UICC World Cancer Congress 2006

Objective: Diagnosis of malignant ovarian tumors is one of the most difficult problems in gynecological oncology. The large amount of data obtained from these studies should be analyzed with the use of advanced statistical methods. The aim of our study was: 1) Creation of the prediction models with the high specificity and sensitivity, based on the method of data mining, which could allow for the precise preoperative differentiation of ovarian masses. 2) Comparision of the results of the best ultrasonographic index classification and obtained models 3) Selection of the best decision based on “k-judges voting” algorithm and prospective verification of the diagnostic effectiveness of the constructed decision models. Methods: The clinical, sonographic features and postoperative histological results of the group of 340 women with adnexal masses, operated in the I-st Department of Gynaecology of the Medical University of Lublin in the years of 2002-2005, were analysed. There were 243 (71.5%) benign ovarian masses and 97 (28.5%) malignant tumors. Results: Data mining techniques can be used to effectively support the initial diagnostics of adnexal tumors. The prognostic model constructed with the use of MLP networks which uses clinical and imaging (2D and 3D) data in the process of the preoperative diagnostics of ovarian tumors is characterised with the higher sensitivity and specificity then the individual diagnostic tests. Multilayer perceptron network allow for more effective prediction of the tumor character then probabilistic neural network and support vector machine method, K-nearest neighbors algorithm, naive Bayes classifiers or these methods connection. SVM method has good generalization effect.