1060 - A Dynamic Priority-Aware Diffusion Model for Clinically-Driven Dose Prediction in Radiotherapy

03:05pm - 03:10pm ET

Room 156

Presenter(s)

Xianrui Yan, MS, M.Eng - Southeast University, Nanjing, shandong

X. Yan^1,2, K. Gao³, C. Li³, Y. Wang³, C. Qinghao⁴, H. Shu¹, and J. Zhu³; ¹Southeast University, Laboratory of Image Science and Technology, The Key Laboratory of Computer Network and Information Integration, Ministry of Education, Jiangsu Provincial Joint International Research Laboratory of Medical Information Processing, Cent, Nanjing, China, ²Department of Radiation Oncology Physics & Technology, Shandong Cancer Hospital of Shandong First Medical University, jinan, China, ³Department of Radiation Oncology Physics & Technology, Shandong Cancer Hospital of Shandong First Medical University, Jinan, China, ⁴Linyi Hospital of Traditional Chinese Medicine, Department of Radiation Oncology Physics & Technology, Linyi, China, Linyi, China

Purpose/Objective(s):

Deep learning-based dose prediction models frequently struggle with the complex trade-offs between target coverage and organ-at-risk (OAR) sparing, often yielding deterministic outputs with limited clinical adaptability. To address this, we proposed a novel generative approach utilizing a Dynamic Priority-Aware Diffusion Model. By explicitly integrating a clinical priority map and optimizing directly within the physical dose space, this study aims to generate highly accurate, clinically executable dose distributions that strictly adhere to physician intent.

Materials/Methods:

A customized multi-channel U-Net architecture was developed as the backbone of a denoising diffusion probabilistic model. The model integrated CTs, structure masks, and a dynamically generated priority map designed to guide the generative trajectory toward clinically optimal trade-offs. Our network computed a hybrid loss function directly within the physical dose space, which comprised Spatial Dose-Volume loss, OAR-specific L1 loss, and Dose Gradient loss to enforce physical realism. To ensure training stability and prevent model collapse, a forced linear warm-up strategy was implemented. The framework was rigorously validated using a dual-institution approach, comprising the public AAPM OpenKBP challenge dataset (head and neck) and a proprietary institutional dataset (thoracic tumors). Dosimetric fidelity was evaluated using standard dose-volume histogram (DVH) metrics and 3D gamma analysis.

Results:

This model demonstrated robust generalization ability across distinct anatomical sites during independent evaluations. On the OpenKBP dataset, the model achieved an overall mean absolute error (MAE) of 0.22±0.15 Gy and a root-mean-square error (RMSE) of 1.37±0.52 Gy. Guided by the priority map, structure-wise mean dose errors for critical OARs were tightly controlled, yielding -0.41±2.61 Gy for the brainstem and 0.24±2.07 Gy for the larynx. On the independent thoracic cohort, the model maintained high predictive accuracy, with an overall MAE of 0.51±0.35 Gy and an RMSE of 2.08±1.19 Gy. Crucially, physical-space optimization effectively constrained doses to adjacent structures, resulting in exceptional mean dose errors of -0.57 ± 2.24 Gy for the lungs and 0.11±2.91 Gy for the heart. Across both cohorts, the dynamic priority mechanism consistently mitigated local dose deviations in critical structures without compromising global dose fidelity.

Conclusion:

The Dynamic Priority-Aware Diffusion Model represents a significant advancement toward intent-driven artificial intelligence in radiotherapy. By explicitly translating clinical priorities into generative guidance and strictly enforcing physical dose constraints, the model produces highly accurate predictions that closely mirror expert planning. This robust framework demonstrates strong potential for seamless integration into automated treatment planning and online adaptive radiotherapy workflows.