S2288
Physics - Machine learning and AI algorithms
ESTRO 2026
Digital Poster Highlight 2046
VQ-DoseNet: A Vector Quantized Model for Stochastic Radiotherapy Dose Prediction Dong Yang 1 , Yao Xu 1 , Zihan Sun 2 , Zhen Zhang 1 , Weigang Hu 1 , Jiazhou Wang 1 1 Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China. 2 Radiation Oncology, The first affiliated hospital of Zhejiang University school of medicine, Hangzhou, China Purpose/Objective: Accurate dose prediction can greatly accelerate radiotherapy planning and improve plan consistency. However, current deep learning–based predictors typically produce deterministic outputs that overlook the inherent variability among clinically acceptable plans. This lack of variability limits their clinical usability, as real-world planning often involves trade- offs between target coverage and organ-at-risk (OAR) sparing. To address this issue, we developed VQ- DoseNet, a stochastic dose prediction framework designed to generate multiple clinically plausible dose distributions that capture realistic planning variability while maintaining dosimetric accuracy. Material/Methods: The proposed VQ-DoseNet combines vector-quantized anatomical encoding with a dose-variability encoder aligned through KL divergence. Controlled perturbations in the latent space are introduced during decoding, enabling stochastic yet anatomically consistent dose predictions.Model performance was evaluated on three anatomical sites—nasopharynx, rectum, and breast—comprising 76, 186, and 231 patients, respectively. Each nasopharyngeal case included five clinically delivered re-plans, allowing assessment of model-derived variability against real clinical variations. The rectum and breast datasets were used for cross-site validation to evaluate generalizability. The public OpenKBP dataset was used for external benchmarking.Prediction accuracy was assessed using mean absolute error (MAE), dose score, and DVH score. Clinical acceptability was evaluated based on PTV coverage and OAR sparing consistency across multiple stochastic predictions.
Results: Our TALLM successfully detected and routed all 50 test cases to the appropriate backend tools. For BED and EQD2 computations, both online models (GPT-4-Plus and Gemini-Pro) and our TALLM produced identical outputs (MAE = 0). However, for NTCP calculations, distinct performance patterns were noticed. Our TALLM maintained perfect numerical accuracy (MAE = 0) by consistently delegating dose-response computations to verified backend functions. In contrast, both online models exhibited hallucinations and numerical errors: GPT-4 showed a MAE of 0.033, with 70% of predictions within ±2 percentage points of reference values, while Gemini-Pro showed a MAE of 0.066, with only 46% of predictions within the same tolerance. Conclusion: Our results demonstrate that while general LLMs can reliably perform simple dose conversions, they lack the precision required for complex parameter- dependent radiobiological calculations. This can lead to both over- and underestimation of side effect risks. Our tool augmentation methodology addresses this limitation by providing deterministic and reproducible results while maintaining data privacy through fully offline operation. References: 1 Jackson A, Marks LB, Bentzen SM, et al. The lessons of QUANTEC: recommendations for reporting and gathering data on dose-volume dependencies of treatment outcome. Int J Radiat Oncol Biol Phys. 2010;76(3 Suppl):S155-S160.2 Niemierko A. A generalized concept of equivalent uniform dose (EUD). Med Phys. 1999;26(6):1100-1109. Keywords: Tool augmentation, LLM, Radiobiology
Made with FlippingBook - Share PDF online