ESTRO 2021 Abstract Book

S769

ESTRO 2021

Conclusion The trained network showed good performance for the bladder. For the rectum, sigmoid and small bowel, the network should be optimized. However the geometrical errors occurred in the cranial-caudal direction where the inter-observer variability is known to be large. For the bladder, rectum and sigmoid the effect of these errors on the D2cc is limited. This is the first step towards a clinically applicable auto-contouring solution for MRI-based gynecological brachytherapy. PD-0928 Deep learning-based classification for standardization of prostate cancer RT structure annotations C. Jamtheim Gustafsson 1,2 , M. Lempart 1,2 , J. Swärd 3 , E. Persson 1,2 , T. Nyholm 4 , J. Scherman 1 1 Skåne University Hospital, Department of Hematology, Oncology and Radiation Physics, Lund, Sweden; 2 Lund University, Department of Translational Sciences, Medical Radiation Physics, Malmö, Sweden; 3 Lund University, Centre for Mathematical Sciences, Mathematical Statistics, Lund, Sweden; 4 Umea University, Department of Radiation Sciences, Umeå, Sweden Purpose or Objective Radiotherapy datasets can suffer from variations in annotation of organ at risk (OAR) and target structures. Annotation standards exist, but their description for prostate targets is limited. For example, it is not always defined whether the prostate gland, vesicles and/or lymph nodes exist together in the same or separate target structures, or what anatomy a CTV/PTV is referring to. This restricts the use of such data for machine learning as data must be correctly annotated. The purpose of this work was to develop a modality independent deep learning-based model for automatic classification and annotation of prostate radiotherapy DICOM structures. Materials and Methods The prostate OAR, support- and target structures (GTV/CTV/PTV), defined with and without separate vesicles