ESTRO 37 Abstract book

S218

ESTRO 37

the surface curvature are included globally per BC. That means a total of 8×2×3+4 = 52 features per BC are used to train and evaluate the RFC. Ground truth annotations are binary per BC: Class 1 means that the BC fits the bladder and the observer is certain about it (positive cases). Class 2 means that either the BC does not fit the bladder or the observer is not certain if the BC location is correct (negative cases). All 844 scans with BCs were annotated by three observers each. Inter-observer variability was examined with the Fleiss’ Kappa measure with a resulting value of 0.51, indicating a moderate agreement. The available three annotations per BC were joined by means of majority voting. RFC performance was measured by cross validation with 40 folds on the majority voted annotations. Results The resulting RFC yields an area-under-curve (AUC) of 0.93 (cf. Fig. 1 for the corresponding ROC curve). See Fig. 2 for positive examples of BC classification.

Conclusion There is variation in the provision of guidance for OAR delineation in UK radiotherapy clinical trials across anatomical categories. Where guidance exists, there are some inconsistencies in OAR outlining definitions, delineation instructions, selection and nomenclature between trials. Besides using standardised nomenclature to avoid mislabelling of OARs, there is a need to encourage the use of published consensus delineation guidelines in radiotherapy trial documentation to improve the provision of instructions and standardise OAR delineation. OC-0422 A novel DIR quality assessment approach for daily clinical radiotherapy routine A. Derksen 1 , L. König 1 , N. Papenberg 1 , T. Gass 2 , B. Haas 2 , C. Behrens 3 1 Fraunhofer MEVIS, Fraunhofer MEVIS, Lübeck, Germany 2 Varian Medical Systems, Varian Medical Systems Imaging Laboratory GmbH, Baden, Switzerland 3 Herlev and Gentofte Hospital- University of Copenhagen, Department of Oncology R Radiotherapy, Herlev, Denmark’ Purpose or Objective This work focuses on increasing radiotherapists’ confidence in deformable image registration (DIR). The goal is to create a tool that automatically assesses a given DIR result by providing the user with feedback regarding sanity or plausibility of DIR-propagated contours. When it comes to assessing DIR quality, a central problem is the inherent lack of ground truth measurements in a clinical setting. Therefore, we have focused on creating indicators for DIR plausibility that can be evaluated by only using image and deformation information as well as already available planning CT contours. Material and Methods We focus on bladder contours (BC) due to their crucial role in achieving accurate DIR results in the pelvic region. Data from three sites was incorporated in algorithm development and evaluation. In total 844 CBCT scans with BC were included from three different clinical sites (686, 130 and 28 scans from site 1, 2 and 3, respectively). The general idea we pursued is: a set of features is extracted per CBCT and propagated BC pair, which is subsequently used to train a random forest classifier (RFC) to distinguish between plausible and non-plausible propagated BCs. Computed features do encode local BC information by a decomposition of the propagated contour in eight sub- contours that are divided by the octants of an axis aligned coordinate system that is placed in each BCs center of mass, i.e. features are computed per sub- contour. Given the eight sub-contours, image gradient mean and std. are sampled at three locations (±10 mm in normal direction of the BC and directly on the BC) per sub-contour. Furthermore, mean, std., min, and max of

Conclusion A novel DIR evaluation approach based on data already available for daily workflows and a RFC is presented and evaluated. Though the trained classifier performance is probably limited by the global decision per BC (in contrast to a local per sub-contour decision) obtained results are still encouraging with an AUC of 0.93. OC-0423 An evaluation of variabilities in organs-at-risk delineation for MR-only head and neck radiotherapy K.Y. Chui 1 , W.W.K. Fung 1 , J. Yuan 2 , A.W.L. Mui 1 , G. Chiu 1 1 Hong Kong Sanatorium & Hospital, Department of Radiotherapy, Happy Valley, Hong Kong SAR China 2 Hong Kong Sanatorium & Hospital, Medical Physics and