ESTRO 2021 Abstract Book
S1401
ESTRO 2021
Figure 2: Measured mean displacement using LSP vs. mean applied displacement; 10 cases shown, each with different marker
Conclusion Displacement measured using LSP agreed well with artificially applied displacement for CBCT and CT image pairs. A cycleGAN was successfully used to create cross domain test pairs, able to demonstrate that LSP is robust to the types of shading and noise variation observed in CBCT data. PO-1677 cGAN-based pseudo-CT generation for prostate MRI-only radiotherapy S. Tahri 1 , C. Cadin 1 , H. Chourak 1 , A. Barateau 1 , S. Ribault 1 , O. Acosta 1 , P. Greer 2,3 , J. Dawling 4 , C. Lafond 1 , R. De Crevoisier 1 , J. Nunes 1 1 Univ Rennes- CLCC Eugène Marquis- INSERM- LTSI - UMR 1099, Laboratoire du traitement du signal et de l'image, Rennes, France; 2 Calvary Mater, Department of Radiation Oncology, , Newcastle, Australia; 3 University of Newcastle, School of Mathematical and Physical Sciences, Newcastle, Australia; 4 Australian e- Health Research Centre, Commonwealth Scientific and Industrial CSIRO, Herston/Queensland, Australia Purpose or Objective Pseudo-CT (pCT) generation from MR images is required for dose calculation in a MR-only radiotherapy workflow. Many deep learning methods (DLMs) have been developed for this purpose. The main advantage of DLMs is the very short calculation time of pCT generation. The aim of this study was to generate pCT with the highest precision, using a conditional generative adversarial network (cGAN) method (Pix2Pix) for prostate MRI-only radiotherapy planning. Materials and Methods For thirty-nine patients, T2-weighted MR images were acquired in addition to the planning CT images. pCTs were generated by the Pix2Pix model using the perceptual loss function and a ResNet 9 blocks generator. To tune the hyper-parameters of Pix2Pix, the mean absolute error (MAE) of Hounsfield units (HU) from voxel-wise comparisons between pCT and reference CT was used in the whole pelvis. The model was trained with a cohort of 25 patients, and was compared in a validation cohort of 14 patients. First, the learning rate (LR or step size) was tested within a range of 0.00001 to 0.07 (11 tests), and then with the LR that gives the lowest MAE, the beta value (called momentum) was tested within a range of 0.4 to 0.9 (5 tests). The beta 2 was set at 0.999, the batch size at 5, the architecture was set on Vanilla-GAN (Binary Cross Entropy loss), and a linear learning rate policy was used (method in which the learning rate decays linearly.) The final evaluation was performed on a 3-fold cross validation (25/25/25 patients in training, 14/14/11 patients in evaluation) with the optimal LR and beta values. Final pCT generation was performed with 50, 100, 200 and 300 epochs. Imaging endpoints were MAE, mean error (ME), mean absolute percentage error (MAPE) and peak signal-to- noise ratio (PSNR) from voxel-wise comparisons between pCT and reference CT in the whole pelvis and bones. Results In the whole pelvis, with different LR, the method showed significant variation of MAE values from 50.4 to 36.1 HU (Table 1). The lower MAE was obtained with a LR of 0.0006. The different values of beta (with LR= 0.0006) showed a variation of MAE from 92.6 to 36.1 HU, with a lower result for a beta value of 0.5 (Table 1). The final pCT generation with 3-fold cross validation with LR = 0.0006 and beta value = 0.5, provided MAE of 30.1 HU. No statistically significant differences were found for the number of epochs (Table 2).
Made with FlippingBook Learn more on our blog