- Email: [email protected]
Fuzzy expert system for predicting pathological stage of prostate cancer
Expert Systems with Applications 40 (2013) 466–470
Contents lists available at SciVerse ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Fuzzy expert system for predicting pathological stage of prostate cancer M.J.P. Castanho a,⇑, F. Hernandes b, A.M. De Ré b, S. Rautenberg b, A. Billis c a
Department of Mathematics, Universidade Estadual do Centro-Oeste, Guarapuava, PR, Brazil Department of Computation Science, Universidade Estadual do Centro-Oeste, Guarapuava, PR, Brazil c School of Medicine, Universidade Estadual de Campinas, Campinas, SP, Brazil b
a r t i c l e
i n f o
Keywords: Fuzzy rule-based system Genetic algorithm Prostate cancer
a b s t r a c t Prostate cancer is the second most common cancer among men, responsible for the loss of half a million lives each year worldwide, according to the World Health Organization. In prostate cancer, definitive therapy such as radical prostatectomy, is more effective when the cancer is organ-confined. The aim of this study is to investigate the performance of some fuzzy expert systems in the classification of patients with confined or non-confined cancer. To deal with the intrinsic uncertainty about the variables utilized to predict cancer stage, the developed approach is based on Fuzzy Set Theory. A fuzzy expert system was developed with the fuzzy rules and membership functions tuned by a genetic algorithm. As a result, the utilized approach reached better precision taking into account some correlated studies. Ó 2012 Elsevier Ltd. All rights reserved.
1. Introduction In Brazil as well as worldwide, prostate cancer is the second most common male malignancy (INCA, 2010). Accurate estimates of the pathologic stage are essential for physicians to decide the appropriate therapy. When the cancer is confined to the prostate, the cure rate after surgery is approximately 80%. Clinical stage, serum prostate-specific antigen (PSA) concentration and Gleason score are among the most recognized factors. A combination of these three parameters leads to a score used to define prognostic groups that are routinely used in daily practice. Lately, numerous tools to aid the medical decision-making have been developed. Particularly, for predicting final pathologic stage, there are statistical approaches and computer systems based on artificial intelligence. In the literature we can find several papers about cancer using soft computing techniques, such as artificial neural networks (Han, Snow, Brandt, & Partin, 2001; Matsui et al., 2002; Saritas, Ozkan, & Sert, 2010; Snow, Smith, & Catalona, 1994), fuzzy logic (Castanho, Barros, Yamakami, & Vendite, 2008; Saritas, Allahverdi, & Sert, 2003; Seker, Odetayo, Petrovic, & Naguib, 2003), genetic algorithms (Baker & Abdul-Kareem, 2007; Ghosh, Mitchell, Tanyi, & Hung, 2010; Ludwig & Roos, 2010; Odusanya et al., 2002; Shah & Kusiak, 2007), neuro-fuzzy systems (Benecchi, 2006; Keles, Hasiloglu, Keles, & Aksoy, 2007; Papageorgiou et al., 2008) and geneticfuzzy systems (Peña-Reyes & Sipper, 1999; Sedighiani & HashemiKhabir, 2009). ⇑ Corresponding author. Address: Caixa Postal 3010, CEP 85015-430, Guarapuava, PR, Brazil. Tel.: +55 04236298348; fax: +55 04236211090. E-mail address: [email protected] (M.J.P. Castanho). 0957-4174/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.07.046
The aim of this study is to investigate the use of a genetic-fuzzy system to predict the pathological stage of prostate cancer, combining preoperative serum PSA, clinical stage, and biopsy Gleason score. In the next section, a literature review is done. Section 3 describes the development of the genetic-fuzzy system. Finally, Section 4 presents the best system found and compares its ability to discriminate prostate-confined cancer and probability tables. 2. Literature review Predictive tools are essential for individualized, evidence-based medical decision making (Shariat, Karakiewicz, & Roehrborn, 2008a). Probability tables are statistical tools to make prostate cancer prognoses. Also, some soft computing components are being used for this purpose. 2.1. Statistical tools There are numerous published probability tables available to help clinicians in the task of predicting the pathologic stage of prostate cancer. However, just some of these were validated and their usefulness is not completely defined (Ross, Scardino, & Kattan, 2001). The first one and the most widely used is the Partin tables. In 1997, Partin et al. (1997), presented probability tables combining preoperative serum prostate-specific antigen (PSA), TNM clinical stage and Gleason grade from the prostate biopsy. Multinomial log-linear regression was performed to provide the likelihood of final pathological stages as follows: organ-confined, capsular penetration, positive seminal vesicle involvement and lymph node
467
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470
involvement. These tables were updated in 2001 (Partin et al., 2001) and 2007 (Makarov et al., 2007). Augustin, Sun, Isbarn, Pummer, and Karakiewicz (2012) compared the three versions and did not identify the best one. The Partin tables represent the most widely used guide towards the selection of definitive therapies in Europe and North America (Bhojani et al., 2009). The European Association of Urology (EAU) guidelines recommend the use of Partin tables and Kattan nomogram for prostate cancer staging (Briganti, Karakiewicz, Joniau, & Van Poppel, 2009). The Partin tables were validated in several studies around the world. Some results are described in Table 1. It has been demonstrated that predictions made by probability tables are more accurate than those by clinicians, regardless of their level of expertise (Chun et al., 2006; Shariat et al., 2008a). 2.2. Soft computing methods Over the last twenty years, soft computing techniques have been applied to develop computational systems for diagnosis and prognosis in Medicine. Soft computing comprises principally artificial neural networks, fuzzy logic and genetic algorithms (Baker, Hassani, & Kareem, 2008). An artificial neural network (ANN) is a system based on the operation of biological neural networks. The use of ANNs in prostate cancer was pioneered by Snow et al. (1994). Some years later, Han et al. (2001) evaluated an ANN for the prediction of the pathologic stage in prostate carcinoma and found results slightly superior to the probability tables. To organ-confined disease the AUC was 0.77 using ANN and 0.72 with the probability tables. In 2001, Sargent (2001) carried out a literature review of 28 studies comparing ANN with standard statistical techniques (regression) and concluded that regression approaches have several desirable features in terms of ease of use and the ability to draw inferences based on their output. In addition, neither method supersedes the other in predictive performance. Matsui et al. (2002) compared ANN for predicting pathological stage of clinically localized prostate cancer in the Japanese population with the Partin tables. The AUC of the ANN with the same three parameters of the tables was 0.825 while with Partin tables it was 0.756 but did not attain statistical significance. Shariat, Karakiewicz, Suardi, and Kattan (2008b) carried out an analysis of the methods available in the literature: nomograms, risk groupings, ANN, probability tables and classification and regression tree analysis. They concluded that the probability tables have the highest accuracy and the best discriminating characteristics for predicting outcomes in prostate cancer patients. Saritas et al. (2010) developed an ANN to predict whether patients have prostate cancer or not before biopsy. The AUC was 0.94. In 2003 they developed a fuzzy expert system to diagnose prostate cancer but the results overestimate the literature (Saritas et al., 2003). Chen, Zhang, Xu, Chen, and Zhang (2012) compare the diagnostic performances of ANN and multivar-
Table 1 Performance of Partin tables validated in different population samples. Area under Receiver Operating Characteristic curve (AUC) values for organ confined disease (OC), extraprostatic extension (EPE), seminal vesicle involvement (SVI) and lymph node involvement (LN). Year
Authors
Samples
OC
EPE
SVI
LN
2005 2005 2006 2008 2008 2010 2010
Eskicorapci et al. Gorziza Ayyathurai et al. Naito et al. Heath el al. Yu et al. Fanning et al.
