ESTRO 2024 - Abstract Book

S3115

Physics - Autosegmentation

ESTRO 2024

Conclusion:

In this study, we proposed a framework to estimate and quantify both epistemic and aleatoric uncertainties in the context of automatic Deep Learning-based segmentation of four OARs for prostate cancer radiotherapy treatment planning. The epistemic uncertainty was successfully estimated by implementing Monte Carlo Dropout in a 3D U-Net. The PU-Net was trained to estimate the aleatoric uncertainty. Results in this scenario were limited due to the weak stochastic nature of the trained model caused by using Monte Carlo Dropout generated pseudo-labels. To quantify the estimated uncertainties, we calculated the standard deviation, predictive entropy and averaged ensemble KL divergence. This study achieved a complete estimation and quantification of the epistemic and aleatoric uncertainties, offering a comprehensive framework encompassing relevant aspects for the clinical application of Deep Learning based OAR segmentation for prostate cancer radiotherapy treatment planning.

Keywords: DL, automatic OAR segmentation, uncertainty

2601

Digital Poster

Guidelines-based automatic segmentation improvements