DVH-Consistent Patient-Specific Dose Prediction for Head-and-Neck Radiotherapy Using a Robust Multi-Input Cascaded U-Net
Abstract
Purpose
To develop a clinically oriented framework for patient-specific 3D dose prediction in head-and-neck radiotherapy, emphasizing DVH-consistent performance for target coverage and OAR sparing.
Methods
A high-quality single-institution cohort of 689 head-and-neck radiotherapy patients was retrospectively collected, consisting of planning CT, clinically used target/OAR structures, beam-related information, and corresponding clinical 3D dose distributions. A multi-input cascaded U-Net was trained to predict 3D dose from CT, PTV, OAR, and beam information. The model was optimized using a combined L1 loss and a Sobel-based gradient loss to improve voxel-wise accuracy while preserving steep dose gradients. Performance was evaluated using voxel-wise mean absolute error (voxel MAE) and a DVH quantile MAE, defined as the mean absolute error of percentile-dose values sampled at fixed volume levels (V = 20%, 40%, 60%, 80%, and 99%) for targets and major OARs.
Results
The proposed model achieved a voxel MAE of 2.16 and a DVH quantile MAE of 3.14. Visual inspection of predicted dose maps demonstrated close spatial agreement with reference dose distributions and preserved high-gradient regions near target–OAR interfaces.
Conclusion
A robust multi-input cascaded U-Net trained on a high-quality clinical cohort with complete planning context (CT/structures/beam/dose) can predict patient-specific 3D dose distributions for head-and-neck radiotherapy with quantitative agreement in both voxel-wise dose error and DVH-based evaluation.