Generalizability and Domain Adaptation of a Deep Learning Dose Prediction Model for Head-and-Neck Radiotherapy
Abstract
Purpose
Deep learning–based dose prediction models have demonstrated high accuracy in multi-institutional challenges; however, their performance under real-world institutional domain shift remains unclear. This study evaluates the generalizability of a state-of-the-art dose prediction model trained for the GDP-HMM challenge on independent institutional head-and-neck (HN) radiotherapy data and quantifies the impact of targeted institutional fine-tuning.
Methods
A MedNeXt-based 3D convolutional neural network previously developed and trained for the GDP-HMM dose prediction challenge was evaluated on retrospective HN cases collected at a single academic institution. The institutional dataset comprised 390 HN treatment plans, of which 350 cases were used for fine-tuning, 40 for validation and model selection, and 37 independent cases for testing. Model performance was first assessed without retraining by applying the GDP-HMM–trained model directly to the institutional test set. The model was then fine-tuned using the institutional training cohort and re-evaluated on the same held-out test set. Prediction accuracy was quantified using voxel-wise mean absolute error (MAE) within the patient body.
Results
When evaluated on the original GDP-HMM validation dataset, the pretrained model achieved an MAE of 2.038 ± 0.640 Gy. Direct application of this model to the institutional test cohort resulted in substantially degraded performance, with an MAE of 7.387 ± 2.715 Gy, indicating pronounced domain shift. After fine-tuning on the institutional training data, test-set performance improved markedly, achieving an MAE of 3.676 ± 1.075 Gy, corresponding to a ~50% reduction in error relative to the non-adapted model.
Conclusion
Although high-performing dose prediction models trained on large multi-institutional datasets may not generalize directly to new clinical environments, modest institutional fine-tuning substantially restores accuracy. These findings highlight both the limitations of out-of-distribution deployment and the practical feasibility of site-specific adaptation, supporting the clinical translation of AI-based dose prediction in radiotherapy.