Turkish Brazilian Welsh Japanese African American American Irish
0.66 0.65 0.73 0.70 0.73 0.68 0.58
– 0.54 – – 0.62 0.62 0.54
0.73 0.63 0.74 – 0.77 0.77 0.66
0.76 0.77 0.78 – – 0.74 0.55
iable logistic regression analysis for differentiating between malignant and benign lung nodules on computed tomography scans. Another soft computing technique is based on fuzzy logic (FL). Fuzzy rule-based system is a mathematical tool for dealing with the uncertainty and the imprecision typical in medical field. The reasoning is based on compositional rule of fuzzy inference and the knowledge of specialists is important to determine the parameters. Seker et al. (2003) investigated the Fuzzy K-Nearest Neighbor (FK-NN) algorithm to estimate the accuracy of breast and prostate cancer prognoses and determined the significance of prognostic markers. Castanho et al. (2008) developed a fuzzy rulebased system to predict the pathologic stage of prostate cancer. For organ-confined disease the AUC was 0.76. The third soft computing component is genetic algorithm (GA), which is a methodology inspired by natural evolution. This component was used by Shah and Kusiak (2007) to identify and classify ovarian, prostate and lung cancer; Ludwig and Roos (2010) and Odusanya et al. (2002) to investigate the prognosis of breast cancer; Baker and Abdul-Kareem (2007) to introduce a new method for the prognosis of nasopharyngeal carcinoma and Ghosh et al. (2010) for prostate cancer treatment planning. Since these three methods are complementary rather than competitive, many researchers have hybridized ANNs, FL and GAs to develop a better performance system. Neuro-fuzzy systems use fuzzy systems to represent and process the knowledge in a clear way with easy interpretation and utilize the learning capacity of ANN. They were used to classify prostate cancer (Benecchi, 2006; Keles et al., 2007) and to diagnose breast cancer (Keles, Keles, & Yavuz, 2011). Karabatak and Ince (2009) presented an automatic diagnosis system for detecting breast cancer based on association rules and neural network. A genetic-fuzzy system is a rule-based fuzzy system where a genetic algorithm is used to optimize the parameters. They were used in breast cancer diagnoses (Peña-Reyes & Sipper, 1999; Sedighiani & HashemiKhabir, 2009). The importance of genetic-fuzzy systems was emphasized in Cordón, Gomide, Herrera, Hoffmann, and Magdalena (2004) and Herrera (2008). Peña-Reyes and Sipper (1999) combined fuzzy-genetic approach and attained high classification performance in breast cancer diagnosis also simplified the resulting system. In this context and believing in improving the literature results, in this study, a genetic-fuzzy system for predicting pathological stage of prostate cancer is proposed.
3. Proposed system Between January 1997 and June 2008, 331 patients ranging from 43 to 76 years old (median 64) were treated with radical prostatectomy for clinically localized prostate cancer at Clinics Hospital located at the Universidade Estadual de Campinas, Brazil. A database description is in Table 2. The surgical specimens (prostate and adjacent structures) from each patient were examined by the same pathologist and the pathologic stage established: 289 patients had organ-confined cancer (TNM, pT2) and 42 patients had extraprostatic cancer (TNM, >pT2). The pathological grade
Table 2 Classification of patients database. PSA
Patients
Gleason score
Patients
Clinical stage
Patients
0–2.5 2.6–4 4.1–6 6.1–8 8.1–10 >10
15 23 72 63 49 109
4 5–6 7¼3þ4 7¼4þ3 8–10
3 211 84 17 16
T1a T1b T1c T2a T2b T2c
9 4 149 118 41 10
468
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470
followed the criteria by the 2005 Consensus Conference (Epstein, Allsbrook, Amin, & Egevad, 2005). To find a system that classifies the patients into two classes: organ-confined or non-confined cancer, a genetic algorithm was developed. In this algorithm each chromosome randomly generated comprises a fuzzy rule based system. Its structure is shown in Fig. 1. During the execution of the system, trapezoidal membership functions are considered and the selection is done by tournament. The fitness is calculated using the sum of two performance measures: sensitivity (probability that the test correctly classifies a patient with non-confined cancer) and specificity (probability of correctly classifying a patient with organ-confined cancer).
Pn
i¼1 TPðxi Þ Pn TPðx iÞ þ i¼1 i¼1 FNðxi Þ Pn i¼1 TNðxi Þ Pn SpecificityðkÞ ¼ Pn i¼1 TNðxi Þ þ i¼1 FPðxi Þ
Sensitiv ityðkÞ ¼ Pn
ð1Þ ð2Þ
where TP is the set True Positive, patients with non-confined cancer classified correctly; FN is the set False Negative, patients with nonconfined cancer classified as confined; TN is the set True Negative, patients with confined cancer correctly classified; FP is the set False Positive, patients with confined cancer classified as non-confined; xi is the test result for the ith individual submitted to it, classified as Positive when xi P k (k is cutoff point) and otherwise as Negative. Each chromosome fitness is obtained by evaluating all patients on basis of the database. The mutation operator was implemented to handle a fuzzy rule or a fuzzy variable. Randomly, when a fuzzy variable is chosen, the mutation occurs in a vertex of the membership function. When a fuzzy rule is selected, the change occurs in the consequent of the rule. The crossover operator influences the fuzzy system in two ways: either on the fuzzy variables or on the rules. In the fuzzy variable crossover, two fuzzy systems are randomly chosen and one of the fuzzy variables (PSA, Gleason score or clinical stage) is selected. Then all membership functions of the selected variable are exchanged between the systems. For the rules, a random number of them is selected to be exchanged between systems. The stopping criterion of the proposed algorithm is the number of generations.
