280 Radiomics based prediction of pathological and molecular features of endometrial cancer managed at an Irish tertiary referral centre

Introduction/Background Endometrial cancer (EC) is the most common gynaecological malignancy in the developed world. Prognostic indicators for EC have recently been revised to include molecular features. Radiomics is the field of extracting quantitative image biomarkers from medical imaging. Radiomics has shown potential to provide imaging biomarkers to predict molecular features. Methodology We performed retrospective analysis on a population treated for EC at a tertiary referral centre. T2-weighted MRI were manually segmented for the endometrial tumour. Feature extraction was performed using PyRadiomics, followed by feature filtering using correlation and Near-Zero Variance filtering. Multivariate logistic regression was performed to develop a model using the 10 most predictive radiomic features. Prediction performance was assessed using Area Under the Curve (AUC) of the receiver operating characteristic curve (ROC) and accuracy. Results 71 patients were included. There were 55 (77%) cases of endometrioid cancer and 16 (23%) cases of non-endometrioid. The mean age was 64.5 years (SD 12.7 years) and the mean Body Mass Index was 34.8 kg/m2 (SD 10.7 kg/m2).The multivariate logistic regression model produced an AUC of 0.85 for lympho-vascular space invasion (LVSI), 0.7 for non-endometrioid histology, 0.62 for tumour grade, 0.64 for (Mismatch Repair (MMR) instability and 0.3 for p53 mutation. ROC curves and confusion matrixes of the logistic regression analysis are demonstrated in Figure 1. Conclusion We developed a radiomics model that has equivocal performance to more complex models in predicting non-endometrioid histology and LVSI. Model performance for MMR instability and p53 mutation may be improved by a larger population. This study adds to the potential use of radiomics in EC. Further study is needed to standardise radiomic processing in EC and identify the most appropriate imaging sequences and predictive models to assess molecular features. Disclosures No conflict of interest to disclose.