Purpose: Accurate prediction of radiotherapy dose distributions could guide the treatment planning process to rapidly and consistently achieve high-quality plans. In this study, we proposed a novel 3D DenseNet for predicting dose distributions of volumetric modulated arc therapy (VMAT) for prostate cancer patients.
Methods: 75 prostate cancer patients were included. Each patient was planned with a prescription dose of 40 Gy in 5 fractions using Eclipse. We built a 3D DenseNet to generate patient-specific dose distributions from the concatenation of the planning CT image, the contours of the external body, planning target volume (PTV), and organs-at-risk (OARs). The proposed DenseNet contains five dense blocks that could alleviate the vanishing gradient problem, encourage feature reuse, and reduce the number of model parameters. The patient cohort was randomly split into 53 patients for training, 11 patients for validation, and 11 patients for independent testing. A 3D U-Net was also trained for dose prediction and model comparison. The model performance was evaluated by comparing the predicted dose with the clinical dose using Dice similarity coefficients (DSCs) of isodose volumes and relevant dosimetric metrics.
Results: The average DSCs of isodose volumes, between 0% and 100% (step size: 1%) of the prescription dose, achieved by the 3D U-Net and 3D DenseNet are 0.939 and 0.948, respectively. The average absolute differences in Dmean of the PTV and OARs are under 1.43 Gy (UNet) and 0.89 Gy (DenseNet). The average absolute difference in PTV V40Gy is under 2% for both models.
Conclusion: The proposed 3D DenseNet could accurately predict the dose distribution of VMAT for a prostate cancer patient in about 8 seconds. It achieved more accurate dose prediction than the state-of-the-art 3D U-Net. The 3D DenseNet is a promising tool for fast dose prediction and guiding plan optimization in an online adaptive workflow.