ESTRO 2020 Abstract book

S995 ESTRO 2020

carotid artery), dyspnoea (due to light stenosis) and chest pain.

PO-1709 Automated organ delineation in T2 head MRI using combined 2D and 3D convolutional neural networks L. Rusko 1 , B. Kolozsvari 1 , P. Takacs 1 , B. Darazs 2 , R. Czabany 2 , B. Gyalai 2 , S. Kaushik 3 , C. Cozzini 3 , V. Paczona 4 , K. Hideghety 4 , H. Mccallum 5 , S.F. Petit 6 , J. Kleijnen 6 , J.A. Hernandez Tamames 6 , F. Wiesinger 3 1 GE Healthcare, Hungary, Budapest, Hungary ; 2 GE Healthcare, Hungary, Szeged, Hungary ; 3 GE Healthcare, Germany, Munich, Germany ; 4 University of Szeged, Department of Oncotherapy, Szeged, Hungary ; 5 Newcastle University, Northern Institute for Cancer Research, Newcastle, United Kingdom ; 6 Erasmus MC, Department of Radiotherapy, Rotterdam, The Netherlands Purpose or Objective T2 weighted MRI images are frequently used for organ delineation in RT planning due to their excellent soft‐ tissue contrast, especially in the head and neck region. To reduce the long time needed for manual contouring of high‐resolution images this work proposes a deep‐learning approach that uses a combination of 2D and 3D convolutional neural networks (CNN) to contour organs of the head and neck in MRI images automatically. The accuracy of the presented approach is demonstrated on a The AAPM RT‐MAC Grand Challenge 2019 published 31 T2 head MRI exams with manually contoured (left/right) parotid (PG) and submandibular (SMG) glands. This image dataset (see cancerimagingarchive.net) was split to 25/6 cases to train/test the presented deep‐learning approach. The proposed solution is a combination of 2D and 3D CNNs. It involves three 2D models which first localize the organ in axial, coronal, and sagittal slices. Subsequently, it involves a 3D model that segments the organ within its 3D bounding box. The center of the bounding box is determined based on the fused output of the 2D models and the size of it is defined for each organ separately. As preprocessing, the MRI image is resampled to uniform resolution (256 3 ) and voxel size (1 mm 3 ) followed by histogram‐based min‐max intensity normalization. For sake of efficiency, the 2D localizer models work at low (128) resolution, while the 3D model works at original (256) resolution, but it is limited to the bounding box (the max dimension of it is less than 100 voxels). Both 2D and 3D models follow the U‐Net architecture [arXiv:1505.04597] involving 4 and 3 pooling layers, respectively. As post‐processing, the largest connected component of the binarized result is taken and up‐sampled to original image resolution. For each organ (PG/SMG, left/right) separate 2D and 3D models were trained using the same (train/test) separation of cases. Each organ model was evaluated, such that for each test case the result was compared with the ground‐truth using DICE‐ score (%) and surface‐distance (SFD, mm) accuracy measures. In addition, the rate (SFV‐2, %) of surface voxels which are located within 2 mm of the ground‐truth contour were also computed. Results public image database. Material and Methods

Conclusion The designed automatic tool could totally detect patients with calcified plaques in heart if a good training in contouring is made. This tool could be used side to try to minimise dose to heart if possible and on the other hand to do a screening and follow‐up of the patients with high risk of developing adverse cardiac events. Acknowledges: This work was financed by the Spanish Association Against Cancer (AECC).