ESTRO 2021 Abstract Book

S1421

ESTRO 2021

2 National and Kapodistrian Univeristy of Athens, Medical Physics Laboratory, Medical School, Athens, Greece

Purpose or Objective Frame-based Gamma Knife radiosurgery is clinically implemented by (i) performing a patient MRI scan after attaching the Leksell stereotactic frame incorporating the N-shaped fiducials (MR-only workflow) or (ii) acquiring a CT scan with the frame and performing an anatomy-based co-registration with the MRI (MR/CT workflow). Moreover, (iii) the mean image method has been proposed for MR distortion correction in MR-only procedures (MR-corrected workflow). In all cases, the target(s) are registered to the stereotactic space by identifying the N-shaped fiducials. This phantom study evaluates and compares the localization uncertainties stemming from registration and inherent MR image distortions. Materials and Methods A custom-made acrylic-based spherical container was filled with 3D polymer gel dosimeter. The phantom was CT-scanned with the Leksell frame on. A treatment plan of 26 individual 4mm-shots was prepared and delivered. Shot centres directly defined in the stereotactic space served as reference target positions (Fig.1). The irradiated phantom was T2w-imaged with the frame on. An extra reversed gradient polarity MR series was acquired to implement the mean image distortion correction method. Targets were independently contoured in the GammaPlan TPS by exploiting the radiation-induced MRI contrast on all images relevant to the three workflows (Fig.1). TPS-calculated structures and transformation matrices were exported and processed using in-house routines. Target centroids were calculated on the dicom coordinate system and then co-registered to the stereotactic space by applying the TPS-calculated transformation(s), following the corresponding workflow. Calculated centroid locations were compared to reference positions for uncertainty evaluation.

Fig.1. Indicative target positions (low-signal areas due to radiation-induced polymerization) in a superior axial slice following the three different workflows. Green circles (10Gy isolines) represent corresponding reference positions. Results Spatial offsets were seen in all workflows (Fig.1). For MR-only, a 0.8mm median spatial uncertainty was estimated, with a max of 1.3mm and great dependence on target location. Reduced uncertainties (median 0.6mm, max 0.9mm) were calculated for the MR/CT workflow. The MR-corrected approach performed best with median/max uncertainties of 0.2mm/0.4mm. Conclusion For both clinically used workflows, target localization uncertainty may compromise treatment efficiency, especially for tiny lesions distant from isocenter. MR distortion correction can improve target localization accuracy in MR-only workflows. This research is co-financed by Greece and the European Union (European Social Fund-ESF) through the Operational Programme «Human Resources Development, Education and Lifelong Learning 2014-2020» in the context of the project “Assessment of spatial uncertainties in target determination related to MRI and their impact on stereotactic radiotherapy treatment planning in multiple brain metastases cases” (MIS 5047965) PO-1695 Accurate H&N 3D segmentation with limited training data using 2-stage CNNs E. Henderson 1 , E. Vasquez Osorio 1 , M. van Herk 1 , C. Brouwer 2 , R. Steenbakkers 2 , A. Green 1 1 The University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom; 2 University Medical Center Groningen, Department of Radiation Oncology, Groningen, The Netherlands Purpose or Objective Auto-segmentation of organs-at-risk (OARs) using convolutional neural networks (CNNs) is an active research area. In radiotherapy large, high quality datasets are scarce, but few methods work well with small sets of training data. A method to train accurate and robust head and neck (HN) 3D auto-segmentation models from scratch with a small dataset is presented. Materials and Methods Two CNN models are trained independently to perform 1) region of interest localisation and 2) 3D auto- segmentation. The localisation model is used to crop to a highly consistent region prior to auto-segmentation (shown in Fig. 1). CNN 1 is trained using 238 CTs from open data¹ to locate the centre-of-mass of the brainstem and parotid glands as an anchor point. CNN 2 is trained with just 34 CTs² for 3D segmentation of the mandible, brainstem, spinal cord and parotid glands. OAR delineations from two doctors are available, where one set of delineations is selected as golden standard and used for training. Model segmentation performance is compared with the inter-observer variation of the two doctors using the 95th percentile Hausdorff distance (HD95) and mean distance-to-agreement (mDTA). The generalisability of the method is tested in an external cohort³ of 6 patients imaged at 2 time points with segmentations of the spinal cord and parotid glands