S1930
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
Digital Poster 3332 Evaluation of AI-Based Segmentation and Synthetic CT Generation for MR-Only Radiotherapy Planning Using the Gold Atlas Dataset Amanda Östensson, Joakim Jonsson Department of Diagnostics and Intervention, Radiation Physics, Umeå University, Umeå, Sweden Purpose/Objective Magnetic resonance imaging (MRI) offers superior soft-tissue contrast compared with computed tomography (CT) and is increasingly used for delineations in radiotherapy. Conventional workflows require CT-MRI registration which introduces geometric uncertainty that can propagate to dose delivery [1,2] . T2w MR-only radiotherapy, based on synthetic CT (sCT) and MRI-based AI-segmentations, offers a streamlined single-modality approach with potential to eliminate registration errors [2,3] , reduce dose to patients [4] , and benefit adaptive radiotherapy with an MR-linac. The objective of this study is to evaluate whether an MR-only workflow using CE- marked models (Workspace+, MVision AI, Helsinki, Finland) achieves equivalent performance to a conventional CT-based workflow. Material/Methods We analysed 19 male pelvis patients from the Gold Atlas dataset [5] , containing deformably co-registered CT and MRI with multi-observer and consensus segmentations. We compared AI-generated contours against five manual delineators using Surface Dice (2mm), Hausdorff distance (HD95), and volumetric difference metrics. Wilcoxon signed-rank tests assess statistical differences. A Turing test was designed to investigate if AI-segmentations could be distinguished from human observers based on multi-metric similarity. We compared synthetic CTs to deformed CTs by mean absolute error (MAE), peak signal-to- noise ratio (PSNR), and structural similarity index (SSIM). We evaluated dosimetric equivalence through VMAT plan recalculation on both CT and sCT using identical parameters, assessing 3D gamma pass rates (1%/1mm, 2%/2mm), voxel-wise dose differences, and DVH parameters for a total of five targets and organs at risk. Results Results show that AI-based segmentation achieved accuracy comparable to, or within the range of inter- observer variability across all organs. Median Dice coefficients exceeded 0.9 for the prostate and bladder, with distance metrics closely matching the manual results. The Turing test showed that AI contours were indistinguishable from human observers for all organs except for the seminal vesicles, with a score slightly above chance. Synthetic CTs showed high similarity to
rate, etc.). The metrics were obtained using a Matlab script, planAnalyzer_v6 (given by San Joan de Reus Hospital). Results: The comparative analysis showed that the MU/cGy and PM indices are similar between both TPSs, while the MCS was consistently higher in Monaco, indicating that the MLC aperture is comparable but Eclipse tends to vary the MLC shape more between control points. Eclipse exhibits a more stable dose rate, generally maintained at its maximum value, which reduces treatment duration. Moreover, the gantry arm speed varies less in Eclipse. These differences may influence the mechanical and dosimetric behavior of the equipment, opening the door to future research related to plan verification.
Conclusion: The solutions provided by both systems for the same cases and treatment unit are different. Monaco tends to generate simpler MLC shapes, while Eclipse appears to offer shorter and more stable results in terms of treatment time, dose rate variability, and gantry speed. Both systems are capable of achieving the planning objectives with the appropriate level of expertise. References: [1]McNiven, Med Phys.37 (2010)[2]Du, Med Phys.41 (2014) Keywords: comparision, monaco, eclipse
Made with FlippingBook - Share PDF online