- (1) University Hospital, LMU Munich, Department of Radiation Oncology, Munich, Germany
- (2) Aarhus University Hospital, grid.154185.c, Central Denmark Region
- (3) University Medical Center Utrecht, grid.7692.a
- (4) Ludwig Maximilian University of Munich, grid.5252.0
Purpose This study investigates for the first time the feasibility of using deep learning for cone-beam CT (CBCT) intensity correction to enable accurate daily dose calculation and treatment adaptation in volumetric-modulated arc therapy (VMAT) and intensity-modulated proton therapy (IMPT). Current CBCT intensity correction approaches often show a lack of either speed or accuracy, which might be overcome by deep learning approaches. Methods Pre-treatment CBCTs and corresponding projections of 30 prostate cancer patients were considered. A previously validated technique for CBCT intensity correction, based on deformable image registration (DIR) of the planning CT to the daily CBCT and scatter estimation in projection space, served as reference (CBCTcor) . Two alternative methods were investigated: A U-shaped deep convolutional neural network (U-Net) was trained to perform scatter correction in projection space, i.e., going from measured to corrected projections before reconstruction (CBCTScatterNet). Moreover, a generative adversarial network (GAN) was trained to perform a translation from the original CBCT (CBCTorg) to CBCTcor in image space, generating a so-called CBCTcorGAN. CBCTScatterNet and CBCTcorGAN were compared to CBCTcor in terms of mean absolute error (MAE) and mean error (ME). For eight exemplary patients, dose calculation accuracy in VMAT and IMPT was evaluated with respect to CBCTorg. Results Both, CBCTScatterNet and CBCTcorGAN, showed a substantially improved agreement to CBCTcor compared to CBCTorg. Mean MAE and ME decreased from 158HU and 152HU for CBCTorg to 39HU and 4HU for CBCTScatterNet and 57HU and −2HU for CBCTcorGAN, respectively. In a 2% dose-difference test, considering only voxels above 50% of the prescribed dose, mean pass-rates were 53% and 64% for CBCTScatterNet and CBCTcorGAN in IMPT. In VMAT, pass-rates of 90% and 97% were obtained for CBCTScatterNet and CBCTcorGAN using a 1% dose-difference criterion. Conclusions CBCT intensity correction using two different implementations of deep learning was found feasible. For VMAT, dose calculation accuracy was high, while for IMPT further improvements may be required. Compared to the reference correction method, deep learning techniques were less affected by DIR inaccuracies and allowed considerably faster CBCT correction within few seconds instead of minutes.