4. Results and discussion The genetic-fuzzy algorithm was implemented using JAVA language. With the aim of finding the best rates of mutation and crossover, the number of membership functions of the input variables (PSA/Gleason/clinical stage) was fixed in 5/3/5. These numbers were chosen after preliminary tests. The crossover and mutation parameters were tested from 0.05 to 0.05 until 0.5. Each variation was executed 100 times. The best results were obtained in the 0.2/ 0.2 combination. This combination was used to find the best membership function number for each input variable. In Table 3 the average values obtained for each variables configuration in 100 tests are presented. The configuration 6/3/4 for the membership functions (PSA/ Gleason/clinical stage) achieved the highest fitness average and the best individual. This individual has fitness equal to 1.677, mutation rate 0.2 and crossover rate equal to 0.2. It is described below. The input variables of the system are PSA level, Gleason score and clinical stage. The serum PSA level is considered Normal (N), Slightly Elevated (SE), Moderately Elevated (ME), Very Elevated (VE), Highly Elevated (HE) or Extremely Elevated (EE). The linguistic variable Gleason score is classified with the following linguistic labels: Well-Differentiated (WD), tumor with less aggressive behavior; Moderately Differentiated (MD) and Poorly Differentiated (PD), with more aggressive behavior. Regarding the linguistic variable clinical stage, labels are given according to the TNM Classification System (Sobin & Wittekind, 1997). The four membership functions clinical stage were cT1 (witch includes cT1a, cT1b, cT1c), cT2a, cT2b and cT2c. From stage T3a on, the tumor is beyond prostate and advanced stages are not considered in this model. The membership function of input variables are represented in Fig. 2. As for the output variable, cancer stage, the following labels are attributed: confined, if all cancer is confined within the prostate and non-confined, if cancer is evidently outside the prostate. The rule base is composed of 47 rules like the following: ‘‘If PSA Level is Moderately Elevated (ME) and Gleason score is Poorly Differentiated (PD) and clinical stage is T2b, then cancer stage is non-confined’’.
Fig. 1. Chromosome representation.
469
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470 Table 3 Some test results with different configurations and mutation/crossover rates equal to 0.2/0.2. PSA/Gleason/stage
Fitness
Sensitivity
Specificity
4/3/3 4/3/4 4/3/5 4/4/3 4/4/4 4/4/5 5/3/3 5/3/4 5/3/5 5/4/3 5/4/4 5/4/5 6/3/3 6/3/4 6/3/5 6/4/3 6/4/4 6/4/5
1.584669 1.596237 1.598746 1.586132 1.600488 1.599233 1.633449 1.657003 1.643206 1.654433 1.649826 1.661603 1.665157 1.666899 1.663275 1.658467 1.658258 1.657213
0.660859 0.664808 0.693031 0.685828 0.680488 0.702091 0.690592 0.705575 0.700348 0.695703 0.706969 0.707317 0.722300 0.709756 0.726132 0.718467 0.726829 0.725784
0.923810 0.931429 0.905714 0.902857 0.920000 0.897143 0.942857 0.951429 0.942857 0.958730 0.942857 0.954286 0.942857 0.957143 0.937143 0.940000 0.931429 0.931429
Table 4 Classification of patients validation database. PSA
Patients
Gleason score
Patients
Clinical stage
Patients
0–4 4.1–6 6.1–8 8.1–10 >10
3 11 9 7 18
4 5–6 7¼3þ4 7¼4þ3 8–10
– 34 9 3 2
T1 T2a T2b T2c
36 8 4 –
The bold values indicate the configuration with the highest average fitness and contains the best individual.
N
1
SE ME VE HE
EE
0.8 0.6 0.4 0.2 0 0
5
10
15
20
Degree of membership
T2b
T2a
To validate the system the database described in Table 4 was used. It contains data of 48 patients with average age of 63, treated with radical prostatectomy at the same hospital between July 2008 and November 2011. Among these patients, 34 had organ-confined disease. The ROC curve displays the sensitivity of a diagnostic test over all possible false-positive rate (1-specificity). A diagnostic test with perfect discrimination has an area under the ROC curve equal to 1 and when there is no discrimination the area is 0.5. In this case, the area under the curve shown in Fig. 3 is 0.693 (95% CI: 0.544–0.818) to Partin probability tables and is 0.824 (95% CI: 0.686–0.918) to
MD
1WD
0.6 0.4 0.2 0 5
Gleason Score T2c
0.8 0.6 0.4 0.2 0 0
1
2
PD
0.8
0
PSA Level 1 T1
Fig. 3. ROC curves to compare Partin tables and genetic-fuzzy system.
Degree of membership
Degree of membership
The inference process was done through zero-order Sugeno method (Sugeno, 1985). The ensemble of fuzzy rules and the fuzzy inference method perform the role of a mathematical function to obtain the system output. This output is a real number. In this way, for each patient whose data are informed as system inputs, a number in the range ½0; 1 points out that the closer to 1 the bigger the involvement of prostate adjacent structures. With the aim of evaluating the performance of this genetic-fuzzy system in discriminating between organ-confined and non-confined prostate cancer, the Receiver Operating Characteristic (ROC) curve (Zweig & Campbell, 1993) was constructed. It is also used to identify a decision threshold (cut-off point) that corresponds to the best combination of sensitivity and specificity. Also, the performance of the developed system and Partin tables (Makarov et al., 2007) were compared using ROC curves.
3
Clinical Stage Fig. 2. Membership functions refer to input variables.
10
470
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470
the genetic-fuzzy system developed. The statistical p value for the comparison of these two curves is 0.045 (P < 0.05). The decision threshold is 0.32 and 0.26, respectively. For the database used in this study, the 2007 version of Partin tables performance is similar to those described in Table 1. This shows that it is in accordance with the other studies. 5. Conclusion In this paper, the use of soft computing components to generate a genetic-fuzzy system was investigated to predict the pathologic stage of prostate cancer. The approach is based on a genetic algorithm to tune both the fuzzy rules and fuzzy sets. The genetic algorithm was effective in adjusting parameters of the fuzzy system. The resulted fuzzy system has better performance than Partin probability tables for the database used. In this way, it is possible to conclude that the use of soft computing techniques can produce good results for prostate cancer prognoses. As future study the application of this approach to other database is proposed. Furthermore, other artificial intelligence techniques can be associated with the developed system. References Augustin, H., Sun, M., Isbarn, H., Pummer, K., Karakiewicz, P. (2012). Decision curve analysis to compare 3 versions of Partin tables to predict final pathological stage. Urologic Oncology: Seminars and Original Investigations 30(4), 396–401. Baker, O., Abdul-Kareem, S. (2007). Using genetic algorithm to evolves algebraic rule-based classifiers for NPC prognosis. In International conference on intelligent and advanced systems (ICIAS) (pp. 71–74). Baker, O. F., Hassani, A., & Kareem, S. (2008). The use of soft computing approaches FL models for medical prognosis NPC. In Proceedings of the 10th international conference on information integration and web-based applications & services (pp. 706–709). New York, NY, USA: ACM. Benecchi, L. (2006). Neuro-fuzzy system for prostate cancer diagnosis. Urology, 68, 357–361. Bhojani, N., Ahyai, S., Graefen, M., Capitanio, U., Suardi, N., Shariat, S. F., et al. (2009). Partin tables cannot accurately predict the pathological stage at radical prostatectomy. European Journal of Surgical Oncology, 35(2), 123–128. Briganti, A., Karakiewicz, P., Joniau, S., & Van Poppel, H. (2009). The motion: Nomograms should become a routine tool in determining prostate cancer prognosis. European Urology, 55, 743–747. Castanho, M. J. P. C., Barros, L. C., Yamakami, A., & Vendite, L. L. (2008). Fuzzy expert system: an example in prostate cancer. Applied Mathematics and Computation, 202, 78–85. Chen, H., Zhang, J., Xu, Y., Chen, B., & Zhang, K. (2012). Performance comparison of artificial neural network and logistic regression model for differentiating lung nodules on ct scans. Expert Systems with Applications, 39(13), 11503–11509. Chun, F. K. H., Karakiewicz, P., Briganti, A., Gallina, A., Kattan, M., Montorsi, F., et al. (2006). Prostate cancer nomograms: an update. European Urology, 50, 914–926. Cordón, O., Gomide, F., Herrera, F., Hoffmann, F., & Magdalena, L. (2004). Ten years of genetic fuzzy systems: current framework and new trends. Fuzzy Sets and Systems, 141, 5–31. Epstein, J. I., Allsbrook, W. C. J., Amin, M. B., & Egevad, L. L.ISUP Grading Committee. (2005). The 2005 International Society of Urological Pathology (ISUP) consensus conference on Gleason grading of prostatic carcinoma. American Journal of Surgical Pathology, 29(9), 1228–1242. Ghosh, P., Mitchell, M., Tanyi, J., Hung, A. (2010). Automatic segmentation of the prostate using a genetic algorithm for prostate cancer treatment planning. In Ninth international conference on machine learning and applications (ICMLA) (pp. 752 – 7570). Han, M., Snow, P., Brandt, J. M., & Partin, A. (2001). Evaluation of artificial neural networks for the prediction of pathologic stage in prostate carcinoma. Cancer Supplement, 1661–1666. Herrera, F. (2008). Genetic fuzzy systems: taxonomy, current research trends and prospects. Evolutionary Intelligence, 1, 27–46.
