Prescription-aware diffusion Transformer for esophageal VMAT dose prediction: a retrospective dual-center validation and target-level analysis

Abstract Background Deep learning-based radiotherapy dose prediction has the potential to support automated treatment planning and plan-quality assessment; however, simultaneous integrated boost (SIB) volumetric modulated arc therapy (VMAT) for esophageal cancer remains challenging because multiple target levels are associated with different prescription doses and complex spatial relationships. We developed a prescription-aware diffusion Transformer framework for esophageal VMAT dose prediction and evaluated voxel-wise prediction accuracy, prescription-specific target behavior, and target-integrity sensitivity in a dual-center cohort. Results The global MAE within the body/union PTV MAE of DoseDiT was 3.15/4.79 Gy in the validation set, 3.07/4.99 Gy in the internal test set, and 4.25/7.47 Gy in the external test set. Target-level analysis revealed a systematic prescription-dependent bias, characterized primarily by persistent underestimation of D95/Rx for Boost targets in both the internal and external cohorts. After RxMap ablation, the PTV MAE increased to 5.35 Gy internally and 8.17 Gy externally, supporting the value of voxel-wise prescription representation for modeling complex target-dose distributions. GlobalPSDM provided no additional accuracy gain, whereas the MMFNet baseline under matched settings achieved lower errors. These findings indicate that the principal value of this study lies in prescription-aware representation and target-level error characterization. Conclusion The voxel-wise RxMap provides local prescription-level information for SIB treatment plans, with its contribution concentrated primarily within target regions. Target-level analysis further demonstrated a systematic prescription-dependent coverage bias, with underestimation of Boost coverage showing good reproducibility in both the internal and external cohorts. GlobalPSDM did not provide an additional prediction-accuracy gain in the current implementation, whereas the matched MMFNet baseline achieved lower values for the major error metrics. These findings indicate that heterogeneous-prescription dose prediction should consider local prescription representation, target-level evaluation, and cross-center model robustness simultaneously. The current DoseDiT can serve as a research-oriented dose prior and analytical framework; further clinical translation will require three-dimensional or physics-aware modeling and prospective multicenter validation.