A nomogram combining clinical features, O-RADS US, and radiomics based on ultrasound imaging for diagnosing ovarian cancer

Background We aimed to develop and validate a nomogram for diagnosing ovarian cancer from ovarian masses based on clinical information, O-RADS US, and radiomics. Methods A total of 981 patients with ovarian masses from two centers were randomly divided into the training cohort (n = 686) and the validation cohort (n = 295). We defined the region of interest (ROI) of the tumor by manually drawing the tumor contour on the ultrasound image of the lesion. The radiomics features were extracted from ultrasound images, and the radiomics score was then calculated. O-RADS US characteristics, radiomics score, and clinical features selected using the LASSO algorithm were used to develop O-RADS US + Radscore + Clinical, Radscore + Clinical, and O-RADS US + Clinical models, respectively. Receiver operating characteristic (ROC), decision curve analysis, and calibration curve were used to evaluate the performance of the nomogram models. Results Age, CA125, O-RADS US, and radiomics score were related to ovarian malignancy through univariate and multivariate logistic regression analyses. In the training and validation datasets, the areas under the ROC curve (AUC) of O-RADS US + Clinical model were 0.830 and 0.815, respectively, and those for the Radscore + Clinical model were 0.876 and 0.867, respectively. The O-RADS US + Radscore + Clinical nomogram model presented improved AUC values of 0.967 in the training group and 0.951 in the validation group, significantly higher than that of Radscore + Clinical and O-RADS US + Clinical models. The calibration curve and the clinical decision curve analysis demonstrated that the nomogram models had high clinical benefits. The O-RADS US + Radscore + Clinical model had the highest net return. Conclusions Combination nomogram model that integrates clinical features, O-RADS US, and radiomics based on ultrasound image analysis could predict ovarian malignancy with high diagnostic accuracy, indicating that this model might have a role in preoperative diagnosis for differentiating benign and malignant ovarian tumors.