INCA. (2010). Instituto Nacional de Câncer - Ministério da Saúde, November 2010. Available from www.inca.gov.br (In Portuguese). Karabatak, M., & Ince, C. (2009). An expert system for detection of breast cancer based on association rules and neural network. Expert Systems with Applications, 36, 3465–3469. Keles, A., Hasiloglu, S. A., Keles, A., & Aksoy, Y. (2007). Neuro-fuzzy classification of prostate cancer using NEFCLASS-J. Computers in Biology and Medicine, 37, 1617–1628. Available from http://portal.acm.org/citation.cfm?id=1287847. 1288045. Keles, A., Keles, A., & Yavuz, U. (2011). Expert system based on neuro-fuzzy rules for diagnosis breast cancer. Expert Systems with Applications, 38(5), 5719–5726. Ludwig, S., & Roos, S. (2010). Prognosis of breast cancer using genetic programming. In R. Setchi, I. Jordanov, R. Howlett, & L. Jain (Eds.), Knowledge-based and intelligent information and engineering systems. Lecture notes in computer science (Vol. 6279, pp. 536–545). Berlin/Heidelberg: Springer. Makarov, D., Trock, B., Humphreys, E., Mangold, L. A., Walsh, P. C., Epstein, J. I., et al. (2007). Update nomogram to predict pathologic stage of prostate cancer given prostate-specific antigen level, clinical stage, and biopsy Gleason score (Partin Tables) based on cases from 2000 to 2005. Urology, 69(6), 1095–1101. Matsui, Y., Egawa, S., Tsukayama, C., Terai, A., Kuwao, S., Baba, S., et al. (2002). Artificial neural network analisys for predicting pathological stage of clinically localized prostate cancer in the Japanese population. Japanese Journal of Clinical Oncology, 32(12), 530–535. Odusanya, A., Odetayo, M., Petrovic, D., Naguib, R., Lakshmi, M., & Sherbet, G. (2002). A genetic algorithm-based model for breast cancer prognosis. In Proceedings of the 6th world multiconference on systemics, cybernetics and informatics (SCI2002) (pp. 394–397). Papageorgiou, E., Spyridonos, P., Glotsos, D., Stylios, C., Ravazoulad, P., Nikiforidis, G., et al. (2008). Brain tumor characterization using the softcomputing technique of fuzzy cognitive maps. Applied Soft Computing, 8(1), 820–828. Partin, A., Kattan, M., Subong, E., Walsh, P., Wojno, K., Oesterling, J., et al. (1997). Combination of prostate-specific antigen, clinical stage, and gleason score to predict pathological stage of localized prostate cancer. Journal of the American Medical Association, 277(18), 1445–1451. Partin, A., Mangold, L., Lamm, D., Walsh, P., Epstein, J. I., & Pearson, J. (2001). Contemporary update of prostate cancer staging nomograms Partin Tables for the new millennium. Urology, 58(6), 843–848. Peña-Reyes, C. A., & Sipper, M. (1999). A fuzzy-genetic approach to breast cancer diagnosis. Artificial Intelligence in Medicine, 17, 131–155. Ross, P., Scardino, P., & Kattan, M. (2001). A catalog of prostate cancer nomograms. Journal of Urology, 162, 1562–1568. Sargent, D. (2001). Comparison of artificial neural networks with other statistical approaches. Cancer, 91(58), 1636–1642. Saritas, I., Allahverdi, N., & Sert, I. U. (2003). A fuzzy expert system design for diagnosis of prostate cancer. In Proceedings of the 4th international conference on computer systems and technologies: e-learning (pp. 345–351). New York, NY, USA: ACM. Saritas, I., Ozkan, I. A., & Sert, I. U. (2010). Prognosis of prostate cancer by artificial neural networks. Expert Systems with Applications, 37, 6646–6650. Sedighiani, K., & HashemiKhabir, S. (2009). Constructing interpretable genetic fuzzy rule-based system for breast cancer diagnostic. Proceedings of the 2009 WRI global congress on intelligent systems (Vol. 01, pp. 441–446). Washington, DC, USA: IEEE Computer Society. Available from http://portal.acm.org/ citation.cfm?id=1603895.1604812. Seker, H., Odetayo, M., Petrovic, D., & Naguib, R. N. (2003). A fuzzy logic basedmethod for prognostic decision making in breast and prostate cancers. IEEE Transactions on Information Technology in Biomedicine, 7(2), 114–122. Shah, S., & Kusiak, A. (2007). Cancer gene search with data-mining and genetic algorithms. Computers in Biology and Medicine, 37, 251–261. Shariat, S., Karakiewicz, P. I., & Roehrborn, C. M. W. K. (2008a). An updated catalog of prostate cancer predictive tools. Cancer, 113(11), 3075–3099. Shariat, S., Karakiewicz, P. I., Suardi, N., & Kattan, M. (2008b). Comparison of nomograms with other methods for predicting outcomes in prostate cancer: a critical analysis of the literature. Clinical Cancer Research, 14(14), 4400–4407. Snow, P., Smith, D., & Catalona, W. (1994). Artificial neural networks in the diagnosis and prognosis of prostate cancer: a pilot study. Journal of Urology, 152, 1923–1926. Sobin, L. H., & Wittekind, C. (Eds.). (1997). TNM classification of malignant tumors (5th ed.. International union against cancer (UICC) (5th ed.). New York: John Wiley & Sons, Inc.. Sugeno, M. (1985). Industrial applications of fuzzy control. Elsevier Science Pub. Co.. Zweig, M. H., & Campbell, G. (1993). Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clinical Chemistry, 39(4), 561–577.
