Differentiating high- and low-grade serous ovarian carcinoma using radiomics: a pilot study

To identify ultrasound-based radiomics features capable of distinguishing between high-grade serous ovarian carcinomas (HGSC) and invasive low-grade serous ovarian carcinomas (LGSC), and to develop machine-learning models that include radiomics features to discriminate between the two.

This was a single-center, analytical, observational, retrospective pilot study of patients referred to Fondazione Policlinico Universitario 'A. Gemelli' IRCCS, Rome, Italy, between January 2014 and June 2022 with a histological diagnosis of HGSC or invasive LGSC, who had DICOM images available from a preoperative ultrasound examination. The extracted radiomics features belonged to two different families: intensity-based statistical features and textural features. Feature selection was performed via univariate analysis using the Wilcoxon-Mann-Whitney statistical test, followed by multivariate selection using the Boruta algorithm, Pearson correlation analysis to reduce collinearity and Akaike information criterion stepwise regression. Selected features were used to train logistic regression models. To distinguish between HGSC and LGSC, two logistic regression models were developed: a radiomics model based solely on radiomics features, and a clinical-radiomics model combining radiomics features with patient age. For comparison, two additional models were constructed: a radiomics-ultrasound model incorporating radiomics features and ultrasound variables that were significantly different between the two histotypes, and an ultrasound model based exclusively on these ultrasound variables. The area under the receiver-operating-characteristics (ROC) curve (ROC-AUC) and the area under the precision-recall (PR) curve (PR-AUC) were calculated. Modeling results were internally validated via the bootstrap resampling method, and optimism-corrected metrics were computed.

A total of 153 patients (120 with a histological diagnosis of HGSC and 33 with a histological diagnosis of invasive LGSC) were recruited, providing a total of 332 images for analysis. For each image, 75 radiomics features were extracted. After feature selection, three textural features, 'gray-level co-occurrence matrix informational measure of correlation 2' ('F_cm.info.corr.2'), 'gray-level run length matrix run percentage' ('F_rlm.r.perc') and 'gray-level size zone matrix zone size variance' ('F_szm.zs.var') were used for the radiomics model; one textural feature (F_szm.zs.var) and age were used for the clinical-radiomics model; one textural feature (F_cm.info.corr.2) and one ultrasound variable (color score) were used for the radiomics-ultrasound model; and three ultrasound variables (color score, presence of papillary projections, presence of the ovarian crescent sign) were used for the ultrasound model. The internally validated (optimism-corrected) ROC-AUC was 0.69 (95% CI, 0.57-0.77) for the radiomics model, 0.80 (95% CI, 0.72-0.90) for the clinical-radiomics model, 0.68 (95% CI, 0.56-0.77) for the radiomics-ultrasound model and 0.78 (95% CI, 0.68-0.88) for the ultrasound model. The clinical-radiomics model performed similarly to the ultrasound model (optimism-corrected ROC-AUC, 0.80 vs 0.78; P = 0.726) and better than the radiomics (optimism-corrected ROC-AUC, 0.80 vs 0.69; P = 0.042) and radiomics-ultrasound (optimism- corrected ROC-AUC, 0.80 vs 0.68; P = 0.048) models.

Our findings indicate that, despite detectable differences in radiomics features between HGSC and LGSC, current radiomics and imaging-based approaches do not provide sufficient added value over conventional ultrasound for reliable preoperative discrimination. Future research may be necessary to develop more accurate and clinically useful artificial intelligence-based predictive models. © 2026 The Author(s). Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of International Society of Ultrasound in Obstetrics and Gynecology.