., 2007). Medical classification problems are among them. In this study we applied an Artificial Immune System (AIS) to the problem of prostate cancer classification. The used AIS models the clonal selection theory in IS as many other AISs. 17 data from healthy people and 33 data from patients were used for this classification process. The features of data are age, gleason score, personal history of cancer and family history of cancer. The 70% of data was taken for training and the remaining 30% was used for the test process. After analysing AIS with different parameters, best classification accuracy on test data was recorded as 93.33%. In other words, only one data was incorrectly classified.

The other type of classification with prostate cancer data was conducted with artificial neural networks (ANN) (Principe, et al., 1999). When this widely used effective classifier was applied the performance has increased to 100% by correctly classifying the whole test data.

1 Materials and Methods
1.1 Dataset formation
Paraffine blocks of prostate pathologies were derived from the archives of the Department of Pathology in Faculty of Medicine at the University of Selcuk, Turkey. Namely, Paraffine-embedded prostate cancer tissue specimens of 50 subjects were used in this study. Age range was 35-72 years, all subjects were males. 10 subjects had family history of cancer and 40 subjects were non family. These subjects went to physicians to demonstrate a variety of serious symptoms of prostate cancer, e.g., difficulty in voiding, urodynia, urgent and frequent urination, and hematuria. Their prostates were examined by one or more of the following means: rectal ultrasound detection, digital rectal examination, computed tomography, and magnetic resonance imaging. Biopsy was performed for the subjects who were suspected to have prostate cancer, and all specimens were from archived paraffin blocks.

Diagnosis of prostate cancer requires the tissue and cell specimens. These specimens are screened and analyzed by a pathologist using a microscope. Optimum medical treatment is decided according to this information gathered by the pathologist. In some cases, correct diagnosis is very hard and there can be 30-40% difference between pathologists’ decisions (Schenck, and Planding, 1998). Dramatic results about wrong diagnosis of cancer cases from biopsy slides can be found in (Kopec et al., 2003). Prostate cancer is evaluated using two staging systems: the Jewett-Whitmore system and the TNM (tumor, node, metastases) system. In TNM system, T refers to the size of the primary tumor, N will describe the extent of lymph node involvement, and M refers to the presence or absence of metastases.

In this study, 50 data with 4 features were used. The features are:
1.Age
2.Gleason score (PSA*)
3.Personal History of Cancer
4.Family History of Cancer

Here, PSA is a protein produced by the prostate gland that can be detected in the blood. Levels rise with age and when the prostate is enlarged. Significantly increased levels of PSA in the blood can indicate prostate cancer. PSA levels are also known to rise in other prostate conditions such as prostatitis (inflammation of the prostate). Normal values of PSA are as the following:

AGE / PSA Value
Under 50 years / < 2.5
50 – 59 years / < 3.5
60 – 69 years / < 4.5
70 years and over / < 6.5

The personal and family history of cancer in 3^rd and 4^th features are taken as 0 if there is no presence of cancer and 1 if there has been a cancer in the subject or his family. The 17 data belong to the healthy subjects while the remaining 33 data were of subjects with prostate cancer. The 70% of the whole data was taken for training, and 30% was taken for test. The division of dataset for training and test is shown in Table 1.

Table 1 The number of healthy and patient data in training and test sets

 
The AIS system was trained using training data and its performance in the test set was recorded. Also, to have an insight about the result of this study, a widely used method- Artificial Neural Networks- was also applied to the same data.

1.2 Used AIS system
The applied method is a simple AIS algorithm that mimics the clonal selection theory in natural immune system as many other clonal selection-based AIS methods (Ada, and Nossal, 1987; Chen, and Mahfouf, 2006; Cutello et al., 2005; L. N. de Castro, and Von Zuben, 2000; Garrett, 2003; Ong et al., 2005; Perelson, and Oster, 1979; Timmis, and Neal, 2001). The biological base for this theory can be found in (Abbas, 1994). The block diagram of used AIS method is shown in Figure 1.

Figure 1 The Flow chart of clonal based AIS system

 
Here, the system units are named as Antibodies (Ab) and inputs that are presented to the system are regarded as Antigens (Ag). A random population is formed in the beginning and the affinity of each member of this population to a presented input (Ag) is calculated using Euclidean distance criterion given below in Equation 1:

Affinity=1-D ;where

 
Here, Ab_k is the k^th feature of Ab vector and Ag_k is the k^th feature of Ag vector.

