Developing and validating ultrasound-based radiomics models for predicting high-risk endometrial cancer.

OBJECTIVE The primary aim of the study was to develop and validate radiomics models, applied to ultrasound images, capable of differentiating high-risk endometrial cancer, as defined by ESMO-ESGO-ESTRO (European Society for Medical Oncology -European Society of Gynaecological Oncology-European Society for Radiotherapy & Oncology) in 2016, from other cancers. The secondary aim was to develop and validate radiomics models for differentiating low risk endometrial cancer from other endometrial cancers. METHODS This is a multicenter retrospective observational study. We identified consecutive patients from two participating centers with histologically confirmed diagnosis of endometrial cancer who had undergone preoperative ultrasound examination by an experienced examiner, between 2016 and 2019. Patients recruited in Center 1 (Rome) were included as training set (n=396), while patients enrolled in Center 2 (Milan) as external validation set (n=102). Radiomics analysis -extraction of a high number of quantitative features from medical images- was applied to the ultrasound images. Clinical (including preoperative biopsy), ultrasound and radiomics features that proved to be statistically significant different between high risk group versus other groups and low risk group versus others at univariate analysis on the training set were considered for multivariate analysis and for developing ultrasound-based machine learning risk prediction models. For discriminating between the high-risk group and the others a random forest model from the radiomics features (radiomics model), a binary logistic regression model from clinical and ultrasound features (clinical-ultrasound model), and another binary logistic regression model from clinical, ultrasound and previously selected radiomics features (mixed model) were created. Similarly, for discriminating between the low risk group and the other groups, a random forest model (radiomics model), a binary logistic regression model (clinical-ultrasound model), and another binary logistic regression model (mixed model) were created. The models developed in the training set were validated in the validation set. RESULTS In the validation set, the radiomics model for predicting high risk showed area under the receiver operating characteristic curve (AUC) 0.80, sensitivity 58.7% and specificity 85.7% (using the optimal cut-off 0.41); the clinical-ultrasound model showed AUC 0.90, sensitivity 80.4% and specificity 83.9% (using the optimal cut off 0.32); and the mixed model showed AUC 0.88, sensitivity 67.3% and specificity 91.0% (using the optimal cut off 0.42). The radiomics model to predict low risk, in the validation set, showed AUC 0.71, sensitivity 65.0% and specificity 64.5% (using the optimal cut off 0.38); the clinical-ultrasound model showed AUC 0.85, sensitivity 70.0% and specificity 80.6% (using the optimal cut off 0.46); and the mixed model showed AUC 0.85, sensitivity 87.5% and specificity 72.5% (using the optimal cut off 0.36). CONCLUSIONS Radiomics seems to have some ability to discriminate between low-risk endometrial cancer and other endometrial cancers and better ability to discriminate between high-risk endometrial cancer and other endometrial cancers; however, the addition of radiomics features to the clinical-ultrasound models did not add any notable increase in performance. Other efficacy studies and then further effectiveness studies are needed to validate the performance of the models. This article is protected by copyright. All rights reserved.