ESTRO 2020 Abstract book

S987 ESTRO 2020

sCT in the test were evaluated with the planning CT. Dose recalculation of clinical plans was performed on sCTs in the Monaco TPS (Elekta AB, 2mm grid). Dose distributions were analysed through voxel-based dose differences and γ-analysis.

indicate marginally better than 3; 3 points indicate no effect of signal void on diagnosis; 2 points indicate slightly poor than 3; and 1 point indicates the lack of evaluation. High score regarded clinically useful.

Results

Mean effect of markers on DWI measured by the radiation oncologist and medical physicist was 4.3 (Standard Deviation (SD) 1.3, range 2–5) points and 4.0 (SD 1.4, range 3–5) points, respectively. The sample is shown. Conclusion Gold markers showed little effect on the quality of DWI. Therefore, despite using iron-containing markers and the size of marker < 0.5 mm being available, MRI, particularly DWI, may be used during follow-up imaging. PO-1698 Deep learning-based sCTs with uncertainty estimation from heterogeneous pediatric brain MRI M. Maspero 1 , L.G. Bentvelzen 1 , M.H.F. Savenije 1 , E. Seravalli 1 , G.O.R. Janssens 1 , C.A.T. Van den Berg 1 , M.E.P. Philippens 1 1 UMC Utrecht, Radiation Oncology, Utrecht, The Netherlands Purpose or Objective To enable MR-only dose planning and accurate MRI-based dose calculations, synthetic-CTs (sCTs) need to be generated. Recently, convolutional networks have been proposed as a general-purpose solution to image-to-image translation problems enabling MRI-based dose calculations. sCTs used in current clinical practice are generated by fixed MRI protocols, which restricts developments on MRI protocols and cross-platform usage. To date, it is unclear how a network would perform with variable imaging protocols, and, additionally, how the quality of sCT can be verified in daily clinical practice. This work aims at assessing (1) the feasibility of dose calculations from MRI acquired with a heterogeneous set of acquisition protocols (2) proposing a combination of networks trained on orthogonal planes to estimate the uncertainty of the generated sCT. Material and Methods All pediatric patients (pts) undergoing brain radiotherapy between 2015/07-2019/05 at our department were considered in this study. Patients underwent a planning CT at our department, but MRI could have been acquired in other departments or centres. Inclusion criteria were the acquisition of a 3D spoiled gradient-recalled echo T1w MRI resulting in a total of 52 pts. Imaging protocols varied among pts (Tab1), and data heterogeneity was maintained in train/validation/test sets. A conditional generative adversarial network (cGAN) was used to learn paired mapping from MRI to CT. Before training, CT images were rigidly registered to MR images (Fig1). Training of cGAN was performed on 30 patients in the three orthogonal planes for 30 epochs (18-36 hrs on a GPU, NVIDIA). The three separately trained networks were applied to the remaining 10+12 pts (validation+test), producing three sCTs per patient. For each patient, median and std of the three views were calculated for each voxel obtaining a combined sCT and an uncertainty map, respectively. The

Results Applying the trained cGAN to the three planes of a single patient (Fig1) required about 20s. A mean absolute error of 75±16HU (mean±1σ, range:42-86) was obtained in the intersection of the body contours between CT and sCT. The combined multi-plane sCTs performed better than sCTs from any single plane. Uncertainty maps highlighted that sCT have low fidelity at the body contours and air cavities. A dose difference of -0.1±0.4% was obtained on the D>50% of the prescribed dose, and mean γ-2%,2mm pass rate of 99.2±1.5% (Tab1).

Conclusion Accurate MR-based dose calculation using a combination of three orthogonal planes for sCT generation is feasible for pediatric brain cancer patients, even when training on