If the affinity of an Ab in the population exceeds a threshold value named as supp, that Ab is selected for cloning. Cloning is done by simply copying the vector of Ab with a number proportional to the Ab’s affinity and after that, a clone population is formed. After cloning, a mutation procedure is applied to the some Ab clones to have diversity in the population. The mutation is done through changing some values of Ab vectors by a new value which is determined randomly. That is, besides of random determination of which Abs will be mutated, the selection of features that will be changed and the determination of new feature values are also conducted in a random fashion. After these processes, a number of best Abs in mutated population is taken for the next iteration (the best Abs are the Abs whose affinities are highest). Also some randomly generated Abs are added to these Abs for the use in next iteration as a beginning population. The iterations are conducted a number of times a memory population is formed using best Abs produced after the iterations conducted for each Ag. The class information of that Ab is also taken with the same class of presented Ag. As a last step, the memory Abs are deleted from the memory population, if the affinity of them is higher that supp value with any other memory Ab (Figure 1).

Here a memory population including memory Abs and their class information is formed in training. The classification procedure on the other hand is done through finding the nearest memory Ab to the presented Ag and class decision about that Ag is given by looking the class information of that nearest Ab.

In our application, we conducted 100 iterations for each presented input data (Ag). The 50 Abs were used in the beginning population in each iteration. The determination of correct supp value was done in experimentally. That is the supp value was changed between 0.9-0.01 and for each experienced supp value, a training-test process was conducted to see the resulted test classification accuracy in percentage. Here this accuracy is calculated as:

 
where T is the set of data items to be classified (the test set), t?T, t.c is the class of the item t, and classify(t)
returns the classification of t by method. The best supp value is determined as the supp value for which the test classification accuracy is the highest.

After conducting the classification of prostate cancer data with AIS, an ANN structure was also trained and
tested for the same data. One-layer ANN architecture was used and gradient descent learning rule was utilized in the training. The optimum value of learning rate parameter, the number of hidden nodes and the momentum constant are also determined in the experiments to give the highest test classification accuracy. The results of ANN were compared with AIS.

2 Results
As stated in the previous section, the first application study was conducted with AIS and the optimum value of supp value was searched through changing its value in 0.01-0.9 interval to have maximum test classification accuracy. Normally, used data and system units are normalized vectors and the affinity value lie only in the [0-1] interval. So, supp value should be selected in that interval and we begun with 0.9 value and decrease it with 0.1. Then, around some values which give high classification accuracies we decreased or increased supp value more tenderly. That is, we first had a rough insight about supp values and scrutinized some good values in a more detailed way. The test classification accuracies for searched supp values are given in Figure 2.

Figure 2 The change of test classification accuracy with regard to the change in supp value

 
As shown from the figure, the maximum test classification accuracy was detected as 93.33% for 0.07, 0.18 and 0.25 values for the supp value. Because 15 data in the test procedure is used, this accuracy means that only one data is misclassified by the system. Figure 2 shows that there isn’t a specific way of determining supp value. That is, we cannot say that the accuracy decreases as supp increases or otherwise. The only way of finding best supp value is to search for supp value through experimentation.

The other application of prostate cancer classification was with ANN method. As stated in Section 2, gradient descent learning algorithm was taken for a one-hidden layered ANN. The searched parameters in this application are optimum number of hidden nodes (hn), learning rate (lr) and momentum constant (mc). In this search procedure, firstly by fixing lr and mc to values of 2 and 0.8, hn is changed between 1 and 50 by steps of 1and for each experimented hn values, the test classification accuracy was recorded. The change of test classification accuracy, according to the hn values is shown in Figure 3.

Figure 3 Change of test classification accuracy according to the changing hn number

 
As can be seen from the figure, 100% test classification accuracy was reached for some hidden node numbers like 16, 35, 40 and 45. Thus we took hn as 16. Searches for best lr and mc parameters were also done but, because the 100% was reached, these are not presented here.

In summary, ANN has reached a higher result than AIS but two points should be emphasized here:

The difference was only for one test data. That is ANN has correctly classified one more data

The number of dataset is not satisfactory to have a confidential comparison between two systems.

Anymore, both methods have well performed for this classification task and encouraged us to use these systems with much more data in real applications of daily life.

3 Discussion
Application of some artificial intelligence methods to the classification of some disease has increasing day by day. While this kind of systems cannot be used in their own in detecting abnormalities, at least they can help physician in selecting suspicious situations or in deciding his resultant decision.

In our study, we classified prostate cancer data of 50 subjects by AIS and ANN. By taking the clonal selection model in immune system, a code for AIS was written and applied to the dataset for classification. With best system parameters, AIS has reached a test classification ratio of 93.33% by misclassifying only 1 test data. On the other hand ANN has reached 100% accuracy. Whereas it is not possible to do a confidential comparison between AIS and ANN methods for 50 data, we can say that ANN can be searched for more crowded datasets for a real-life application to help doctors.

Acknowledgement
This work is supported by the Coordinatorship of Selçuk University's Scientific Research Projects Grant.

References
Abbas A.K., Lichtman A.H., and Pober J.S., 1994, Celluar and Molecular Immunology, United States of America, W. B. Sounders Com.

