Sonographic prediction of malignancy in adnexal masses using an artificial neural network

Objective To generate a neural network algorithm which computes a probability of malignancy score for pre-operative discrimination between malignant and benign adnexal tumours. Design A retrospective analysis of previously collected data. Information from 75% of the study group was used to train an artificial neural network and the remainder was used for validation. Setting The Gynaecological Ultrasound Research Unit at King's College Hospital, London. Population Sixty-seven women with known adnexal mass who had been examined using transvaginal B-mode ultrasonography and colour Doppler imaging with pulse spectral analysis immediately before surgery. The excised masses were classified histologically as benign (n = 52) or malignant (n = 15), of which three were borderline. Methods The variables that were put into the artificial neural network were: age, menopausal status, maximum tumour diameter, tumour volume, locularity, the presence of papillary projections, the presence of random echogenecity, the presence of analysable blood flow velocity waveforms, the peak systolic velocity, time-averaged maximum velocity, the pulsatility index, and resistance index. Histological classification, categorised as benign or malignant, was the output result. Results A variant of the back propagation method was selected to train the network. The overall architecture of the network with the best performance contained an input layer with four variables (age, time-averaged maximum velocity, papillary projection score and maximum tumour diameter), a hidden layer with three units and an output layer with one. The sensitivity and specificity at the optimum diagnostic decision value for the artificial neural network output (0.45) were 100% (95% CT 78.2%-100%) and 98.1% (95% CI 89.5%-100%), respectively. These values were significantly better than those obtained from the independent use of the resistance index, pulsatility index, time-averaged maximum velocity or peak systolic velocity at their optimum decision values (P < 0.01). Conclusion Artificial neural networks may be used on clinical and ultrasound derived end-points to accurately predict ovarian malignancy. There is a need for a prospective evaluation of this technique using a larger number of patients.