ESTRO 2022 - Abstract Book

S689

Abstract book

ESTRO 2022

Conclusion This work exhibits the efficacy of uncertainty estimation for HNSCC GTV-DL auto-segmentation. Using the estimated uncertainty, the uncertainty regions disclose potential erroneous regions of predictions. The feasibility of using uncertainty regions for patient-level failure detection was also demonstrated with a primitive approach. This study contributes substantially to the clinical applicability of DL-based GTV segmentation.

OC-0772 Deep learning-based 4D synthetic CT for lung radiotherapy

M. Maspero 1 , K. Keijnemans 1 , S.L. Hackett 1 , B.W. Raaymakers 1 , J.J.C. Verhoeff 1 , M.F. Fast 1 , C.A.T. van den Berg 1

1 UMC Utrecht, Radiotherapy, Utrecht, The Netherlands

Purpose or Objective Synthetic-computed tomography (sCT) generation is crucial to enable MR-only radiotherapy (RT) and accurate MR-based dose calculations[1]. To date, sCT generation was scarcely performed in the thoracic area[2,3]. Obtaining sCT in a region strongly affected by breathing motion is crucial to facilitate adaptive MR-guided radiotherapy (MRgRT), possibly reducing the time from patient positioning to irradiation and enabling dose accumulation based on 4D MRI. Also, dose accumulation requires fast respiratory-sorted sCT generation (<1min)[4,5]. Recently, convolutional neural networks (CNNs) were able to generate sCTs (< 20 s) quickly[6,7]. However, no previous work focused on the generations of sCT for anatomies affected by respiratory motion. This work aims to assess the feasibility of sCT generation for patients with thoracic cancer using 4D MRI sorted according to the respiratory phase and investigate the dosimetric accuracy of MR-based calculation with these sCTs. Materials and Methods Thirteen patients undergoing lung radiotherapy were considered in this study. Patients underwent a planning 4D-CT, and a 4D-MRI was acquired with a simultaneous multi-slice (SMS) sequence obtaining ten respiratory phases and a midposition. Additional 98 thoracic patients imaged with 3D CT and a 2D turbo spin-echo MRI were considered to enlarge the training set. Patients were split in train (4D/3D=7/70)/validation (2/7) /test (4/11) sets. A 2D reversible adversarial network (revGAN) was used to learn paired mapping from MRI to CT. Before training, CT images were rigidly registered to MR images ( Fig1 ). Training of revGAN was performed in the three orthogonal planes and over ten respiratory phases and midposition for 30 epochs. After hyperparameter optimisation on the validation set, the network was inferred in the three orthogonal directions to the four patients (test), producing three sCTs per patient. A voxel-wise median was calculated for each patient, obtaining a combined sCT [8]. The midposition sCTs in the test were evaluated against the midposition planning CT after matching body contours. Dose recalculation of clinical plans was performed on sCTs in Monaco (Elekta AB, 3mm grid) with 1.5 T magnetic field. Dose distributions were analysed through voxel-based dose differences and gamma-analysis.