BLUE RIBBON POSTER MULTI-DISCIPLINARY: Reconstructing Delivered Dose In Real Time: A Beam Physics-Embedded, Language-Model-Driven Approach
Abstract
Purpose
Existing deep learning-based dose prediction methods primarily learn empirical mappings between anatomy and dose, without modeling beam delivery physics. This gap may limit their robustness and accuracy, especially in heterogeneous regions where dose deposition is highly sensitive to beam configuration and tissue composition. We propose a dose reconstruction framework that integrates beam physics via a large language model (LLM) into the calculation model to improve its accuracy and robustness.
Methods
The proposed approach consists of a two-stage pipeline. In Stage 1, beam delivery physics (e.g., fluence modulation and energy characteristics) is retrieved from AAPM guidelines and institutional protocols and tokenized using an LLM into deterministic representations that serve as beam-modality priors. In Stage 2, beam-wise dose reconstruction is performed in beam-eye-view (BEV) space using a multimodal network that fuses BEV-CT images and LLM-derived physical tokens. Model training employs a composite loss combining mean absolute error, mean squared error, edge-aware loss, and high-dose region-weighted loss.
Results
A total of 50 lung dosimetry patients were included, with 38/4/8 used for training, testing, and validation. Qualitative comparisons demonstrate close agreement between predicted and reference dose distributions, with residual errors largely confined to low-dose regions and magnitude of within ±3 Gy out of ~75Gy (max dose). Quantitatively, the proposed method achieves a mean 3D gamma pass rate (2mm/10%) of 0.975 ± 0.056, outperforming the comparison methods of 3D U-Net (0.931 ± 0.113), Dose-Net (0.939 ± 0.120), and CLIP-Unet (0.945 ± 0.087) (p < 0.05). More restricted, its 2D gamma pass rate (2mm/10%) reaches 0.962 ± 0.073. Performance remains stable across heterogeneous datasets.
Conclusion
By explicitly integrating LLM-encoded physical information into beam dose reconstruction, the proposed framework improves dosimetry accuracy significantly over existing learning-based methods. This physics-driven AI approach supports future development of intra-fractional dose reconstruction and has potential in real-time 3D in-vivo dosimetry.