Contents lists available at SciVerse ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Fuzzy expert system for predicting pathological stage of prostate cancer M.J.P. Castanho a,⇑, F. Hernandes b, A.M. De Ré b, S. Rautenberg b, A. Billis c a
Department of Mathematics, Universidade Estadual do Centro-Oeste, Guarapuava, PR, Brazil Department of Computation Science, Universidade Estadual do Centro-Oeste, Guarapuava, PR, Brazil c School of Medicine, Universidade Estadual de Campinas, Campinas, SP, Brazil b
a r t i c l e
i n f o
Keywords: Fuzzy rule-based system Genetic algorithm Prostate cancer
a b s t r a c t Prostate cancer is the second most common cancer among men, responsible for the loss of half a million lives each year worldwide, according to the World Health Organization. In prostate cancer, definitive therapy such as radical prostatectomy, is more effective when the cancer is organ-confined. The aim of this study is to investigate the performance of some fuzzy expert systems in the classification of patients with confined or non-confined cancer. To deal with the intrinsic uncertainty about the variables utilized to predict cancer stage, the developed approach is based on Fuzzy Set Theory. A fuzzy expert system was developed with the fuzzy rules and membership functions tuned by a genetic algorithm. As a result, the utilized approach reached better precision taking into account some correlated studies. Ó 2012 Elsevier Ltd. All rights reserved.
1. Introduction In Brazil as well as worldwide, prostate cancer is the second most common male malignancy (INCA, 2010). Accurate estimates of the pathologic stage are essential for physicians to decide the appropriate therapy. When the cancer is confined to the prostate, the cure rate after surgery is approximately 80%. Clinical stage, serum prostate-specific antigen (PSA) concentration and Gleason score are among the most recognized factors. A combination of these three parameters leads to a score used to define prognostic groups that are routinely used in daily practice. Lately, numerous tools to aid the medical decision-making have been developed. Particularly, for predicting final pathologic stage, there are statistical approaches and computer systems based on artificial intelligence. In the literature we can find several papers about cancer using soft computing techniques, such as artificial neural networks (Han, Snow, Brandt, & Partin, 2001; Matsui et al., 2002; Saritas, Ozkan, & Sert, 2010; Snow, Smith, & Catalona, 1994), fuzzy logic (Castanho, Barros, Yamakami, & Vendite, 2008; Saritas, Allahverdi, & Sert, 2003; Seker, Odetayo, Petrovic, & Naguib, 2003), genetic algorithms (Baker & Abdul-Kareem, 2007; Ghosh, Mitchell, Tanyi, & Hung, 2010; Ludwig & Roos, 2010; Odusanya et al., 2002; Shah & Kusiak, 2007), neuro-fuzzy systems (Benecchi, 2006; Keles, Hasiloglu, Keles, & Aksoy, 2007; Papageorgiou et al., 2008) and geneticfuzzy systems (Peña-Reyes & Sipper, 1999; Sedighiani & HashemiKhabir, 2009). ⇑ Corresponding author. Address: Caixa Postal 3010, CEP 85015-430, Guarapuava, PR, Brazil. Tel.: +55 04236298348; fax: +55 04236211090. E-mail address: [email protected] (M.J.P. Castanho). 0957-4174/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2012.07.046
The aim of this study is to investigate the use of a genetic-fuzzy system to predict the pathological stage of prostate cancer, combining preoperative serum PSA, clinical stage, and biopsy Gleason score. In the next section, a literature review is done. Section 3 describes the development of the genetic-fuzzy system. Finally, Section 4 presents the best system found and compares its ability to discriminate prostate-confined cancer and probability tables. 2. Literature review Predictive tools are essential for individualized, evidence-based medical decision making (Shariat, Karakiewicz, & Roehrborn, 2008a). Probability tables are statistical tools to make prostate cancer prognoses. Also, some soft computing components are being used for this purpose. 2.1. Statistical tools There are numerous published probability tables available to help clinicians in the task of predicting the pathologic stage of prostate cancer. However, just some of these were validated and their usefulness is not completely defined (Ross, Scardino, & Kattan, 2001). The first one and the most widely used is the Partin tables. In 1997, Partin et al. (1997), presented probability tables combining preoperative serum prostate-specific antigen (PSA), TNM clinical stage and Gleason grade from the prostate biopsy. Multinomial log-linear regression was performed to provide the likelihood of final pathological stages as follows: organ-confined, capsular penetration, positive seminal vesicle involvement and lymph node
467
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470
involvement. These tables were updated in 2001 (Partin et al., 2001) and 2007 (Makarov et al., 2007). Augustin, Sun, Isbarn, Pummer, and Karakiewicz (2012) compared the three versions and did not identify the best one. The Partin tables represent the most widely used guide towards the selection of definitive therapies in Europe and North America (Bhojani et al., 2009). The European Association of Urology (EAU) guidelines recommend the use of Partin tables and Kattan nomogram for prostate cancer staging (Briganti, Karakiewicz, Joniau, & Van Poppel, 2009). The Partin tables were validated in several studies around the world. Some results are described in Table 1. It has been demonstrated that predictions made by probability tables are more accurate than those by clinicians, regardless of their level of expertise (Chun et al., 2006; Shariat et al., 2008a). 2.2. Soft computing methods Over the last twenty years, soft computing techniques have been applied to develop computational systems for diagnosis and prognosis in Medicine. Soft computing comprises principally artificial neural networks, fuzzy logic and genetic algorithms (Baker, Hassani, & Kareem, 2008). An artificial neural network (ANN) is a system based on the operation of biological neural networks. The use of ANNs in prostate cancer was pioneered by Snow et al. (1994). Some years later, Han et al. (2001) evaluated an ANN for the prediction of the pathologic stage in prostate carcinoma and found results slightly superior to the probability tables. To organ-confined disease the AUC was 0.77 using ANN and 0.72 with the probability tables. In 2001, Sargent (2001) carried out a literature review of 28 studies comparing ANN with standard statistical techniques (regression) and concluded that regression approaches have several desirable features in terms of ease of use and the ability to draw inferences based on their output. In addition, neither method supersedes the other in predictive performance. Matsui et al. (2002) compared ANN for predicting pathological stage of clinically localized prostate cancer in the Japanese population with the Partin tables. The AUC of the ANN with the same three parameters of the tables was 0.825 while with Partin tables it was 0.756 but did not attain statistical significance. Shariat, Karakiewicz, Suardi, and Kattan (2008b) carried out an analysis of the methods available in the literature: nomograms, risk groupings, ANN, probability tables and classification and regression tree analysis. They concluded that the probability tables have the highest accuracy and the best discriminating characteristics for predicting outcomes in prostate cancer patients. Saritas et al. (2010) developed an ANN to predict whether patients have prostate cancer or not before biopsy. The AUC was 0.94. In 2003 they developed a fuzzy expert system to diagnose prostate cancer but the results overestimate the literature (Saritas et al., 2003). Chen, Zhang, Xu, Chen, and Zhang (2012) compare the diagnostic performances of ANN and multivar-
Table 1 Performance of Partin tables validated in different population samples. Area under Receiver Operating Characteristic curve (AUC) values for organ confined disease (OC), extraprostatic extension (EPE), seminal vesicle involvement (SVI) and lymph node involvement (LN). Year
Authors
Samples
OC
EPE
SVI
LN
2005 2005 2006 2008 2008 2010 2010
Eskicorapci et al. Gorziza Ayyathurai et al. Naito et al. Heath el al. Yu et al. Fanning et al.
