VQ-DoseNet: A vector quantized model for stochastic radiotherapy dose prediction

Radiotherapy treatment planning is a time-consuming process, and dose prediction models are crucial for improving efficiency. Traditional deep learning-based models often produce deterministic outputs, which fail to capture the inherent variability in dose distributions required for clinical acceptance. To address this limitation, we propose a novel vector quantized model (VQ-DoseNet) to introduce stochasticity in the dose prediction process. This method explicitly captures the probabilistic nature of dose variations by perturbing input features, generating multiple plausible dose distributions that reflect observed clinical variability. Experiments demonstrate the superiority of our model compared to state-of-the-art (SOTA) approaches, achieving a mean absolute error (MAE) of 0.106 Gy on in-house dataset and dose and DVH scores of 3.608 ± 1.267 Gy and 1.329 ± 1.934 Gy on OpenKBP dataset, respectively. Multiple stochastic predictions on in-house and OpenKBP datasets demonstrated that the model consistently generates dose distributions in close agreement with the ground truth, with DVH parameters remaining within clinical constraints. These results highlight the potential of our approach to maintain high prediction accuracy while incorporating dose variability, offering a promising avenue for enhancing treatment planning in radiotherapy.