Room: Track 2
Purpose: This study aims to develop a novel method for generating synthetic relative proton-stopping-power (sRPSP) images from MRI with consistent-cycle Generative Adversarial Networks (ccGAN). Performance on MRI of individual sequences and different combinations was compared.
Methods: Planning CT and multi-sequence MR images of 151 brain tumor patients (aged 1 to 20 years, 105 males) were retrospectively analyzed. CT was converted into RPSP via the clinical calibration curve, serving as one of the model inputs. Rigidly registered to planning CT, T1-weighted (T1W), T2-weighted (T2W), and T2 FLAIR MRI were permutated to form 7 combinations for separate model training with RPSP. Compared to conventional cycle-GAN models, our ccGAN model introduces the consistent loss to provide strong constraints for enforcing the prediction closer to the target by using paired inputs. The training and testing sets consist of data from 126 randomly selected patients and the remaining 25 patients, respectively. The mean absolute error (MAE), absolute error (ME), peak signal-to-noise ratio (PSNR), structural similarity (SSIM) index were calculated between RPSP and sRPSP. Dosimetric difference was also calculated on 10 patients in the test group.
Results: Generating sRPSP (256×256×128) on a high-performance GPU cluster took 20s after one-week model training. The combination of T1W and T2W MRI achieved the highest PSNR (29.5±1.6) and SSIM (0.93±0.02) while T1W resulted in the lowest MAE (0.073±0.02). For the sRPSP generated from T1W MRI, the 2%/2mm gamma passing rate was 99.1%±1.5%. Differences in CTV V95 and D95 between proton dose distributions calculated on RPSP and sRPSP were 0.57%±0.64% and 0.07%±0.3%, respectively.
Conclusion: A novel ccGAN method has been developed for directly generating sRPSP from MRI for proton therapy in brain tumor patients. Despite that the combination of MR sequences may produce the best overall performance, individual sequences such as T1W appeared to be adequate in achieving accurate dosimetry.