PT  - JOURNAL ARTICLE
AU  - F Moro
AU  - M Albanese
AU  - L Boldrini
AU  - V Chiappa
AU  - J Lenkowicz
AU  - F Bertolina
AU  - F Mascilini
AU  - R Moroni
AU  - MA Gambacorta
AU  - F Raspagliesi
AU  - G Scambia
AU  - AC Testa
AU  - F Fanfani
TI  - 387 Developing and validating ultrasound-based radiomics models for predicting high-risk endometrial cancer
AID  - 10.1136/ijgc-2021-ESGO.136
DP  - 2021 Oct 01
TA  - International Journal of Gynecologic Cancer
PG  - A88--A89
VI  - 31
IP  - Suppl 3
4099  - http://ijgc.bmj.com/content/31/Suppl_3/A88.short
4100  - http://ijgc.bmj.com/content/31/Suppl_3/A88.full
SO  - Int J Gynecol Cancer2021 Oct 01; 31
AB  - Introduction/Background*Transvaginal ultrasound examination is the first imaging investigation for endometrial cancer. Ultrasound-based models for predicting high risk endometrial cancer have recently been published. However, none of these models includes radiomics features. Radiomics is an innovative high throughput technique extracting and translating high numbers of features from medical images into mineable data.Aim of this study was to develop and validate ultrasound-based radiomics models, aiming to differentiating high risk category, as defined by ESMO-ESGO-ESTRO in 2016, versus the remaining categories of risk.Methodology This is a multicenter retrospective observational study. Patients with histologically confirmed diagnosis of endometrial cancer who had undergone preoperative ultrasound examination between 2016 and 2019 were identified from two centers. 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 was applied to the ultrasound images. Clinical (including preoperative biopsy), ultrasound and radiomics features that proved to be different at the univariate analysis on the training set were considered for multivariate analysis and for developing ultrasound-based machine learning assessment models.Result(s)*For discriminating high risk category versus the other categories one random forest model from the radiomics features (radiomics model), one 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.In the validation set, the radiomics model for predicting high risk showed AUC 0.80, sensitivity 58.7%, specificity 85.7%, positive likelihood ratio (LR+) 4.10 and negative likelihood ratio (LR-) 0.48; the clinical-ultrasound model showed AUC 0.87, sensitivity 67.3%, specificity 89.2%, LR+ 6.29 and LR- 0.37; and the mixed model showed AUC 0.88, sensitivity 67.3%, specificity 91.0%, LR+ 7.55 and LR- 0.36 (table 1).View this table:Abstract 387 Table 1 Logistic regression models for discriminating high risk group versus other three groups, and low risk group versus the other risk groups for validation setConclusion*The mixed model including radiomics, clinical (including preoperative biopsy) and ultrasound features provided the best performance, even if the accuracy was slightly higher in comparison with the model based only on clinical and ultrasound variables. Interestingly, the model based only on radiomics features was able to provide good accuracy to discriminate high risk group versus the others.