ESTRO 2021 Abstract Book

S1551

ESTRO 2021

93 features were extracted from the GTV on each reconstruction using pyRadiomics and directly compared with a one-way ANOVA test. Unsupervised feature selection removed volume-correlated and redundant features, using Spearman’s rank correlation coefficient. Remaining features were assessed for their ability to predict DR. Bootstrapped univariable Cox analysis was performed and significant features (p<0.05) were included in a multivariable Cox model with variables: tumour volume, tumour motion, lobe location, performance status, age and T stage. Over 500 resamples, selected features were recorded and concordance index (CI) was calculated. Finally, models were built using the top ranked features across resamples, with number of features determined by the median signature size. Each model was assessed in terms of Akaike Information Criterion (AIC) and prognostic features. Results 223 patients remained in the analysis. Over 60% of selected features differed between image types (Fig.1B). Following feature selection, the percentage of remaining features was: 11%, 13%, 15% and 14% for T50, T0, Mean MC and Median MC. Median signature length was consistent, and FirstOrder_90Percentile was selected regardless of model. Both MC reconstructions offered virtually the same information. Overall, T50 (exhale) had the highest number of prognostic features remaining in multivariable analysis (Fig.2A). Fig.2B demonstrates improved performance (higher CI and lower AIC) in the single phase models.