Foundation Model-Based Transfer Learning for Data-Efficient and Deliverable Automated Treatment Planning
Abstract
Purpose
To address the data scarcity bottleneck in deep learning (DL) dose prediction by establishing a foundation model-based transfer learning framework. We aim to demonstrate that adapting large-scale public priors to institutional data significantly enhances learning efficiency and prediction accuracy, directly translating into clinically deliverable plans for complex head-and-neck (HN) and pancreas cases.
Methods
A foundation model was established by pretraining three diverse convolutional neural network and transformer-based DL backbones on 3,232 public HN and lung plans to learn generic anatomical and dosimetric features. This foundation was adapted via site-specific transfer learning to institutional cohorts of 256 HN and 72 pancreas patients. Model performance and training convergence rates were benchmarked against training-from-scratch baselines across varying data fractions (5%–100%). Finally, predictions were converted into deliverable volumetric modulated arc therapy (VMAT) plans via an automated 15-minute MIM and RayStation dose-mimicking workflow. Statistical significance was assessed using the Wilcoxon signed-rank test.
Results
The foundation model approach demonstrated superior generalization capabilities. In the HN cohort, fine-tuning on merely 12 cases (5%) achieved accuracy equivalent to training from scratch on 116 cases (50%), demonstrating a 10-fold improvement in data efficiency. Furthermore, the transfer learning strategy accelerated training convergence by approximately 10-fold compared to the baseline. For the data-scarce pancreas cohort, transfer learning reduced the mean absolute error (MAE) by 15% compared to the baseline. The downstream automated plans maintained high fidelity to DL predictions and were fully deliverable, passing patient-specific quality assurance (PSQA) within 1%, while offering significantly better organ-at-risk (OAR) sparing (p<0.05) compared to manual clinical plans.
Conclusion
Leveraging foundation models pre-trained on large-scale public data provides a robust solution to data scarcity in radiotherapy artificial intelligence. This paradigm enables institutions to deploy high-precision, clinically deliverable automated planning with minimal local data requirements and zero manual intervention.