Turkish Brazilian Welsh Japanese African American American Irish
0.66 0.65 0.73 0.70 0.73 0.68 0.58
– 0.54 – – 0.62 0.62 0.54
0.73 0.63 0.74 – 0.77 0.77 0.66
0.76 0.77 0.78 – – 0.74 0.55
iable logistic regression analysis for differentiating between malignant and benign lung nodules on computed tomography scans. Another soft computing technique is based on fuzzy logic (FL). Fuzzy rule-based system is a mathematical tool for dealing with the uncertainty and the imprecision typical in medical field. The reasoning is based on compositional rule of fuzzy inference and the knowledge of specialists is important to determine the parameters. Seker et al. (2003) investigated the Fuzzy K-Nearest Neighbor (FK-NN) algorithm to estimate the accuracy of breast and prostate cancer prognoses and determined the significance of prognostic markers. Castanho et al. (2008) developed a fuzzy rulebased system to predict the pathologic stage of prostate cancer. For organ-confined disease the AUC was 0.76. The third soft computing component is genetic algorithm (GA), which is a methodology inspired by natural evolution. This component was used by Shah and Kusiak (2007) to identify and classify ovarian, prostate and lung cancer; Ludwig and Roos (2010) and Odusanya et al. (2002) to investigate the prognosis of breast cancer; Baker and Abdul-Kareem (2007) to introduce a new method for the prognosis of nasopharyngeal carcinoma and Ghosh et al. (2010) for prostate cancer treatment planning. Since these three methods are complementary rather than competitive, many researchers have hybridized ANNs, FL and GAs to develop a better performance system. Neuro-fuzzy systems use fuzzy systems to represent and process the knowledge in a clear way with easy interpretation and utilize the learning capacity of ANN. They were used to classify prostate cancer (Benecchi, 2006; Keles et al., 2007) and to diagnose breast cancer (Keles, Keles, & Yavuz, 2011). Karabatak and Ince (2009) presented an automatic diagnosis system for detecting breast cancer based on association rules and neural network. A genetic-fuzzy system is a rule-based fuzzy system where a genetic algorithm is used to optimize the parameters. They were used in breast cancer diagnoses (Peña-Reyes & Sipper, 1999; Sedighiani & HashemiKhabir, 2009). The importance of genetic-fuzzy systems was emphasized in Cordón, Gomide, Herrera, Hoffmann, and Magdalena (2004) and Herrera (2008). Peña-Reyes and Sipper (1999) combined fuzzy-genetic approach and attained high classification performance in breast cancer diagnosis also simplified the resulting system. In this context and believing in improving the literature results, in this study, a genetic-fuzzy system for predicting pathological stage of prostate cancer is proposed.
3. Proposed system Between January 1997 and June 2008, 331 patients ranging from 43 to 76 years old (median 64) were treated with radical prostatectomy for clinically localized prostate cancer at Clinics Hospital located at the Universidade Estadual de Campinas, Brazil. A database description is in Table 2. The surgical specimens (prostate and adjacent structures) from each patient were examined by the same pathologist and the pathologic stage established: 289 patients had organ-confined cancer (TNM, pT2) and 42 patients had extraprostatic cancer (TNM, >pT2). The pathological grade
Table 2 Classification of patients database. PSA
Patients
Gleason score
Patients
Clinical stage
Patients
0–2.5 2.6–4 4.1–6 6.1–8 8.1–10 >10
15 23 72 63 49 109
4 5–6 7¼3þ4 7¼4þ3 8–10
3 211 84 17 16
T1a T1b T1c T2a T2b T2c
9 4 149 118 41 10
468
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470
followed the criteria by the 2005 Consensus Conference (Epstein, Allsbrook, Amin, & Egevad, 2005). To find a system that classifies the patients into two classes: organ-confined or non-confined cancer, a genetic algorithm was developed. In this algorithm each chromosome randomly generated comprises a fuzzy rule based system. Its structure is shown in Fig. 1. During the execution of the system, trapezoidal membership functions are considered and the selection is done by tournament. The fitness is calculated using the sum of two performance measures: sensitivity (probability that the test correctly classifies a patient with non-confined cancer) and specificity (probability of correctly classifying a patient with organ-confined cancer).
Pn
i¼1 TPðxi Þ Pn TPðx iÞ þ i¼1 i¼1 FNðxi Þ Pn i¼1 TNðxi Þ Pn SpecificityðkÞ ¼ Pn i¼1 TNðxi Þ þ i¼1 FPðxi Þ
Sensitiv ityðkÞ ¼ Pn
ð1Þ ð2Þ
where TP is the set True Positive, patients with non-confined cancer classified correctly; FN is the set False Negative, patients with nonconfined cancer classified as confined; TN is the set True Negative, patients with confined cancer correctly classified; FP is the set False Positive, patients with confined cancer classified as non-confined; xi is the test result for the ith individual submitted to it, classified as Positive when xi P k (k is cutoff point) and otherwise as Negative. Each chromosome fitness is obtained by evaluating all patients on basis of the database. The mutation operator was implemented to handle a fuzzy rule or a fuzzy variable. Randomly, when a fuzzy variable is chosen, the mutation occurs in a vertex of the membership function. When a fuzzy rule is selected, the change occurs in the consequent of the rule. The crossover operator influences the fuzzy system in two ways: either on the fuzzy variables or on the rules. In the fuzzy variable crossover, two fuzzy systems are randomly chosen and one of the fuzzy variables (PSA, Gleason score or clinical stage) is selected. Then all membership functions of the selected variable are exchanged between the systems. For the rules, a random number of them is selected to be exchanged between systems. The stopping criterion of the proposed algorithm is the number of generations.
