Answers are grounded in this paper and cite their sources.
Objective. Diffusion models show strong potential for high-fidelity radiotherapy dose prediction, but their performance depends heavily on the conditioning mechanism. This study introduces a novel conditioning strategy that improves volumetric consistency in slice-wise dose generation while reducing reliance on manually annotated organ-at-risk (OAR) masks. Approach. We propose a diffusion-based dose prediction framework conditioned on features from the full 3D computed tomography (CT) volume and a pre-trained medical image segmentation model. A 3D feature extractor encodes global anatomical information from the entire CT volume to preserve spatial coherence across 2D dose predictions. Rather than relying on manual OAR masks, the model uses anatomical representations from the pre-trained segmentation network, integrated via cross-attention to enable anatomy-aware dose prediction. Main results. The framework, Diff-Seg3D, was evaluated on three lung cancer radiotherapy datasets representing Volumetric Modulated Arc Therapy (VMAT), Fixed Gantry Intensity-Modulated Radiation Therapy (FGIMRT), and 3D Conformal Radiation Therapy (3DCRT) planning techniques. It achieved the numerically lowest mean absolute error (MAE) on all three datasets, with values of 3.16, 4.38, and 4.02 Gy, respectively, and the best dose-volume histogram (DVH) scores (2.64 and 3.82) on VMAT and FGIMRT, while remaining competitive on 3DCRT. Pre-trained 3D anatomical and segmentation encoders improved the baseline diffusion model, with segmentation features improving performance without requiring explicit OAR contours at inference. Significance. This conditioning strategy improves volumetric consistency and anatomical awareness in diffusion-based dose prediction. The framework can serve as a tool for generating treatment-specific, anatomy-conditioned dose priors to guide radiotherapy planning.
No takes yet. Share an insight, caveat, or question.
Cobilean et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: