GANs covert CBCT to CT for head-neck, lung and breast: paired vs unpaired; single-site vs generic
Maspero, M., Savenije, M. H. F., Van Heijst, T. C. F., Kotte, A. N. T. J., Houweling, A. C., Verhoeff, J. J. C., Seevinck, P. R., Van den Berg, C. A. T.
Radiotherapy & Oncology
133
p. S1105-S1106
Abstract
Purpose or Objective CBCT offers a representation of daily anatomy that may be used for online dose calculation and adaptation. However, dose calculations cannot be performed on CBCT due to lack of HU calibration, limited FOV and presence of image artefacts. This study investigates the use of generative adversarial networks (GANs) to convert CBCT into CT. Such networks allow very fast image conversion and thus facilitate online adaptation. However, CBCT to CT conversion using paired learning is problematic given the possible anatomical interscan differences breaking the “paired” (PA) assumption of data consistency. Here, we investigate the use of unpaired (UP) training, which obviates the need for consistent pairwise datasets (e.g. registered, anatomically matched) in the training. In this explorative study, we investigated the performance of PA vs UP learning and compared site-specific vs generic trained network. Material and Methods CBCTs of 88 patients diagnosed with head-neck (HN, 31), lung (29) and breast (28) cancer undergoing radiotherapy were rigidly registered according to the clinical procedure and resampled to the planning CT (XVI, Elekta). PA vs UP: To perform PA training, we used a conditional GAN (cGAN), while for UP training, we used a cycle-consistent GAN (cycleGAN). The two networks were trained in 2D transverse planes mapping CBCT to CT Hounsfield Units. For each anatomical site, the networks were trained on 15 patients (training set) and evaluated on the remaining patients (test set). Single vs generic: To verify whether a single network could be used for all the patients independently of the anatomical site, we trained both the cGAN and the cycleGAN on the data of 45 patients. Image comparison in terms of mean absolute error (MAE) and mean error (ME) in the FOV of the CBCT vs planning CT was performed on the test set for both the experiments. Results cGAN and cycleGAN training required about 1 and 5 days, respectively, on a GPU Tesla P100 (NVIDIA). Forward evaluation took about 20 s. PA vs UP: For all the three sites, discontinuities between 2D transverse slices were more visible after PA compared to UP training (Fig1). In the case of UP training, some residual image artefacts were present in the transverse plane, especially for the breast cases. Mean MAE and ME were for UP and PA are reported in Fig2. Single vs generic: For all the anatomical sites, training with all the patients resulted in mean MAE and ME within 1σ respect to training on patients of each site (Fig2). Conclusion Visually, UP training resulted in slightly better image quality although residual artifacts were present. Quantitatively PA and UP were comparable; however, a fundamental problem in the evaluation is the lack of a good reference considering that anatomical changes between CT and CBCT may have taken place. Investigations to verify the accuracy of dose calculations is still needed to justify the use of GANs to enable CBCTbased dose calculations. In general, the use of a generic network for all the sites seems to be a viable option and the time necessary to convert CBCT into CT justifies the use of GANs for online ART.