4. Results and discussion The genetic-fuzzy algorithm was implemented using JAVA language. With the aim of finding the best rates of mutation and crossover, the number of membership functions of the input variables (PSA/Gleason/clinical stage) was fixed in 5/3/5. These numbers were chosen after preliminary tests. The crossover and mutation parameters were tested from 0.05 to 0.05 until 0.5. Each variation was executed 100 times. The best results were obtained in the 0.2/ 0.2 combination. This combination was used to find the best membership function number for each input variable. In Table 3 the average values obtained for each variables configuration in 100 tests are presented. The configuration 6/3/4 for the membership functions (PSA/ Gleason/clinical stage) achieved the highest fitness average and the best individual. This individual has fitness equal to 1.677, mutation rate 0.2 and crossover rate equal to 0.2. It is described below. The input variables of the system are PSA level, Gleason score and clinical stage. The serum PSA level is considered Normal (N), Slightly Elevated (SE), Moderately Elevated (ME), Very Elevated (VE), Highly Elevated (HE) or Extremely Elevated (EE). The linguistic variable Gleason score is classified with the following linguistic labels: Well-Differentiated (WD), tumor with less aggressive behavior; Moderately Differentiated (MD) and Poorly Differentiated (PD), with more aggressive behavior. Regarding the linguistic variable clinical stage, labels are given according to the TNM Classification System (Sobin & Wittekind, 1997). The four membership functions clinical stage were cT1 (witch includes cT1a, cT1b, cT1c), cT2a, cT2b and cT2c. From stage T3a on, the tumor is beyond prostate and advanced stages are not considered in this model. The membership function of input variables are represented in Fig. 2. As for the output variable, cancer stage, the following labels are attributed: confined, if all cancer is confined within the prostate and non-confined, if cancer is evidently outside the prostate. The rule base is composed of 47 rules like the following: ‘‘If PSA Level is Moderately Elevated (ME) and Gleason score is Poorly Differentiated (PD) and clinical stage is T2b, then cancer stage is non-confined’’.
Fig. 1. Chromosome representation.
469
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470 Table 3 Some test results with different configurations and mutation/crossover rates equal to 0.2/0.2. PSA/Gleason/stage
Fitness
Sensitivity
Specificity
4/3/3 4/3/4 4/3/5 4/4/3 4/4/4 4/4/5 5/3/3 5/3/4 5/3/5 5/4/3 5/4/4 5/4/5 6/3/3 6/3/4 6/3/5 6/4/3 6/4/4 6/4/5
1.584669 1.596237 1.598746 1.586132 1.600488 1.599233 1.633449 1.657003 1.643206 1.654433 1.649826 1.661603 1.665157 1.666899 1.663275 1.658467 1.658258 1.657213
0.660859 0.664808 0.693031 0.685828 0.680488 0.702091 0.690592 0.705575 0.700348 0.695703 0.706969 0.707317 0.722300 0.709756 0.726132 0.718467 0.726829 0.725784
0.923810 0.931429 0.905714 0.902857 0.920000 0.897143 0.942857 0.951429 0.942857 0.958730 0.942857 0.954286 0.942857 0.957143 0.937143 0.940000 0.931429 0.931429
Table 4 Classification of patients validation database. PSA
Patients
Gleason score
Patients
Clinical stage
Patients
0–4 4.1–6 6.1–8 8.1–10 >10
3 11 9 7 18
4 5–6 7¼3þ4 7¼4þ3 8–10
– 34 9 3 2
T1 T2a T2b T2c
36 8 4 –
The bold values indicate the configuration with the highest average fitness and contains the best individual.
N
1
SE ME VE HE
EE
0.8 0.6 0.4 0.2 0 0
5
10
15
20
Degree of membership
T2b
T2a
To validate the system the database described in Table 4 was used. It contains data of 48 patients with average age of 63, treated with radical prostatectomy at the same hospital between July 2008 and November 2011. Among these patients, 34 had organ-confined disease. The ROC curve displays the sensitivity of a diagnostic test over all possible false-positive rate (1-specificity). A diagnostic test with perfect discrimination has an area under the ROC curve equal to 1 and when there is no discrimination the area is 0.5. In this case, the area under the curve shown in Fig. 3 is 0.693 (95% CI: 0.544–0.818) to Partin probability tables and is 0.824 (95% CI: 0.686–0.918) to
MD
1WD
0.6 0.4 0.2 0 5
Gleason Score T2c
0.8 0.6 0.4 0.2 0 0
1
2
PD
0.8
0
PSA Level 1 T1
Fig. 3. ROC curves to compare Partin tables and genetic-fuzzy system.
Degree of membership
Degree of membership
The inference process was done through zero-order Sugeno method (Sugeno, 1985). The ensemble of fuzzy rules and the fuzzy inference method perform the role of a mathematical function to obtain the system output. This output is a real number. In this way, for each patient whose data are informed as system inputs, a number in the range ½0; 1 points out that the closer to 1 the bigger the involvement of prostate adjacent structures. With the aim of evaluating the performance of this genetic-fuzzy system in discriminating between organ-confined and non-confined prostate cancer, the Receiver Operating Characteristic (ROC) curve (Zweig & Campbell, 1993) was constructed. It is also used to identify a decision threshold (cut-off point) that corresponds to the best combination of sensitivity and specificity. Also, the performance of the developed system and Partin tables (Makarov et al., 2007) were compared using ROC curves.
3
Clinical Stage Fig. 2. Membership functions refer to input variables.
10
470
M.J.P. Castanho et al. / Expert Systems with Applications 40 (2013) 466–470
the genetic-fuzzy system developed. The statistical p value for the comparison of these two curves is 0.045 (P < 0.05). The decision threshold is 0.32 and 0.26, respectively. For the database used in this study, the 2007 version of Partin tables performance is similar to those described in Table 1. This shows that it is in accordance with the other studies. 5. Conclusion In this paper, the use of soft computing components to generate a genetic-fuzzy system was investigated to predict the pathologic stage of prostate cancer. The approach is based on a genetic algorithm to tune both the fuzzy rules and fuzzy sets. The genetic algorithm was effective in adjusting parameters of the fuzzy system. The resulted fuzzy system has better performance than Partin probability tables for the database used. In this way, it is possible to conclude that the use of soft computing techniques can produce good results for prostate cancer prognoses. As future study the application of this approach to other database is proposed. Furthermore, other artificial intelligence techniques can be associated with the developed system. References Augustin, H., Sun, M., Isbarn, H., Pummer, K., Karakiewicz, P. (2012). Decision curve analysis to compare 3 versions of Partin tables to predict final pathological stage. Urologic Oncology: Seminars and Original Investigations 30(4), 396–401. Baker, O., Abdul-Kareem, S. (2007). Using genetic algorithm to evolves algebraic rule-based classifiers for NPC prognosis. In International conference on intelligent and advanced systems (ICIAS) (pp. 71–74). Baker, O. F., Hassani, A., & Kareem, S. (2008). The use of soft computing approaches FL models for medical prognosis NPC. In Proceedings of the 10th international conference on information integration and web-based applications & services (pp. 706–709). New York, NY, USA: ACM. Benecchi, L. (2006). Neuro-fuzzy system for prostate cancer diagnosis. Urology, 68, 357–361. Bhojani, N., Ahyai, S., Graefen, M., Capitanio, U., Suardi, N., Shariat, S. F., et al. (2009). Partin tables cannot accurately predict the pathological stage at radical prostatectomy. European Journal of Surgical Oncology, 35(2), 123–128. Briganti, A., Karakiewicz, P., Joniau, S., & Van Poppel, H. (2009). The motion: Nomograms should become a routine tool in determining prostate cancer prognosis. European Urology, 55, 743–747. Castanho, M. J. P. C., Barros, L. C., Yamakami, A., & Vendite, L. L. (2008). Fuzzy expert system: an example in prostate cancer. Applied Mathematics and Computation, 202, 78–85. Chen, H., Zhang, J., Xu, Y., Chen, B., & Zhang, K. (2012). Performance comparison of artificial neural network and logistic regression model for differentiating lung nodules on ct scans. Expert Systems with Applications, 39(13), 11503–11509. Chun, F. K. H., Karakiewicz, P., Briganti, A., Gallina, A., Kattan, M., Montorsi, F., et al. (2006). Prostate cancer nomograms: an update. European Urology, 50, 914–926. Cordón, O., Gomide, F., Herrera, F., Hoffmann, F., & Magdalena, L. (2004). Ten years of genetic fuzzy systems: current framework and new trends. Fuzzy Sets and Systems, 141, 5–31. Epstein, J. I., Allsbrook, W. C. J., Amin, M. B., & Egevad, L. L.ISUP Grading Committee. (2005). The 2005 International Society of Urological Pathology (ISUP) consensus conference on Gleason grading of prostatic carcinoma. American Journal of Surgical Pathology, 29(9), 1228–1242. Ghosh, P., Mitchell, M., Tanyi, J., Hung, A. (2010). Automatic segmentation of the prostate using a genetic algorithm for prostate cancer treatment planning. In Ninth international conference on machine learning and applications (ICMLA) (pp. 752 – 7570). Han, M., Snow, P., Brandt, J. M., & Partin, A. (2001). Evaluation of artificial neural networks for the prediction of pathologic stage in prostate carcinoma. Cancer Supplement, 1661–1666. Herrera, F. (2008). Genetic fuzzy systems: taxonomy, current research trends and prospects. Evolutionary Intelligence, 1, 27–46.
