Neural networks in the diagnosis of malignant ovarian tumours

Objective To assess the role of neural networks in predicting the likelihood of malignancy in women presenting with ovarian tumours. Design Retrospective case study. Setting University Department of Obstetrics and Gynaecology, St James's Hospital, Leeds. Methods Information from 217 cases with histologically proven benign, borderline or malignant tumours was extracted for study. Four variables (age, ultrasound findings with and without colour Doppler imaging and CA125) were entered in the neural network classifier. The neural network results were compared with logistic regression analysis. Results When used in the neural network the variables of age, CA125 and ultrasound score produced the best result with a sensitivity of 95% and a corresponding specificity of 78% in predicting malignancy. Logistic regression gave a sensitivity or 82% for a specificity of 51%. Conclusion The neural network is a good method of combining diagnostic variables and may be a useful predictor of malignancy in women presenting with ovarian tumours. A comparison of the performance of the neural network with conventional diagnostic methods would be warranted prior to use in clinical practice.