ESTRO 2021 Abstract Book

S1381

ESTRO 2021

Conclusion Our atlas-based workflow has proved to be effective for the clinical workflow since it reduces contouring time and improves reproducibility of contours. The result of this study should encourage each institution to creates its own. PO-1660 Investigating the generation of synthetic CT for abdominal tumors treated with particle therapy G. Meschini 1 , D. Calabrese 1 , F. De Mori Bajolin 1 , A. Vai 2 , G. Fontana 3 , S. Molinelli 2 , A. Pella 3 , S. Imparato 4 , V. Vitolo 5 , A. Barcellini 5 , E. Orlandi 5 , C. Paganelli 1 , G. Baroni 1,3 1 Politecnico di Milano, Department of Electronics, Information and Bioengineering, Milan, Italy; 2 National Center for Oncological Hadrontherapy, Medical Physics, Pavia, Italy; 3 National Center for Oncological Hadrontherapy, Bioengineering, Pavia, Italy; 4 National Center for Oncological Hadrontherapy, Radiology, Pavia, Italy; 5 National Center for Oncological Hadrontherapy, Radiotherapy, Pavia, Italy Purpose or Objective In particle therapy of abdominal tumors, the use of synthetic Computed Tomography (sCT) generated from 3D Magnetic Resonance (MR) imaging can support treatment monitoring and adaptation, by verifying inter- fraction changes and reducing the non-therapeutic dose provided by repeated CT scans. Materials and Methods The end-exhale CT and the 3D breath-hold T1-weighted VIBE MR acquisition were available for 14 patients with abdominal lesions, for a total of 22 scans and 1366 transversal slice pairs. Preprocessing included rigid and deformable image registration, air-bubbles overwriting, MRI denoising and data augmentation (x8). A conditional Generative Adversarial Network (cGAN) was trained on slices from 16 volume pairs and validated on 3 different volumes (166 slices). The use of both single and three input channels (air, bones, soft tissue) was evaluated in terms of mean absolute error (MAE) and Dice coefficient of a clearly identifiable structure (kidney). Six-fold cross validation was performed, then the metrics were computed on 3 patient volumes (192 slices) considered as testing dataset. Results On the validation set, the MAE (mean±standard deviation) was 68±8 HU and 56±10 HU, whereas the mean Dice coefficient was 0.84 and 0.88 in the single and three channels configurations, respectively. Due to the higher accuracy of the three channel cGAN, this was cross validated (MAE 66±4 HU) and tested. The accuracy on the test set (MAE 75±3 HU) was comparable to other published methods for sCT generation. A qualitative result is illustrated in the figure, where the sCT generated with the three channels cGAN (3-ch.) performed better than that derived with one channel (1-ch), showing proper matching of bones (yellow arrow) and air bubbles (pink arrow) and slight blurring of soft tissues. This suggests sCT feasibility for accurate dose recalculations in particle therapy, as dose variations are mainly due to high density gradients along the beam path. Increasing the training dataset is expected to improve soft tissues contrast in the generated sCTs.