Dose Prediction Deep Learning-Based Model for VMAT of Prostate Cancer Applying Magnetic Resonance Image (MRI) in Versa HD Linear Accelerator

Background: Prostate cancer patients are commonly undergoing Radiotherapy (RT) and treatment planning system have a prominent role for dose calculation, while this would seem that dose distribution uncertainties of treatment planning system (TPS) may effect on RT results. Therefore, this study aimed to design a Dose prediction deep learning-based model for prostate cancer volumetric arc therapy (VMAT) applying MRI in Versa HD linear accelerator (linac). Materials and Methods: In this work, MRI of 45 patients who underwent VMAT was acquired, and cycle-consistent GAN (CycleGAN) (that allow image-to-image translation) and U-net deep learning (DL) framework for prostate were employed. The synthetic CT (sCT) images were generated from MR images. The predicted dose among CycleGAN, U-net and Monaco TPS (that calculate dose distribution based on CT simulation images) was compared to each other. Results: The sCT that was generated employing CycleGAN illustrated more obvious boundaries than the sCT of U-net (sCTU-net). The gamma passing rate of cycleGAN and U-net was exceeded 97% and 90%, respectively, in all areas. Conclusion: The results of this study illustrates that deep learning models including CycleGAN and U-net are good alternative for dose prediction of VMAT in Versa HD linac, while it seems that CycleGAN may be more accurate compared to U-net.