OC118: Mathematical models to evaluate ovarian masses?can they beat an expert operator?

Abstract

For persistent ovarian tumours it is crucial to distinguish between malignant tumours and benign tumours, which may be suitable for minimal access surgery. The aim of this study was to cross-validate prospectively logistic regression and artificial neural network (ANN) models to predict malignancy in patients presenting with adnexal masses. Ultrasonographic features of the adnexal mass, age of the patient and serum CA125 levels were encoded as features for two ANN models. Previously published ANN models were prospectively tested on a set of 92 patient records. The area under the receiver operating characteristic curves were 0.89, 0.89, and 0.92 for the logistic regression model, ANN1, and ANN2, respectively. At a decision threshold of 50% the logistic regression model reached a sensitivity of 68% and specificity of 92%; ANN1 had a sensitivity of 77% and specificity of 82%, whereas ANN2 had a sensitivity of 81% and specificity of 85%. However, subjective assessment of the sonologist had a sensitivity of 97% and specificity of 97%. A second moderately experienced gynaecologist independently assessed all ultrasound images and clinical records and reached a sensitivity of 94% and specificity of 92%. Simple logistic regression and ANN models can be trained to predict malignancy with a reasonable degree of accuracy. However, current models can certainly not beat an expert operator. Signs of malignancy include the presence of papillary structures, irregular solid areas, septa and a strong vascularization at colour Doppler imaging. Further refinement of mathematical models might help the clinician in making an adequate preoperative diagnosis. However, the results of multicentre trials like IOTA (International Ovarian Tumour Analysis) need to be awaited before clinical use of mathematical models could be advocated. Future research should focus on difficult tumours and early cancers.