Purpose or Case Report: Pelvic ultrasound is commonly used to detect ovarian torsion, but the diagnosis remains challenging as there is no single pathognomonic feature. This retrospective case-controlled study aims to identify an algorithm to detect torsion based on common ultrasound imaging features.
Methods & Materials: In this IRB approved retrospective study, patients who underwent pelvic ultrasound because of abdominal or pelvic pain between 2005 and 2015 were identified and classified as either torsion-absent or torsion-confirmed, based on final radiologic and surgical evidence. Of patients with torsion, patients less than 1 year of age were excluded as infants with torsion have different presentation and imaging features than older patients. In total, 99 torsion-confirmed and 331 sequential torsion-absent cases from 2015 were included. Radiologic features extracted from the ultrasound images included binary variables of presence of Doppler flow, free fluid, peripheral follicles, as well as ovary medialization and the continuous variables of right and left ovarian size. These features were fed into supervised learning systems to find viable decision algorithms. Data was divided into 60% training and 40% validation data sets and performance was assessed using sub-sets of the validation set.
Results: All variables had statistically significant differences between the torsion-confirmed and torsion-absent groups with p-values < 0.005 (Table 1). Using single variables to identify torsion provided only modest detection performance with areas under the curve (AUC) for medialization, peripheral follicles, and absence of flow of 0.76+/-.16, 0.66+/-0.14, and 0.82 +/-.14 respectively. The best decision tree (Fig. 1) using a combination of variables yielded an AUC of 0.96 +/- 0.07 and required knowledge of the presence of flow, peripheral follicles, the volume of both ovaries, and presence of cysts.
Conclusions: An algorithm combining multiple ultrasound imaging features associated with ovarian torsion performs better than simple approaches relying on singular features. While complex combinations using multiple interaction models provide slightly better performance, a clinically pragmatic decision tree can be employed to detect torsion and provide sensitivity levels of 95+/-14% with a specificity of 92 +/-2%.