S2315
Physics - Machine learning and AI algorithms
ESTRO 2026
differences were observed between CT and sCT for mean doses in any target or organ at risk (all p > 0.05).Mean dose differences were < 0.2 Gy for all structures, including the PTV and OARs (duodenum, stomach, colon, spinal cord, small bowel).Gamma pass rates demonstrated excellent concordance: 99.8 ± 0.3 % (3%/3 mm), 99.4 ± 0.6 % (2%/2 mm) and 93.9 ± 2.4 % (1%/1 mm).These results confirm the high geometric and dosimetric fidelity of the synthetic CTs generated by MVision for abdominal MR-only planning. Conclusion: AI-based synthetic CT generation from TrueFISP MR images on the MRIdian system provided accurate dose calculations for abdominal SBRT without dosimetric compromise.This study represents, to our knowledge, one of the first clinical validations of a commercial sCT solution for MR-only liver and pancreatic SBRT.The high gamma pass rates and non-significant dose differences support clinical implementation of this workflow for adaptive MR-guided radiotherapy. Keywords: MR-only radiotherapy, synthetic CT, abdominal SBRT Quad-modal Digital Patient Phantom Generation via Latent Rectified Flow: Toward a Large Cohort Creator for Virtual Trials Muheng Li 1,2 , Jintao Ren 3,4 , Antony John Lomax 1,2 , Stine Sofia Korreman 3,4 , Ye Zhang 1 1 Center for Proton Therapy, Paul Scherrer Institute, Villigen, Switzerland. 2 Department of Physics, ETH Zurich, Zurich, Switzerland. 3 Danish Centre for Particle Therapy, Aarhus University Hospital, Aarhus, Denmark. 4 Department of Clinical Medicine, Aarhus University, Aarhus, Denmark Purpose/Objective: Digital patient phantoms enable robust validation of new physics models, AI algorithms and treatment planning systems. They are also of great use in Virtual Clinical Trials (VCTs), a game-changing approach in radiotherapy research [1]. VCTs leverage computer simulations and real-world data to enhance trial design, de-risk new technologies in silico, and Proffered Paper 4521 accelerate research timelines. However, the potential of VCTs is fundamentally constrained by the scarcity of large-scale, diverse, and realistic virtual patient cohorts. This study presents a proof-of-concept generative engine to develop methods for synthesising geometrically representative and automatically co- registered quad-modal patient phantoms (CT, T1w- MRI, T2w-MRI, and PET). To our knowledge, this is the first attempt at 3D generative synthesis across multiple modalities, enabling data generation for future VCTs.
Conclusion: This meta-analysis provides strong quantitative evidence that auto-planning achieves comparable target coverage and plan quality to manual planning, while delivering superior and more consistent OAR sparing. Coupled with anticipated planning efficiency, the implementation of AP systems is strongly supported to enhance plan quality, reduce inter- planner variability, and optimize clinical workflow. Keywords: Automated Planning, Meta-Analysis, Workflow AI-based synthetic CT on MRIdian enables accurate abdominal SBRT without dosimetric compromise Daniel NGUYEN 1 , Cristina BARCA 1 , Sena YOSSI 2 , Guillaume LARGERON 1 , Mustapha KHODRI 1 1 Physics, ORLAM, LYON, France. 2 Radiotherapy, ORLAM, Lyon, France Purpose/Objective: To evaluate the dosimetric accuracy of a commercial AI-based synthetic CT (sCT) solution generated from low-field MRI (TrueFISP, MRIdian 0.35 T) for abdominal stereotactic body radiotherapy (SBRT) in liver and pancreatic cancer patients. Material/Methods: Digital Poster 4436 Thirty consecutive patients treated between 2023 and 2025 on the MRIdian MR-Linac were retrospectively analysed: 20 pancreatic and 10 liver SBRT treatments. Synthetic CTs were generated from daily TrueFISP images using the commercial MVision sCT module (MVision AI, Finland). For each case, the treatment plan was recalculated on the corresponding sCT using the same dose grid and monitor units as on the clinical CT. Mean, minimum and maximum doses (Gy) were compared for the PTV, PTV High, duodenum, colon, spinal cord, small bowel, stomach and external contour. Paired t-tests assessed statistical differences. 3D gamma analyses (3%/3 mm, 2%/2 mm, 1%/1 mm) were performed to evaluate dose distribution agreement between CT and sCT. Results: Across all 30 patients, no statistically significant
Made with FlippingBook - Share PDF online