S2167
Physics - Inter-fraction motion management and daily adaptive radiotherapy
ESTRO 2026
Digital Poster 3852 Diffusion-based synthetic-CT generation from head and neck CBCT Lorenzo Colombo 1 , Audrey Duran 2 , Quentin Spinat 2 , Pierre Olléon 2 , Sami Romdhani 2 , Olivier Teboul 2 , Nikolaos Paragyios 3 , Pascal Fenoglietto 4 , Pauline Maury 5 , Charlotte Robert 6 1 Clinical affairs, TheraPanacea, Paris, France. 2 AI engineering, TheraPanacea, Paris, France. 3 CEO, TheraPanacea, Paris, France. 4 Department of radiation oncology, Institut du cancer de Montpellier, Montpellier, France. 5 Department of radiation oncology, Institut Gustave Roussy, Villejuif, France. 6 Department of radiation oncology, Institut Gustave Roussy, Villejuid, France Purpose/Objective: Cone-beam CT (CBCT) imaging is routinely used in radiation therapy for patient positioning; however, its broader clinical application is constrained by suboptimal image quality, spatial resolution, and field of view. Expanding CBCT beyond setup verification could significantly advance adaptive radiotherapy, provided image fidelity is enhanced to enable accurate organ-at-risk delineation, dose calculation, and treatment replanning. This study presents and clinically evaluates an AI–based method that generates synthetic-CTs (sCTs) from CBCT images. The proposed solution mitigates key technical limitations of CBCT and demonstrates its potential to fully exploit CBCT imaging for adaptive radiotherapy in head and neck cancer management. Material/Methods: Training a 3D multimodal translation model for medical imaging presents several challenges, including the computational burden of volumetric data, the scarcity of high-quality paired datasets, and the need to preserve quantitative accuracy for clinical use. To address these, a 3D latent diffusion model was developed. Its variational autoencoder (VAE) component provides strong compression while maintaining high reconstruction fidelity and computational efficiency. The diffusion model was first trained in an unsupervised manner on unpaired data, then fine-tuned for cross-modality translation using limited paired samples. A ControlNet-based conditioning strategy was implemented to achieve anatomically consistent and quantitatively reliable image synthesis. The unsupervised model was trained on approximately 34000 CT scans (50% pelvis, 37.5% head and neck, 12.5% thorax) to ensure anatomical diversity, followed by fine-tuning with 306 head and neck cancer patients (4414 paired datasets). An independent cohort of 20 head and neck cancer patients treated across three European cancer centres was used for clinical evaluation. Planning-CTs were
Conclusion: Plan complexity in MR-guided oART exhibits distinct inter- and intra-patient variability signatures. While inter-patient differences are mainly determined by initial setup and planning preparation, the largest fluctuations arise from extreme anatomical variations, where single-fraction changes can surpass the variability observed across patients. Keywords: Plan Complexity, MRIgRT, pancreatic cancer
Made with FlippingBook - Share PDF online