ESTRO 2021 Abstract Book

S402

ESTRO 2021

dose to the contralateral SCR region, again next to dose to the oral cavity.

Conclusion Next to the contralateral SCR region, sparing the oral cavity is important for prevention of xerostomia. The SCR region was more predictive for development of dayXER12M and physXER12M than the entire parotid gland. Our study results are in line with the physiology of the parotid glands, that are active during the day and during eating. Therefore, evaluation of strategies to preserve the parotid gland should be based on parotid specific endpoints. (Funded by KWF project 11350 / 2017-2) OC-0518 Impact of observer knowledge on AI delineation assessments: Bias in clinical acceptability testing J.W. Kim 1,2 , J. Marsilla 2 , J. Weiss 3 , D. Tkachuk 2 , J.K. Jacinto 4,5 , J. Cho 4,5 , E. Hahn 4,5 , S. Bratman 4,5 , B. Haibe- Kains 2,6 , A. Hope 4,5 1 Gangnam Severance Hospital, Yonsei University College of Medicine, Department of Radiation Oncology, Seoul, Korea Republic of; 2 University of Toronto, Department of Medical Biophysics, Toronto, Canada; 3 Princess Margaret Cancer Centre, University Health Network, Department of Biostatistics, Toronto, Canada; 4 University of Toronto, Department of Radiation Oncology, Toronto, Canada; 5 Princess Margaret Cancer Centre, University Health Network, Radiation Medicine Program, Toronto, Canada; 6 University of Toronto, Department of Computer Science, Toronto, Canada Purpose or Objective To determine if observer beliefs about the source of delineation creates unconscious bias in rating the quality of AI-delineated contours. Materials and Methods We trained a 3D UNET-based model to automatically contour 19 organs at risk (OARs) using radiation treatment planning data from 582 head and neck (HN) cancer patients treated at a large tertiary cancer center. A 80/10/10 split was used for training, validating and testing the model. To evaluate model quality, both manually delineated, gold standard clinical contours (human) and deep learning-based contours (AI) were presented to trained HN observers blinded to the true source of the contours using a publicly available web tool. Observers reported on their belief about the origin of the contour (“human”, “AI”, or “unknown”) and ranked clinical acceptability (5 for “perfect, not requiring editing” to 1 for “unusable for planning purpose”) for all OARs. Mixed effect regression modelling with a random intercept for each rater and OAR was utilized to assess the difference in rating between predicted “AI” and “human” for all contours and when contours were human and AI. Mixed effects models were also run for each OAR in ‘incorrect’ origin attributions to assess the