3D SE-ResUNet for Accurate Dose Prediction in Head and Neck IMRT: a Structure-Specific Ensemble Approach

Accurate three-dimensional dose prediction is important for automated treatment planning and radiotherapy in head and neck cancer. This study investigates an ensemble-based deep learning framework for dose prediction from computed tomography (CT) images and anatomical structures, with particular emphasis on structure-specific model weighting. A dataset of 340 head and neck cancer patients from the Open Knowledge-Based Planning (OpenKBP) dataset was used for model development and evaluation. Three three-dimensional convolutional neural network (CNN) architectures, including U-Net, SE-UNet, and SE-ResUNet, were trained for dose prediction. Their predictions were first combined using a global weighted ensemble. To account for differences in prediction performance across anatomical structures, a structure-specific ensemble was subsequently developed. For each structure, the relative contribution of the three models was optimized independently on the validation set using an exhaustive grid search with weight increments of 0.1. Performance was evaluated using the Dose Score and dose-volume histograms (DVH) Score. Among the individual models, SE-ResUNet achieved the lowest DVH Score (2.067 Gy), while the global weighted ensemble achieved a Dose Score of 2.77 Gy and a DVH Score of 2.068 Gy. The proposed structure-specific ensemble substantially reduced the DVH Score to 1.75 Gy, although its overall Dose Score increased slightly to 2.81 Gy. The proposed structure-specific ensemble provides an anatomy-aware strategy for combining complementary dose predictions from heterogeneous deep learning models. While it does not improve the overall Dose Score relative to the global weighted ensemble, it substantially improves the DVH Score and provides a more structure-aware characterization of dose prediction performance. The approach may therefore be useful for deep learning-based dose prediction in head and neck IMRT. .