Accuracy of ultrasound subjective 'pattern recognition' for the diagnosis of borderline ovarian tumors.

Abstract

To assess the value of pattern recognition for the preoperative ultrasound diagnosis of borderline ovarian tumors (BOTs). This was a prospective study of women who were referred to our regional cancer center with the diagnosis of an adnexal mass on a Level II (routine) gynecological ultrasound scan. Women with lesions of uncertain nature were referred for a Level III (expert) ultrasound scan in our tertiary center. The tumor pattern recognition method was used to differentiate between various types of ovarian tumors. Morphological features suggestive of BOTs were: unilocular cyst with a positive ovarian crescent sign and extensive papillary projections arising from the inner wall, or a cyst with a well defined multilocular nodule. The ultrasound findings were compared with the final histological diagnosis. A total of 224 women with an adnexal mass of uncertain nature were referred for an expert scan, 166 (74.1%) of whom underwent surgery. In this group of women the final histological diagnoses were: 99 (60%) benign lesions, 32 (19%) invasive ovarian cancer and 35 (21%) BOTs. Using pattern recognition combining the different morphological features, a correct preoperative diagnosis of BOT was made in 24/35 (68.6%) women: area under the receiver-operating characteristics curve 0.812 (standard error 0.049; 95% CI, 0.716-0.908), sensitivity 0.69 (95% CI, 0.52-0.81), specificity 0.94 (95% CI, 0.88-0.97), positive likelihood ratio 11.3 (95% CI, 5.53-22.8) and negative likelihood ratio 0.34 (95% CI, 0.21-0.55). Ultrasound diagnosis of BOTs is highly specific. However, typical features are absent in one-third of cases, which are typically misdiagnosed as benign lesions.