Artificial neural network models for the preoperative discrimination between malignant and benign adnexal masses

Objective The aim of this study was to generate and evaluate artificial neu ral network (ANN) models from simple clinical and ultrasound-derived criter ia to predict whether or not an adnexal mass will have histological evidenc e of malignancy. Design The data were collected prospectively from 173 consecutive patients who were scheduled to undergo surgical investigations at the University Hos pitals, Leuven, between August 1994 and August 1996. The outcome measure wa s the histological classification of excised tissues as malignant (includin g borderline) or benign. Methods Age, menopausal status and serum CA 125 levels and sonographic feat ures of the adnexal mass were encoded as variables. The ANNs were trained o n a randomly selected set of 116 patient records and tested on the remainde r (n = 57). The performance of each model was evaluated using receiver oper ating characteristic (ROC) curves and compared with corresponding data from an established risk of malignancy index (RMI) and a logistic regression mo del. Results There were 124 benign masses, five of borderline malignancy and 44 invasive cancers (of which 29% were metastatic); 37% of patients with a mal ignant or borderline tumor had stage I disease. The best ANN gave an area t inder the ROC curve of 0.979 for the whole dataset, a sensitivity of 95.9% and specificity of 93.5%. The corresponding values for the RMI were 0.882, 67.3% and 91.1%, and for the logistic regression model 0.956, 95.9% and 85. 5%, respectively. Conclusion An ANN can be trained to provide clinically accurate information , on whether or not an adnexal mass is malignant, from the patient's menopa usal status, serum CA 125 levels, and some simple ultrasonographic criteria .