A physics-informed deep learning model for predicting beam dose distribution of intensity-modulated radiation therapy treatment plans.

Background and purpose: We aimed to develop a physics-informed deep learning model for beam dose prediction in intensity-modulated radiation therapy (IMRT) for patients with nasopharyngeal cancer. Materials and methods: A total of 100 nine-beam IMRT cases are enrolled in this study retrospectively, divided into training set (72), validation set (8), and test set (20). CT images and contour inputs are preprocessed to generate multiple feature maps for each beam angle, incorporating the dose fall-off principles in water for 6MV photons. Four beam dose prediction models using different loss are built using the U-Net framework to predict each beam dose simultaneously. Beam dose mean absolute error (MAE), beam dose gradient Euclidean distance, total dose MAE, and total dose gradient Euclidean distance are calculated to evaluate model performance. Results: The dose prediction model with beam dose loss, gradient loss, and masked loss achieves total dose MAE of 2.92 Gy, total dose gradient Euclidean distance of 1.35, beam dose MAE of 0.96 Gy, and beam dose gradient Euclidean distance of 0.30. Conclusions: This study proposes a physics-informed deep learning network specifically for the task of beam dose prediction. Additionally, this study addresses the interpretability challenges in deep learning models by employing a crosshair sampling scheme to validate the relationships between input and output channels.