ESTRO 2021 Abstract Book

S1535

ESTRO 2021

Conclusion We presented a framework for detailed mapping of surviving colonies following GRID irradiation together with predicted survival levels from homogeneous irradiation. For the given cell line, our findings indicate that GRID irradiation, especially at high peak doses, causes reduced survival compared to an open field configuration. Our novel colony segmentation procedure developed for the project will open for new applications where area per colony (not only colony as a dichotomous score) is used to further assess GRID effects. PO-1807 Automated detection of online auto-segmentation deviations by an independent segmentation algorithm M. Claessens 1,2 , P. Dirix 1,2 , D. Verellen 1,2 1 Iridium Network, Radiation Oncology, Antwerp, Belgium; 2 University of Antwerp, Integrated Personalized and Precision Oncology Network , Antwerp, Belgium Purpose or Objective Despite the implementation of deep learning segmentation models in radiotherapy departments to automate and optimize the delineation of tumour volumes and organs at risk, the quality of these structures is mainly verified manually by clinical experts. To facilitate this procedure, the effectivity of a secondary, independent auto-segmentation algorithm has been investigated to flag in advance inferior segmentations of an in-house trained model. Materials and Methods Five different organs (prostate, bladder, anorectum, femoral head left and right) were auto-delineated for 48 prostate patients both by an in-house trained (online, RayStation) model and an independent (offline, Mirada DLCExpert) algorithm (Fig. 1). Neither model was assumed correct, rather different quantitative comparison measures were calculated using the models’ segmentations and used as input features for a machine learning (ML) classification model to predict the quality of the online contours. This quality was assessed by an experienced radiation oncologist who categorised the online segmentations into four classes characterized by a time adjustment criterion: ‘from scratch, ‘5-10 min adaptations’, ‘<5 min adaptations’ or ‘no adaptations’. Subsequently, two different scenarios were set up investigating whether this approach could distinguish online segmentations that needed further adaptation (scenario 1) and make a distinction between the different estimated time-adjustment related classes (scenario 2).