Objective: To assess the accuracy and reproducibility of ultrasound 'pattern recognition' for the diagnosis of borderline ovarian tumors by asking experienced ultrasound operators to evaluate representative images of different types of adnexal tumor.
Methods: Digitally stored static two-dimensional B-mode images of representative cases of benign, borderline and invasive malignant ovarian tumors were independently assessed by three expert sonologists who had not performed the original real-time ultrasound examination. The outcome measures included diagnostic accuracy and interobserver agreement in the diagnosis of benign, borderline or invasive malignant ovarian tumors.
Results: One hundred and sixty-six cases were included in the final data analysis. A correct classification was made by all three experts in 83% of the primary invasive cancers, 76% of the benign masses and in 44% of the borderline malignant tumors (P < 0.01). The experts showed a tendency to misclassify borderline tumors as benign rather than primary invasive (ratio of 8 : 1 for Expert A, 4 : 1 for B and 6 : 1 for C). The interobserver agreement between any two experts was very good when they were tested for their ability to discriminate between invasive and non-invasive (benign and borderline) ovarian tumors (Cohen's kappa 0.85-0.88), but poorer for the discrimination between malignant (invasive and borderline) and benign tumors (kappa 0.70-0.78).
Conclusions: The accuracy of ultrasound diagnosis of borderline tumors is lower in comparison with benign and invasive malignant lesions. The diagnostic performance and interobserver agreement are better when the outcomes are dichotomized into non-invasive and invasive malignant lesions, as opposed to the traditional diagnosis of benign and malignant tumors.