Author: Theodore Higgins Arsenault, Beatriz Guevara, Rojano Kashani, Raymond F. Muzic, Gisele Castro Pereira, Alex T. Price 👨🔬
Affiliation: University Hospitals Seidman Cancer Center, Case Western Reserve University Department of Biomedical Engineering 🌍
Purpose: Accurate dose prediction in radiotherapy is essential for treatment planning. This study evaluates four nnUnet-based models using the OpenKBP head and neck dataset: a baseline model (Model 1) trained on all anatomical structures, a reduced model for planning target volumes (PTVs) (Model 2), and two medically informed models focusing on PTVs coverage (Model 3) and PTVs coverage with gradients (Model 4). The aim is to assess the effect of incorporating domain knowledge on predictive accuracy and efficiency.
Methods: The dataset was split into training (n=200), validation (n=40), and testing (n=100) sets. Performance was evaluated using Mean Absolute Error (MAE) and Structural Similarity Index Measure (SSIM) between the ground truth and predicted dose distribution, as well as prediction time. Medically informed models prioritized critical regions such as PTVs and gradients around them, while the baseline and PTV-only models used mean squared error (MSE) across the entire matrix volume in the loss function. All models were trained using an Nvidia A100 GPU.
Results: Model 1 (CT + All Structures, MSE) achieved an MAE of 5.348±0.881 Gy and an SSIM of 0.933±0.070 with a prediction time of 0.010±0.003s. Model 2 (CT + PTV only, MSE) had a lower MAE of 5.109±0.597 Gy and similar SSIM while also predicting in 0.010±0.004s. Model 3 (CT + PTV only, PTV coverage loss) excelled with the lowest MAE of 4.482±0.625 Gy, highest SSIM of 0.940±0.068, and the fastest time of 0.010±0.002s. Model 4 (CT + PTV only, PTV coverage + gradient loss) had an MAE of 5.210±1.112 Gy but maintained competitive SSIM (0.935±0.068) and fast prediction times. Overall, Model 3 provided the best accuracy-efficiency balance.
Conclusion: Medically informed models, particularly the PTV-coverage approach, improved predictive accuracy and spatial fidelity. These findings highlight the value of integrating clinical expertise into neural network architectures for radiotherapy dose prediction.