Artificial neural networks in sonographic characterization of adnexal tumors: a preliminary study.

Objectives: To apply computer generated artificial neural networks (ANNs) for the development of a statistical model, helpful in better preoperative discrimination between benign and malignant adnexal tumors. Materials and methods: The retrospective study included data of 162 patients who were examined with gray scale and color Doppler sonography (B-K Medical, 2002 ADI, Denmark and Voluson 530, and 730 Kretz, Austria) because of adnexal tumors. Following variables were analyzed with the use of ANN: patient's age, menopausal status, years since menopause, tumor size and volume, presence of solid parts, papillary projections >3 mm, localization of blood vessels (central or peripheral), resistive and pulsatility indices, peak systolic velocity and color Doppler score according to Timmerman et al. (1999). The output variable was classified as either benign or malignant. The ANNs were generated with the use of a statistical package (Statistical Neural Networks, Statsoft, USA) and included linear, radial basis function and multiple layer perceptron networks. The networks were trained on randomly selected 81 cases, verified (n = 40), and tested on the remaining 41 cases. Sensitivity and specificity of generated statistical models were compared with those calculated for all individual sonographic parameters. Results: In 43 women, the cysts resolved spontaneously whereas 119 patients were subsequently operated. Removed tumors were classified histologically as either benign (n = 93) or malignant (n = 26), including four borderline; five stage I ovarian cancers. The best performance was found for the radial basis function network with six input neurons and one hidden neuron. The specificity and sensitivity at the best decision cut-off point were 96.1% (25 of 26) and 94.8% (129 of 136), respectively. These percentages were significantly better than respective prognostic values derived from independent use of each tested variable (P < 0.05). Conclusion: Evaluation of... [ABSTRACT FROM AUTHOR]