Objective: To construct and cross-validate logistic regression models used for the prediction of ovarian malignancies in two groups of women with adnexal tumors., Materials and Methods: Preoperative clinical, gray-scale and color Doppler ultrasound data of 307 women treated in the Ist Dept of Gynecology of the Medical University in Lublin (group I) and 464 of women treated in the Dept of Surgical Gynecology, Medical University in Poznan (group II) were analyzed retrospectively. These data were used to construct predictive models which were developed for both groups separately and then cross-validated (12 cases in each group, six malignant and 6 benign) between the groups. Multiple logistic regression analysis was chosen to calculate probability of malignancy in each examined mass., Results: There were 228 (74.2%) benign tumors and 79 (25.7%) malignant tumors in group I. Group II consisted of 299 (64.4%) benign tumors and 165 (35.6%) malignant masses. Only six variables were included in the logistic regression model in group I. These were age, bilaterality, septae, papillary projections, volume and color score. In group II there were also 6 variables included in the regression model (menopausal status, septae, bilaterality, ascites, blood vessel localization and PI). At 50% probability of malignancy the model constructed for group I had a sensitivity and specificity of 74.6% i 94.7%, respectively. At the same probability level in group II sensitivity and specificity were 86.1% i 93.6%, respectively. Cross-validation of the predictive model constructed for group I in randomly selected cases from group II had sensitivity of 66.4% and specificity of 79.2%. Sensitivity and specificity of the group II model tested in cases from group I were 64.3% and 75.2%, respectively., Conclusions: We conclude that predictive models created with the use of multiple logistic regression analysis may be useful in preoperative discrimination of adnexal tumors. However, much better definition of diagnostic criteria, especially color Doppler score must be achieved before these models could be widely used in clinical practice.