S1972
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
by reducing planning time. Most existing ATP approaches rely on direct aperture optimization (DAO) with dose-mimicking objectives [2], but their application is often limited to specific prescriptions and tumor sites. To overcome these limitations, we introduce a pipeline integrating a generalizable dose prediction model coupled with an efficient optimization engine, enabling the generation of clinically deliverable, quality assurance–passing treatment plans. Material/Methods: Our ATP pipeline (Figure1) consists of deep-learning based dose prediction followed by volumetric arc modulated therapy (VMAT) DAO using a custom collapsed-cone convolution dose engine [3]. The optimization stage employs a 3D inverse planning framework leveraging voxel- wise dose constraints and aperture smoothing to ensure plan deliverability and dosimetric robustness. The training dataset included 403 patients treated using VMAT (2017-2024), with prescriptions of ≥ 50 Gy for one to three PTV levels, encompassing all tumor locations. The test cohort consisted of 16 patients with oropharyngeal, hypopharyngeal, and laryngeal tumors, with two-level prescriptions – tumor sites that represent ~80% of H&N cases. Our dose prediction model is a latent diffusion model, adapted from [4]. Pipeline performance was evaluated based on plan metrics and quantitative DVH metrics. Statistical significance was assessed using the Wilcoxon signed rank test. Automated and clinical plans (CP) were reviewed in a blind comparison study by two radiation oncologists (one junior, one senior) who assessed plan acceptability and stated their preference. Results were aggregated across both
technique.
Conclusion: PRV-SHS, a purely coplanar VMAT-based technique was able to achieve dosimetric results comparable to TomoTherapy 1 cm and superior to 2.5 cm jaws widths while maintaining full hippocampal and OAR constraints. It offers a practical, accessible alternative with reduced treatment time for HS-WBRT. This is particularly useful in centers without TomoTherapy or non-coplanar planning capabilities. Further prospective validation with formal non-inferiority testing is warranted. References: Gondi V, Pugh SL, Tome WA, et al. Preservation of memory with conformal avoidance of the hippocampal neural stem-cell compartment during whole-brain radiotherapy for brain metastases (RTOG 0933): a phase II multi-institutional trial. J Clin Oncol. 2014;32(34):3810-3816. Keywords: hippocampal sparing, VMAT, tomotherapy End-to-End Head and Neck VMAT Planning Using a Deep Learning–Based Latent Diffusion and 3D Robust Optimization Framework Jeanne Boyer-Chammard 1,2 , Rémi Vauclin 3 , Elie Mengin 3 , Madalina Costea 4 , Adelina Brezae 4,5 , Roger Sun 6 , Pauline Maury 6 , Charlotte Robert 6 , Marie-Claude Biston 5 , Nikos Komodakis 1,7 , Nikos Paragios 8,9 , Vincent Grégoire 5 , Eric Deutsch 6 1 R&D Artificial Intelligence, TheraPanacea, Paris, France. 2 Joint Collaboration : Gustave Roussy INSERM 1030 / Department of Radiation Oncology, Université Paris-Saclay / Centre Léon Bérard, Villejuif / Lyon, France. 3 Physics, TheraPanacea, Paris, France. 4 Clinical Affairs, TheraPanacea, Paris, France. 5 Department of Radiation Oncology, Centre Léon Bérard, Lyon, France. 6 Gustave Roussy, INSERM 1030, Université Paris- Saclay, Villejuif, France. 7 Computer Science Department, University of Crete, Rethymno, Greece. 8 CEO, TheraPanacea, Paris, France. 9 Centrale-Supélec, Université Paris-Saclay, Gif-sur-Yvette, France Purpose/Objective: Automated treatment planning (ATP) for Head-and- Neck (H&N) cancer reduces inter-observer variability, improves treatment quality [1] and clinical efficiency Digital Poster Highlight 3996
experts. Results:
ATP achieved superior organs-at-risk (OAR) sparing (Table1), with significantly reduced doses to the brainstem, spinal cord, and parotid glands (p<0.002), as well as to the mandible, esophagus, and trachea (p<0.004). Target coverage remained equivalent between plans (no significant difference in D95%, p>0.027), while the homogeneity index in the primary target volume was significantly improved for ATP (p<0.002). CP showed higher overall maximum dose (p<0.003), whereas the total number of monitor units was comparable (p=0.099). ATP demonstrated superior clinical acceptability (94%: 78% as is, 16% minor edits) compared to CP (91%: 81% as is, 10% minor edits), with clinicians preferring ATP in 62.5% of cases. Conclusion: Our framework demonstrated significant OAR dose reduction and improved dose homogeneity, while preserving target coverage and increasing planning
Made with FlippingBook - Share PDF online