INCA. (2010). Instituto Nacional de Câncer - Ministério da Saúde, November 2010. Available from www.inca.gov.br (In Portuguese). Karabatak, M., & Ince, C. (2009). An expert system for detection of breast cancer based on association rules and neural network. Expert Systems with Applications, 36, 3465–3469. Keles, A., Hasiloglu, S. A., Keles, A., & Aksoy, Y. (2007). Neuro-fuzzy classification of prostate cancer using NEFCLASS-J. Computers in Biology and Medicine, 37, 1617–1628. Available from http://portal.acm.org/citation.cfm?id=1287847. 1288045. Keles, A., Keles, A., & Yavuz, U. (2011). Expert system based on neuro-fuzzy rules for diagnosis breast cancer. Expert Systems with Applications, 38(5), 5719–5726. Ludwig, S., & Roos, S. (2010). Prognosis of breast cancer using genetic programming. In R. Setchi, I. Jordanov, R. Howlett, & L. Jain (Eds.), Knowledge-based and intelligent information and engineering systems. Lecture notes in computer science (Vol. 6279, pp. 536–545). Berlin/Heidelberg: Springer. Makarov, D., Trock, B., Humphreys, E., Mangold, L. A., Walsh, P. C., Epstein, J. I., et al. (2007). Update nomogram to predict pathologic stage of prostate cancer given prostate-specific antigen level, clinical stage, and biopsy Gleason score (Partin Tables) based on cases from 2000 to 2005. Urology, 69(6), 1095–1101. Matsui, Y., Egawa, S., Tsukayama, C., Terai, A., Kuwao, S., Baba, S., et al. (2002). Artificial neural network analisys for predicting pathological stage of clinically localized prostate cancer in the Japanese population. Japanese Journal of Clinical Oncology, 32(12), 530–535. Odusanya, A., Odetayo, M., Petrovic, D., Naguib, R., Lakshmi, M., & Sherbet, G. (2002). A genetic algorithm-based model for breast cancer prognosis. In Proceedings of the 6th world multiconference on systemics, cybernetics and informatics (SCI2002) (pp. 394–397). Papageorgiou, E., Spyridonos, P., Glotsos, D., Stylios, C., Ravazoulad, P., Nikiforidis, G., et al. (2008). Brain tumor characterization using the softcomputing technique of fuzzy cognitive maps. Applied Soft Computing, 8(1), 820–828. Partin, A., Kattan, M., Subong, E., Walsh, P., Wojno, K., Oesterling, J., et al. (1997). Combination of prostate-specific antigen, clinical stage, and gleason score to predict pathological stage of localized prostate cancer. Journal of the American Medical Association, 277(18), 1445–1451. Partin, A., Mangold, L., Lamm, D., Walsh, P., Epstein, J. I., & Pearson, J. (2001). Contemporary update of prostate cancer staging nomograms Partin Tables for the new millennium. Urology, 58(6), 843–848. Peña-Reyes, C. A., & Sipper, M. (1999). A fuzzy-genetic approach to breast cancer diagnosis. Artificial Intelligence in Medicine, 17, 131–155. Ross, P., Scardino, P., & Kattan, M. (2001). A catalog of prostate cancer nomograms. Journal of Urology, 162, 1562–1568. Sargent, D. (2001). Comparison of artificial neural networks with other statistical approaches. Cancer, 91(58), 1636–1642. Saritas, I., Allahverdi, N., & Sert, I. U. (2003). A fuzzy expert system design for diagnosis of prostate cancer. In Proceedings of the 4th international conference on computer systems and technologies: e-learning (pp. 345–351). New York, NY, USA: ACM. Saritas, I., Ozkan, I. A., & Sert, I. U. (2010). Prognosis of prostate cancer by artificial neural networks. Expert Systems with Applications, 37, 6646–6650. Sedighiani, K., & HashemiKhabir, S. (2009). Constructing interpretable genetic fuzzy rule-based system for breast cancer diagnostic. Proceedings of the 2009 WRI global congress on intelligent systems (Vol. 01, pp. 441–446). Washington, DC, USA: IEEE Computer Society. Available from http://portal.acm.org/ citation.cfm?id=1603895.1604812. Seker, H., Odetayo, M., Petrovic, D., & Naguib, R. N. (2003). A fuzzy logic basedmethod for prognostic decision making in breast and prostate cancers. IEEE Transactions on Information Technology in Biomedicine, 7(2), 114–122. Shah, S., & Kusiak, A. (2007). Cancer gene search with data-mining and genetic algorithms. Computers in Biology and Medicine, 37, 251–261. Shariat, S., Karakiewicz, P. I., & Roehrborn, C. M. W. K. (2008a). An updated catalog of prostate cancer predictive tools. Cancer, 113(11), 3075–3099. Shariat, S., Karakiewicz, P. I., Suardi, N., & Kattan, M. (2008b). Comparison of nomograms with other methods for predicting outcomes in prostate cancer: a critical analysis of the literature. Clinical Cancer Research, 14(14), 4400–4407. Snow, P., Smith, D., & Catalona, W. (1994). Artificial neural networks in the diagnosis and prognosis of prostate cancer: a pilot study. Journal of Urology, 152, 1923–1926. Sobin, L. H., & Wittekind, C. (Eds.). (1997). TNM classification of malignant tumors (5th ed.. International union against cancer (UICC) (5th ed.). New York: John Wiley & Sons, Inc.. Sugeno, M. (1985). Industrial applications of fuzzy control. Elsevier Science Pub. Co.. Zweig, M. H., & Campbell, G. (1993). Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clinical Chemistry, 39(4), 561–577.