A Controlled Evaluation of Architectural Enhancements to 3D U-Net for Automated Radiotherapy Dose Prediction
Abstract
Purpose
To systematically assess whether commonly proposed architectural enhancements provide measurable benefits for deep learning-based radiotherapy dose prediction, using controlled comparisons of 3D U-Net variants to support evidence-based model selection and establish a practical baseline for clinic-specific development.
Methods
Four 3D U-Net architectures were evaluated using the OpenKBP challenge dataset comprising 340 head-and-neck radiotherapy patients (training n = 200, validation n = 40, test n = 100). The models included a baseline U-Net with standard convolution blocks, an Attention U-Net, a ResNet U-Net, and a Transformer-based U-Net. All models were trained under identical training parameters - including optimizer, loss function, learning rate schedule, batch size, input resolution, and random seed - to isolate the effects of architectural modifications. Final performance was assessed on the test cohort using established OpenKBP metrics: dose score, dose-volume histogram (DVH) score, and mean squared error (MSE).
Results
The baseline U-Net significantly outperformed all enhanced architectures across all evaluation metrics (paired T-test, all p < 0.001). Mean dose scores were 2.77 ± 1.04 Gy for the baseline, compared with 3.20 ± 1.16 Gy for Attention U-Net, 3.34 ± 1.14 Gy for ResNet U-Net, and 3.31 ± 1.21 Gy for Transformer U-Net. Corresponding DVH scores were 2.20 ± 1.32 Gy, 2.98 ± 1.96 Gy, 2.72 ± 1.52 Gy, and 2.40 ± 1.29 Gy, respectively.
Conclusion
Under controlled training conditions, increased architectural complexity did not yield performance gains over a baseline 3D U-Net. This study provides a reproducible reference for architecture benchmarking and suggests that, for datasets of this size and scope, simpler architectures may represent a stable and interpretable baseline for clinic-specific dose prediction research.