ESTRO 2024 - Abstract Book

S3114

Physics - Autosegmentation

ESTRO 2024

Figure 1 depicts a bar plot of the epistemic and aleatoric uncertainties for the four OARs quantified by the standard deviation, predictive entropy and averaged ensemble KL divergence. For a fair comparison, the MCDO and PU-Net models were trained on the same fold. The epistemic uncertainty is larger in magnitude compared to the aleatoric uncertainty for every metric and OAR, although with a relatively smaller difference for the predictive entropy. In all cases, the rectum and bladder have the highest uncertainties, as expected since they have the largest surface and inter-organ boundaries. Figure 2 depicts uncertainty maps for two patients, which show the anatomical location of the highest sources of uncertainty, represented by red colors. For the two patients, both the aleatoric and epistemic uncertainty maps indicate that inter-organ boundaries correspond to the highest sources of uncertainty. The aleatoric uncertainty maps (second and fourth rows in Figure 2) show a high amount of scattered pixels for the standard deviation and KL divergence quantifications, with more defined uncertainty regions for the entropy.