Multitask Prediction of Radiotherapy Dose and Angle Based on Neural Network
Abstract
Purpose
Accurate radiotherapy dose prediction largely depends on beam configuration, but most deep learning-based dose prediction models rely on explicit beam Angle input, which is not feasible in the early stages of planning. This study proposes a unified framework that can simultaneously predict the beam flux distribution and 3D dose, so that dose inference can be performed without clear beam parameters to assist physicians in plan implementation.
Methods
The research established a 3D deep learning model that takes CT scans and anatomical structures as inputs to jointly predict beam distribution and dose. The network comprises three key components: a CT-based encoder, a beam generation branch, and a dose prediction branch built on 3D U-Net. Instead of relying on explicit beam angle conditions, we adopted a knowledge distillation strategy: teacher models trained with beam angle information learn and provide latent beam representations, while student models are trained to infer these representations directly from CT anatomical data. The predicted beam distribution was then fused with CT features to guide dose prediction.
Results
The model achieved a Mean Absolute Error (MAE) of 3.12Gy and Root Mean Square Error (RMSE) of 8.56 Gy for dose prediction, demonstrating moderate accuracy. DVH analysis for the PTV showed a D95 of 98.5, while beam prediction achieved a Dice@0.5 of 0.8402 and an IoU@0.5 of 0.7244, indicating strong agreement with the ground truth beam patterns. Overall, the model effectively predicts both dose and beam distributions, with clinically acceptable accuracy.
Conclusion
This study demonstrates that simultaneous beam and dose prediction with knowledge distillation is feasible and effective for radiotherapy dose modeling. By inferring beam information directly from patient anatomy, the proposed approach enhances physical interpretability and supports dose prediction in scenarios where beam parameters are unavailable, offering potential value for automated planning and clinical decision support.