S1872
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
Digital Poster 2075
Halcyon breast radiotherapy: how the deployment of AI-generated structures reshapes IMRT–VMAT strategies. Caroline MOREAU-NOBLET, Marie DUTHY, Anne- Sophie FAUVEL, Christine CAILLET, Pascal GUINAUDEAU, Frédéric COSTE, Matthieu MOREAU- NOBLET Medical Physics Department, Clinique Mutualiste de l'Estuaire, Saint-Nazaire, France Purpose/Objective: In France, Halcyon (Varian) accelerators are widely deployed for intensity-modulated breast radiotherapy, representing 37% of breast-dedicated systems versus 16% in Europe [1]. These machines exclusively deliver VMAT or fixed-beam IMRT. Meanwhile, the growing use of AI-based auto-contouring has introduced new substructures, particularly cardiac ones, into optimization constraints, influencing planning strategies. This study aimed to evaluate IMRT and VMAT strategies on Halcyon, in a context of AI- generated structures, according to breast laterality and nodal involvement. Material/Methods: Sixty previously treated patients were retrospectively selected: 20 simple breast (SB), 20 with axillary- supraclavicular involvement (B + Ax/Sc), and 20 with Ax/Sc plus internal mammary chain involvement (B + Ax/Sc + IMC), equally divided between left and right sides. Left cases were planned in deep inspiration breath-hold, right in free breathing. The prescription was 40 Gy in 15 fractions, using De Rose et al. [2] constraints. All OARs were segmented with Limbus AI (RadFormation). Five techniques were compared for SB [3]: IMRT (4-6 beams), VMAT with 2, 3 or 4 partial arcs (2a, 3a, 4a), and butterfly VMAT (4a_btf). For nodal cases, IMRT, VMAT 3a and 4a were evaluated. All plans were optimized (Varian Eclipse PO v16.1) with an 8 mm PTVbreast expansion and recalculated on the original anatomical structures (AcurosXB v16.1, dose-to- medium). Dosimetric indices for PTVs and OARs (heart, LAD, left ventricle, lungs, contralateral breast, liver) were compared using Wilcoxon tests. Results: Target coverage was similar accross all techniques. For left SB, IMRT better spared contralateral OARs and heart (Dmean − 0.4 ± 0.5 Gy) but increased LAD Dmax by 8.9 ± 7.3 Gy versus VMAT 2a/3a/4a. For left B + Ax/Sc, IMRT reduced lung V4Gy by 5.8 ± 4.6 % with a similar heart Dmean as VMAT 3a/4a, while VMAT lowered LAD Dmax by 8.5 ± 6.5 Gy. For left B + Ax/Sc + IMC, only IMRT met pulmonary constraints, whereas VMAT 3a/4a reduced LAD Dmax by 11.5 ± 3.7 Gy. For right-sided cases, IMRT consistently improved OAR sparing compared with VMAT.
Results: Experimental validation confirmed the superior performance of OctreeFormer. Computationally, it achieved a ~30% reduction in GPU memory consumption and a 2.5x speedup in inference compared to state-of-the-art models. Moreover, OctreeFormer significantly outperformed leading 2D and 3D prediction methods in accuracy on the OpenKBP[2] and in-house (nasopharyngeal cancer) datasets. It demonstrated significantly superior spatial dose agreement, quantified by Dose Score[4]: 2.425 ± 1.054 (vs. DoseDiff[3]: 2.654 ± 1.040) on OpenKBP and 2.835 ± 1.153 (vs. DoseDiff: 3.134 ± 1.182) on the nasopharyngeal dataset. Conclusion: OctreeFormer successfully overcomes the key limitations of existing methods by integrating an efficient octree representation with a hierarchical transformer and PTV-guided masking. The model reduces computational costs while improving prediction accuracy, demonstrating strong potential for clinical deployment. This approach can shorten planning time, standardize plan quality, and reduce reliance on manual tuning. References: [1] Chung C V, Khan M S, Olanrewaju A, et al. Knowledge-based planning for fully automated radiation therapy treatment planning of 10 different cancer sites[J]. Radiotherapy and Oncology, 2025, 202: 110609.[2] Babier A, Zhang B, Mahmood R, et al. OpenKBP: the open - access knowledge - based planning grand challenge and dataset[J]. Medical Physics, 2021, 48(9): 5549-5561.[3] Zhang Y, Li C, Zhong L, et al. DoseDiff: distance-aware diffusion model for dose prediction in radiotherapy[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3621-3633.[4] Kui X, Liu F, Yang M, et al. A review of dose prediction methods for tumor radiation therapy[J]. Meta-Radiology, 2024, 2(1): 100057. Keywords: Dose Prediction, Automated Planning, Deep Leaning
Made with FlippingBook - Share PDF online