28 An MRI radiomic signature for predicting brachytherapy outcomes in locally advanced cervical cancer

Abstract

Introduction The standard treatment of locally advanced cervical cancer (LACC) consists of concomitant chemoradiation followed by brachytherapy. The recent implementation of image-guided adaptive brachytherapy (IGABT) has shown a significant improvement in local control while limiting toxicity. We identified a radiomic signature of LACC relapse based on per-brachytherapy MRI. Methods 100 patients with LACC treated with pulsed-dose-rate IGABT after initial chemoradiation were retrospectively included (TS: N = 60, training; VS: N = 40, validation). Using LIFEx, we extracted 30 textural features from the clinical target volume delineated on T2 FSE MRI. A methodological study was carried out to identify the most robust and informative features. The ability of remaining features to predict relapse was assessed through univariate and multivariate statistical analysis. Results 9 clusters including 24 robust features were considered for the clinical analysis. In univariate analysis, five features were statistically significant predictors of overall recurrence (TS: p 0.05). A 5-feature signature predicting relapse was identified and validated (TS: AUC = 0.88, p 0.00001, VS: AUC = 0.78, p 0.01) and performed better than the five features determined from univariate analysis and the tumor volume (TS, Delong’s test, p 0.05). Conclusion Our radiomics signature has been validated in an institutional cohort. MRI-based texture analysis could provide clinically relevant information to adapt brachytherapy planning. However, standardization of MRI protocols or development of normalization methods is still needed to perform multi-center MRI-based radiomics studies.