Logistic regression models to distinguish between benign, primary invasive, borderline malignant and metastatic ovarian tumors

Objectives: To develop mathematical models that go beyond the classification of ovarian tumors as either benign or malignant. Methods: This study included the 1066 patients from the International Ovarian Tumor Analysis (IOTA) group s dataset. These patients were recruited at nine centers across Europe and underwent transvaginal gray-scale as well as color Doppler ultrasound examination. More than 40 measurements were available to develop mathematical models. The outcome was the classification of the tumor as benign, primary invasive, borderline malignant or metastatic. Logistic regression was used to develop models to confront each pair of outcomes (six models). This allowed identification of the most important variables for each pair. The results of the six models were combined to arrive at estimated probabilities for each class. Data were split into a training set (754 patients) and a test set (312 patients). Results: Eight hundred patients had a benign tumor (75.0%), 169 had a primary invasive tumor (15.9%), 55 had a borderline malignant tumor (5.2%) and 42 had a metastatic tumor (3.9%). In general, 10 variables were used in the six logistic regression models: age, maximal diameter of the lesion and of the largest solid component, blood flow in the papillary projections, personal history of ovarian cancer, irregular internal cyst walls, ascites, bilateral lesions, unilocular tumor and entirely solid tumor. The presence of ascites and the maximal diameter of the largest solid component were the most important variables that distinguished the four types of tumor. Test set areas under the receiver-operating characteristic curve were between 0.834 (borderline tumors) and 0.933 (primary invasive tumors). Conclusions: Logistic regression models were very good at distinguishing between benign, primary, invasive, borderline malignant and metastatic tumors.