G. Baldassari

Radiomics‐based ultrasOund Model for differentiating Uterine Sarcomas from leiomyomas ( ROMUS ): a retrospective pilot Multicenter Italian Trials in Ovarian Cancer ( MITO ) study

ABSTRACT Objective To develop machine‐learning models that incorporate clinical information and radiomics features extracted from ultrasound images to distinguish uterine sarcomas from leiomyomas. Methods This retrospective, multicenter, pilot case–control study included 200 patients (100 with a uterine sarcoma and 100 with a usual‐type leiomyoma, i.e. including no benign leiomyoma variants) who underwent preoperative ultrasound examination between January 2010 and June 2022. The patient cohort was split (70:30) into training and validation sets, with the same proportion of leiomyomas and sarcomas in each subset. We extracted radiomics features belonging to different families: intensity‐based statistical features and textural features. The variables used in model building were patient age and the radiomics features that differed statistically significantly between sarcomas and leiomyomas and that were not redundant based on Spearman's correlation coefficient. Logistic regression, random forest, extreme gradient boosting (XGBoost) and support vector machine models were tested in the model development process. We evaluated the performance of the models in differentiating between sarcomas and leiomyomas using the area under the receiver‐operating‐characteristics curve (AUC), accuracy, sensitivity and specificity. We compared these results to those of subjective assessment by the original ultrasound examiner and to those of two independent expert ultrasound examiners who, blinded to clinical history, reviewed the same grayscale ultrasound images as those used for the radiomics analysis. Results Sixty‐three radiomics features were extracted. Of these, eight differed statistically significantly between sarcomas and leiomyomas and were not correlated, so were selected for inclusion in model building. In the validation set, the model that performed best in differentiating between sarcomas and leiomyomas was an XGBoost model integrating patient age and radiomics features. In the validation set, this model had an AUC of 0.93, sensitivity of 0.93 and specificity of 0.83, at a risk‐of‐malignancy cut‐off of 47% (the cut‐off that yielded the highest number of correct classifications based on Youden's index in the training set). The corresponding results for the model integrating only the radiomics features were: AUC of 0.87, sensitivity of 0.87 and specificity of 0.83. Subjective assessment by the original ultrasound examiner had a sensitivity of 0.87 and specificity of 1 in the validation set, while retrospective review of grayscale ultrasound images by ultrasound experts had a sensitivity of 0.87 and specificity of 0.80 (same results for both reviewers). Conclusion A model including eight radiomics features and patient age demonstrated reasonably good discriminative and classification performance for distinguishing uterine sarcomas from leiomyomas. Its classification ability was similar to that of subjective assessment by the original ultrasound examiner, being more sensitive but less specific. To confirm the role of radiomics for discriminating between uterine sarcomas and leiomyomas, large prospective studies including benign leiomyoma variants are needed. If good performance of radiomics models can be confirmed, integrating automated radiomics analysis into ultrasound machine software may help ultrasound examiners to discriminate between sarcomas and benign leiomyomas. © 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.

Developing and validating ultrasound‐based machine‐learning models incorporating radiomics features to predict malignancy in adnexal masses

ABSTRACT Objective The primary aim of this study was to develop and internally validate ultrasound‐based radiomics models to discriminate between all types of benign and malignant adnexal masses. The secondary aim was to compare the performance of the radiomics models with that of the Assessment of Different NEoplasias in the adneXa (ADNEX) model. Methods This was a retrospective, observational, single‐center study, for which all patients with an adnexal mass that were included in the ongoing International Ovarian Tumor Analysis phase‐5 and phase‐7 studies and were examined using ultrasound between January 2012 and December 2023 at Fondazione Policlinico Universitario A. Gemelli IRCCS, Rome, Italy, were eligible for inclusion. Inclusion criteria were: adnexal mass detected by ultrasound; surgical removal of the adnexal mass within 180 days after the ultrasound examination; histological confirmation of an adnexal mass; and absence of a synchronous malignant tumor. Patients without digital ultrasound images saved in DICOM format were excluded. The patient cohort was split randomly into training and validation sets using a stratified split with a ratio of 70:30, to preserve the proportion of benign and malignant cases in the two sets. Two machine‐learning models for discriminating between benign and malignant adnexal masses were built using one image per tumor, with 5‐fold cross‐validation for hyperparameter tuning, and were tested on the validation set. The variables used in model building were patient age, serum CA 125 level and the radiomics features that differed significantly between benign and malignant tumors (determined using the Mann–Whitney U ‐test with Benjamini–Hochberg correction) and were not redundant based on Pearson correlation analysis. Histology was the reference standard. We assessed the discriminative performance of the radiomics models using the area under the receiver‐operating‐characteristics curve (AUC) and classification performance using sensitivity and specificity at the optimal cut‐off of each model to classify the mass as malignant, as determined by Youden's index. The diagnostic performance of the developed radiomics models was compared with that of the ADNEX model (AUC, sensitivity and specificity at the 10% risk‐of‐malignancy cut‐off, which is the recommended threshold for clinical use of the ADNEX model). Results In total, 4501 patients met the inclusion criteria. Among these, 2428 patients were excluded owing to an absence of ultrasound images or images unsuitable for radiomics analysis. Overall, a total of 2073 patients were included in the analysis, of whom 803 (38.7%) had a histologically confirmed malignant tumor. In the validation set ( n  = 622, including 254 malignancies), the clinical–radiomics model trained using the eXtreme Gradient Boosting algorithm, including age, serum CA 125 level and 14 selected radiomics features, achieved the highest performance, with an AUC of 0.89 (95% CI, 0.86–0.92), sensitivity of 0.83 (95% CI, 0.79–0.88) and specificity of 0.81 (95% CI, 0.77–0.85) at the optimal cut‐off (31% risk of malignancy, based on Youden's index). At a 10% risk‐of‐malignancy cut‐off, it had a sensitivity of 0.94 (95% CI, 0.91–0.97) and specificity of 0.48 (95% CI, 0.42–0.53). The ADNEX model had an AUC of 0.95 (95% CI, 0.93–0.97), sensitivity of 0.97 (95% CI, 0.95–0.99) and specificity of 0.72 (95% CI, 0.68–0.77) at the 10% risk‐of‐malignancy cut‐off in the validation set. Conclusions Our results support further exploration of radiomics analysis for distinguishing between benign and malignant adnexal masses in larger study populations. Future studies should consider using multiple images per tumor and testing alternative model‐building methods, and should perform external validation to assess the generalizability of the radiomics 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.