ESTRO 38 Abstract book

S267 ESTRO 38

Purpose or Objective In adaptive photon and proton radiotherapy (RT and PT) it is desirable to utilize pre-treatment CBCT images for online dose calculation and adaptation. Due to various imaging artifacts CBCTs are, however, not suitable for accurate dose determination and current correction methods lack speed for online application. This work aimed at investigating the feasibility of using a cycle- consistent generative adversarial network (cycleGAN) for CBCT intensity correction to enable fast and accurate RT and PT dose calculation. The dedicated loss-function of the network allowed for using unpaired (not registered) training data despite inter-scan anatomical deviations. Material and Methods Planning CT (pCT) and daily CBCT imaging data of 25 prostate cancer patients were used for training in this study. A cycleGAN was trained using 18/25 patients and 4- fold cross-validation, aiming at translating the uncorrected CBCT images (CBCT org ) to a pCT-like image (CBCT cycleGAN ) in 2D (slice-by-slice). No initial matching of the data was performed (unpaired). A previously validated CBCT correction method (CBCT cor ) applying corrections in projection space on the basis of a prior virtual CT, obtained from pCT to CBCT deformable image registration (DIR), served as reference. CBCT cycleGAN was compared to CBCT cor in terms of mean Hounsfield Unit (HU) error within the body outline. Dose calculation accuracy was evaluated using volumetric modulated arc therapy (VMAT) photon and opposing single field uniform dose (OSFUD, 90°/270° gantry angle) proton plans generated on CBCT cor and recalculated on CBCT cycleGAN . Single-sided SFUDs were utilized to compare the proton range in beam’s eye view (BEV). Results The average (over all patients) HU error comparing CBCT cor and CBCT cycleGAN was -0.5 HU. In comparison, CBCT org showed an average deviation of 33HU with respect to CBCT cor (Fig. 1). For VMAT the average pass-rates for a 2%/1% dose-difference criterion were 100%/94%. For the OSFUD plans the 2% pass-rate was lower, at 77% (Fig. 2). Using a 2%/2mm gamma criterion the pass-rate increased to 94%. In terms of the proton range 87% of all analyzed BEV profiles agreed better than 3mm on CBCT cor and CBCT cycleGAN . The average range difference was 0.4mm. Application of the cycle consistent GAN allowed for a considerable increase in speed: time to correct a 3D CBCT was reduced from 8-10 min (CBCT cor ) to about 3 s (CBCT cycleGAN ). Conclusion Our study demonstrated for the first time the feasibility of using a cycleGAN and unpaired training for CBCT intensity correction. Results suggest high RT dose calculation accuracy. In PT, agreement to the reference CBCT cor was reduced due to the sensitivity of the proton range on HU values. The substantial speed-up with respect to the reference method renders CBCT cycleGAN particularly interesting in the scope of online adaptive therapy approaches. Due to unpaired training the method is independent from anatomical inconsistencies in the training data. Acknowledgements: German Cancer Aid, DFG-MAP, IMDI

OC-0514 Unsupervised deep learning for fast and accurate CBCT to CT deformable image registration S.R. Van Kranen 1 , T. Kanehira 1 , R. Rozendaal 1 , J. Sonke 1 1 Netherlands Cancer Institute, Radiation Oncology, Amsterdam, The Netherlands Purpose or Objective CBCT based online adaptive radiotherapy requires fast deformable image registration (DIR) to facilitate plan adaptation within a reasonable timeslot. Deep learning convolutional neural networks (CNNs) have demonstrated fast and accurate alignment of MR datasets of brain images[1]. The goal of this study is to train such networks for head & neck CBCT-to-CT DIR and to compare the accuracy with our clinically implemented bSpline method. Material and Methods VoxelMorph [1], a U-net based CNN with a spatial transformation block for image registration, was modified to capture deformations up to 3 cm. Networks were trained unsupervised on 7899 CBCT-CT scan pairs of 232 patients. The loss function consisted of local cross correlation regularized on the bending energy to enforce smooth deformations. Several trainings were performed to determine the optimal weighting (λ=0.1-10). An independent set of 10 CBCT-CT pairs was regularly evaluated during training to determine convergence (no improvement during 200 iterations). Training typically took 1.5 days on a single GTX 1080 GPU.