ESTRO 2021 Abstract Book

S1398

ESTRO 2021

Conclusion We adapted STAPLE to identify consistently biased observers in deep learning training data. We showed that auto-segmentation models, trained to segment skeletal muscle at the L3 vertebral level, were unaffected by the presence of biased observers. Using multiple observers in the training set led to models that were not significantly different to those trained on single expert data. Finally, we found that reducing the number of images in the training data while increasing the number of delineations per image did not affect accuracy. Our results therefore suggest a potential route towards cheaper data curation, while maintaining model performance, by employing multiple trained observers delineating the same small image set instead of using more expensive expert annotators on a larger dataset. PO-1674 An MRI sequence independent Convolutional Neural Network for head sCT generation in proton therapy B. Knäusl 1 , L. Zimmermann 1,2 , M. Stock 3 , C. Lütgendorf-Caucig 3 , D. Georg 1 , P. Kuess 1 1 Medical University of Vienna, Department of Radiation Oncology, Vienna, Austria; 2 University of Applied Sciences, Faculty of Engineering, Competence Center for Preclinical Imaging and Biomedical Engineering , Wiener Neustadt, Austria; 3 Medaustron Center for Ion Beam Therapy and Research, Medical, Wiener Neustadt, Austria Purpose or Objective A magnetic resonance imaging (MRI) sequence independent deep learning technique was developed and validated to generate synthetic computed tomography (sCT) scans for MR guided proton therapy. Materials and Methods 47 meningioma patients previously undergoing proton therapy based on pencil beam scanning were divided into training (33), validation (6), and test (8) cohorts. T2, T1, and contrast enhanced T1 (CM) MRI sequences were used in combination with the planning CT data to train a 3D U-Net architecture with ResNet-Blocks. A hyperparameter search was performed including two loss functions, two group sizes of normalisation, and depth of the network. Training outcome was compared between models trained for each individual MRI sequence and for all sequences combined. The performance was evaluated based on a metric and dosimetric analysis as well as spot difference maps. Furthermore, the influence of immobilisation masks which are not visible on MRIs is investigated. Results Based on the hyperparameter search, the final model was trained with fixed features per group for the group normalisation, six down-convolution steps, an input size of 128x192x192, and feature loss. For the test data set for body/bone the mean absolute error (MAE) values were on average 82.9/236.4 Hounsfield units (HU) when trained using T1CM images, 79.8/216.3 HU for T1, and 71.1/186.1 HU for T2. The structural similarity metric (SSIM) ranged from 0.95 to 0.98 for all sequences. The investigated dose parameters of the target structures agreed within 1% between original proton treatment plans and plans recalculated on sCTs. The spot difference maps had peaks at +/-0.2 cm and for 98% of all spots the difference was less than 1 cm. Conclusion A novel MRI sequence independent sCT generator was developed, which suggest that the training phase of neural networks can be disengaged from specific MRI acquisition protocols. In contrast to previous studies, the patient cohort consisted exclusively of actual proton therapy patients (i.e., “real world data”). PO-1675 Automated delineation for MR-only prostate radiotherapy using a 2.5D convolutional neural network C. Holland 1 , J. Wyatt 1 , R. Pearson 2 , T. Wintle 2 , R. Maxwell 1 1 Newcastle University, Institute for Translational and Clinical Research, Newcastle, United Kingdom;