An Ultra-Fast Dose Calculation Algorithm with Transformer Neural Networks: Model Fine-Tuning and Validation
Abstract
Purpose
Fast and accurate dose calculation is a key requirement for real-time adaptive radiotherapy, particularly for emerging image-guided radiation therapy systems. Monte Carlo methods and existing treatment planning system (TPS) algorithms do not achieve the computational speeds required for real-time application, necessitating development of an ultra-fast dose calculation algorithm. An existing transformer-based dose prediction algorithm (iDoTA) demonstrated sub-second calculation times, but limited generalizability across photon energies. Here, we investigate the performance of a fine-tuned iDoTA model using clinical treatment plans with 6, 10, and 15MV photon beams.
Methods
The iDoTA model combines a U-net convolutional neural network with a transformer backbone to predict a 3D dose distribution from a patient CT and ray-tracing volume. Starting from the open-source, published iDoTA weights, the model was trained using an additional 309 individual static fields from 59 breast field-in-field treatment plans with 6, 10, and 15MV photon energies. The ground-truth dose distributions were calculated using AcurosXB from the Varian Eclipse TPS. The retrained model’s performance was evaluated on 105 fields from 10 breast patients unseen during training, and the dosimetric agreement between iDoTA and AcurosXB was quantified using average relative error (ARE).
Results
The fine-tuned model demonstrated improved ARE compared to predictions generated with the original iDoTA weights. Across all test fields, the mean ARE decreased from 7.75% to 5.27%. Per beam energy, the fine-tuned iDoTA achieved mean AREs of 5.50%, 4.43%, and 5.35% for 6, 10, and 15MV beams, respectively, outperforming the original model at all energies. The relative differences between iDoTA and the TPS are greatest in the build-up and penumbra regions.
Conclusion
Fine-tuning an existing transformer-based photon dose prediction model using clinical treatment plans improved the ARE across multiple photon energies without modifying the underlying architecture. These results support the feasibility of extending fine-tuning to a variety of treatment sites.