S1871
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
quality. Knowledge-Based Planning (KBP)[1] offers data-driven automation for patient-specific plans. While traditional KBP methods use predefined features to predict dose-volume histograms, they struggle to model complex spatial dose patterns. Deep learning-based KBP improves prediction capabilities but faces dimensionality trade-offs: 2D approaches lose 3D spatial consistency, which compromises accuracy, while 3D approaches improve volumetric consistency but incur prohibitive computational and memory costs due to processing redundant voxels. To overcome these limitations, we developed OctreeFormer, a novel sparse-voxel-based 3D prediction network built upon octree- representation transformers. Material/Methods: As illustrated in Fig.1, the proposed framework leverages octree-based transformers for dose prediction. It encodes the input 3D volume into a hierarchical octree, selectively processing only relevant regions to reduce computational redundancy and GPU memory demands. The architecture incorporates two main innovations: (1) A hierarchical framework that supervises 3D feature learning using multi-stage contextual cues. It employs specialized Octree- Transformer Blocks (Fig.2) where an attention mechanism enables efficient token interaction based on the node hierarchy, allowing dynamic focus on multi-scale features. (2) A PTV-guided masking strategy that stochastically masks input tokens based on a probability distribution derived from the PTV's geometry and vicinity. This prioritizes interactions between the PTV and nearby OARs, helping the model capture clinically critical spatial relationships. The method was evaluated on an in-house dataset and the public OpenKBP dataset for radiotherapy dose prediction.
Conclusion: Algorithm-dependent dose variations can be
substantial for small lung targets. Among the tested algorithms RS MC and AXB are recommended because they showed the closest agreement with the in-house MC reference. Furthermore, it is preferable to report dose as Dm to avoid Dm-to-Dw conversion uncertainties in heterogeneous lung tissue. Keywords: Dose algorithms, SBRT, Lung cancer
Poster Discussion 2044
OctreeFormer: Efficient Radiotherapy Dose Prediction Using Hierarchical Octree-based Transformers Yuan Dong, Hongyi Chen, Jiahao Wang, Long Sun, Xue Bai, Binbing Wang, Jiping Liu Department of Radiation Physics, Zhejiang Cancer Hospital, HangZhou, China Purpose/Objective: Modern radiation therapy techniques (IMRT/VMAT) requires precise treatment planning to deliver adequate radiation doses to Planning Target Volumes (PTVs) while sparing Organs at Risk (OARs). Precise treatment planning relies on time-consuming, iterative manual optimization by physicists, and thus introduces significant interobserver variability in plan
Made with FlippingBook - Share PDF online