Automatic Prediction of 3D Radiation Dose Distribution in Prostate Cancer Treated with Volumetric Modulated Arc Therapy (VMAT) Using a Conditional Generative Adversarial Network (cGAN)
Zhu S, Elshaikh MA, Movsas B, and Wen N. Automatic Prediction of 3D Radiation Dose Distribution in Prostate Cancer Treated with Volumetric Modulated Arc Therapy (VMAT) Using a Conditional Generative Adversarial Network (cGAN). Int J Radiat Oncol Biol Phys 2021; 111(3):e147.
Int J Radiat Oncol Biol Phys
Purpose/Objective(s): VMAT is a commonly used technique for the treatment of prostate cancer, but the planning process can be time-consuming due to the inherent complexity of the iterative optimizations. In this study, we aim to show the efficacy of a conditional generative adversarial network (cGAN) in automatically predicting the 3D radiation dose distribution trained with previously delivered plans.
Materials/Methods: In an IRB-approved analysis, we obtained the data of 167 patients with prostate adenocarcinoma who received definitive RT at our institution. All included cases were planned with VMAT technique with a prescribed dose of 78 Gy in 39 fx or 79.2 Gy in 44 fx to the prostate gland +/- seminal vesicles. For each patient, the inputs to the model were 6 structure masks (PTV and 5 OARs including bladder, rectum, penile bulb, and femoral heads), all resampled to voxel size of 2.5mm x 2.5mm x 3.0mm to be consistent with the grid size for dose, and the model predicts a 3D radiation dose distribution. The cGAN's generator was a modified 3D U-net, and its discriminator was a 5-layer 3D convolutional neural network. For training, mean absolute error (MAE) and binary cross entropy were used as the loss functions for the generator and discriminator, respectively, with Adam as the optimizers. Forty cases were randomly selected for testing only, and the remaining 127 formed the training set. The algorithm was implemented in Python 3.8 with PyTorch 1.7 as the framework. Model training was performed on NVIDIA Tesla V100 GPU. Model performance was assessed by comparing dosimetric metrics (based on institutional and RTOG 0415 guidelines) between the predicted doses (normalized to PTV maximal dose of 107%) and clinical doses with two-tailed paired t-test. Results: The model was trained for 300 epochs. On the independent test set, the mean MAE for the predicted 3D dose distribution was 1.6 Gy per voxel (range, 0.9 – 4.5 Gy). Dosimetric comparison (Table 1) demonstrated that the predicted doses had similar to improved PTV coverage and comparable dose to OARs, except for the femoral heads for which the model predicted higher D5's that still fell within dose constraint of < 50 Gy.
Conclusion: We demonstrated that, with the structures of the PTV and OARs as inputs, the 3D cGAN is excellent at the rapid prediction of radiation dose distributions of VMAT plans for prostate cancer that closely resemble clinically delivered doses. With further work on the deliverability of the predicted doses, our method can significantly increase the efficiency of treatment planning.