EP01.10: Interoperator variability and accuracy in fetal renal parenchyma assessment versus machine learning model in unilateral obstructive uropathies

To assess accuracy and interoperator variability in the subjective evaluation of fetal renal aspect and explore the role of radiomic techniques to predict the outcome in fetuses with unilateral urinary tract obstructions. A retrospective study was conducted to re-assess third trimester renal ultrasound images in fetuses with unilateral obstructive uropathies. Two fetal medicine experts with >10 years of clinical experience were asked to predict postnatal outcomes based on ultrasound renal images. Radiomic shape features of fetal urinary tract were extracted from the same images and a machine-learning model was developed. Adverse composite outcome was defined by the necessity of postnatal surgery and/or abnormal kidney function on nuclear scan. Cohen's kappa was used to study interobserver variability and accuracy for predicting postnatal outcome was assessed by sensitivity and specificity. Renal images of 86 fetuses with unilateral obstructive uropathy were evaluated. Moderate agreement in the subjective assessment of fetal renal parenchyma was found for presence of dilated calices (K0.3 – p < 0.003) and loss of corticomedullary differentiation (K0.33– p < 0.001), while poor agreement was observed for the evaluation of renal echogenicity (K0.18, p0.04). Concerning the prediction of postnatal outcome, scarce agreement was found with a Cohen's kappa of 0.4 (p < 0.001) for prediction of necessity of postnatal surgery, and a K0.41 (p < 0.001) for prediction of postnatal abnormal renogram. Fetal expert's prediction of adverse outcome yielded a mean sensitivity of 57% and specificity of 70%, while the machine learning model showed an AUC of 0.78, with sensitivity of 76% and specificity of 67%. Poor to moderate agreement was observed in the subjective evaluation of fetal renal parenchyma between two expert operators. Accuracy in evaluating postnatal renal outcomes was improved by the use of a machine-learning algorithm, compared to the subjective prediction of fetal experts.