Ada G.L. and Nossal G.J.V, 1987, The Clonal Selection Theory, Scientific American, 257: (2), 50-57.
http://dx.doi.org/10.1038/scientificamerican0887-62

Chaudhuri K., Saha S., Azeem R., Balachandran S., Yu S., Majumdar N., and Nino F., 2007, Artificial Immune Systems: A Bibliography, Compiled by Dasgupta, D., CS Technical Report, No. CS-07-004.

Chen J., and Mahfouf, M., 2006, A Populastion Adaptive Based Immune Algorithm for Solving Multi-objective Optimization Problems, Lecture Notes in Computer Science, 4163, 280-293.
http://dx.doi.org/10.1007/11823940_22

Cutello V., Narzisi G., Nicosia G., Pavone M., “Clonal Selection Algorithms: A Comparative Case Study Using Effective Mutation Potentials”, Lecture Notes in Computer Scienvce, Vol. 3627, pp. 13-28, 2005.
http://dx.doi.org/10.1007/11536444_2

Dasgupta D., 1998, Artificial Immune Systems and Their Applications, Berlin, Springer-Verlag.

Garrett S.M., 2003, A Paratope is not an Epitope: Implications for Immune Network Models and Clonal Selection”, Lecture Notes in Computer Scienvce, 2787: 217-228.
http://dx.doi.org/10.1007/978-3-540-45192-1_21

Goddard A.D., Solomon B., 1993, “Genetic aspects of cancer”, Advances in Human Genetics, 21: 321-376.

Kopec D., Kabir, M.H., Reinharth D., Rothschild O., and Castiglione J.A., 2003, Human Errors in Medical Practice: Systematic Classification and Reduction with Automated Information Systems, J. of Medical Systems, 27: (4) 297-313.
http://dx.doi.org/10.1023/A:1023796918654

Li J., Yen C., Liaw D., Podsypanina K., Bose S.,Wang, S.I., 1997, PTEN a putative protein, Am. J. Hum. Genet. 1943-1947.

L. N. de Castro, Timmis J., 2002, Artificial Immune Systems: A New Compoutational Intelligence Approach, U.K., Springer Ed.

L. N. de Castro and Von Zuben F.J., 1999, Artificial Immune Systems: Part I- Basic Theory and Applications, Technical Report - DCA-RT 02/00.

L. N. de Castro and Von Zuben F.J., 2000, An Evolutionary Immune Network for Data Clustering”, Proc. of the IEEE Brazilian Symposium on Artificial Neural Networks, 84-89.
http://dx.doi.org/10.1109/SBRN.2000.889718

Magnusson C., Baron, J., Persson, I., Wolk, A., Bergstrom R., Trichopoulos D., Adami H.O., 1998, “Body size in different periods of life and breast cancer risk in postmenopausal women”, Int J Cancer, 76: 29-34.
http://dx.doi.org/10.1002/(SICI)1097-0215(19980330)76:1%3C29::AID-IJC6%3E3.0.CO;2-#

Marshall, C.J., 1991, Tumor suppressor genes, Cell, 64: 313-326.
http://dx.doi.org/10.1016/0092-8674(91)90641-B

Ong, Z.X., Tay J.C., Kwoh C.K., 2005, Applying the Clonal Selection Principle to Find Flexible Job-Shop Schedules, Lecture Notes in Computer Science, 3627: 442-455.
http://dx.doi.org/10.1007/11536444_34

Perelson, A.S. and Oster, G.F., 1979, “Theoretical Studies of Clonal Selection: Minimal Antibody Repertuarie Size and Reliability of Self-Nonself Discrimination”, J. Theor. Biol., 81: 645-670.
http://dx.doi.org/10.1016/0022-5193(79)90275-3

Principe, J.C., Euliano, N.R., Lefebvre, W.C., 1999, Neural and Adaptive Systems (Fundemantals Through Simulations), USA, John WILEY & SONS, Inc.

Schenck U., and Planding W., 1998, Quantitation of visual screening technique in cytology, Proc. Image Analysis in Medicine, II. National Symposium, 7-14.

Steck, P.A., Pershouse, M.A., Jasser, S.A., Yung, W.K., Lin H., Ligon, A.H., 1997, Identification of a candidate tumour suppressor gene, MMAC1, at chromosome 10q23.3 that is mutated in multiple advanced cancers, Nat. Genet., 15: 356-362.
http://dx.doi.org/10.1038/ng0497-356

Timmis J., and Neal M., 2001, A Resource Limited Artificial Immune System”, Knowledge Based Systems, 14: 121-130.
http://dx.doi.org/10.1016/S0950-7051(01)00088-0

Waite K.A., and Protean C., 2002, PTEN: form and function, Am J Hum Genet, 70: 829-844.
http://dx.doi.org/10.1086/340026