S2351
Physics - Quality assurance and auditing
ESTRO 2026
managing inter- and intrafractional anatomical variations during radiotherapy. Central to oART is the synthetic CT (sCT), derived from CBCT with corrected Hounsfield Units (HU) for accurate dose computation and optimization. While several sCT generation algorithms have been clinically validated, patient- specific quality assessment (PSQA) methods remain under development, often relying on in-house solutions. Building on prior work proposing fast PSQA for MRI-based sCT [1], we present a rapid PSQA approach for CBCT-based sCT and evaluate its sensitivity to deliberately introduced errors. Material/Methods: The study included 10 prostate patients, each with five high-quality CBCTs (50 images total). To assess the sCT image quality, we compared the dose distributions on the sCT with the one on the CBCT. First, sCT images were generated using RaySearch’s validated corrected CBCT algorithm. Key anatomical structures were segmented via deep learning. A deformable image registration (DIR) between the planning CT and sCT, guided by those structures, was used to propagate the target. Structures were rigidly copied from sCT to CBCT, and densities were assigned via bulk density mapping. Gamma passing rates (GPR, 2%/2mm) and relative dose differences ( Δ D = dose on sCT – dose on CBCT, normalized to the fraction dose) for PTV mean, PTV2%, PTV98%, Rectum2%, Bowel mean, and Bladder2% were calculated. Acceptance criteria were defined from the Δ D distribution. The workflow was repeated on two intentionally faulty cases: one with a hip prosthesis and another with an error in the external segmentation. Results: For the 50 images, mean and standard deviation (std) of GPR and Δ D were calculated. Δ D for the bowel bag mean showed wide variability, likely due to air bubble artifacts, and was excluded from criteria. Acceptance thresholds were defined as μ ± 3 σ for Δ D and >97% for GPR. In the faulty cases, Δ D exceeded 4% and GPR dropped to 90% and 92%, clearly appearing as outliers in the distribution and boxplots.As a guideline for clinical implementation, a common threshold of [- 2%, 3%] for Δ D could be used.
sensitivity versus miss rate per error type, highlighting that model-related deviations were more frequently missed.
Conclusion: This study shows that QA of AI segmentation can detect clinically relevant errors in prostate cancer radiotherapy, with effectiveness varying by organ and error type. Anatomy related deviations were generally identified more reliably than subtler model related over or under segmentations. All anatomy related errors in the femoral heads were successfully detected, while for the bladder and anorectum, pronounced anatomy related errors were more consistently identified than model related deviations. Combining binary classification with error type analysis provides insight into the strengths and limitations of current quality assurance approaches and supports safer and more efficient clinical workflows. Further studies focusing on bladder and anorectum are warranted to validate these findings. References: [1] van Aalst JE, Maruccio FC, Simo ẽ s R, Janssen TM, Wolterink JM, van Ooijen PMA, Brouwer CL. Reliability of uncertainty quantification methods for deep learning auto-segmentation in head and neck organs at risk. Phys Med Biol. 2025 Oct 17;70(20). doi: 10.1088/1361-6560/ae110c. PMID: 41061736.[2] Claessens M, Vanreusel V, De Kerf G, Mollaert I, Löfman F, Gooding MJ, Brouwer C, Dirix P, Verellen D. Machine learning-based detection of aberrant deep learning segmentations of target and organs at risk for prostate radiotherapy using a secondary segmentation algorithm. Phys Med Biol. 2022 May 27;67(11). doi: 10.1088/1361-6560/ac6fad. PMID: 35561701. Keywords: Prostate cancer, AI segmentation, Error detection Digital Poster 1587 Patient-specific quality assurance method for pelvic cone beam CT-based synthetic CT Joachim Marichal 1 , Michaël Claessens 2 1 Faculty of Medicine and Health Sciences, University of Antwerp, Antwerp, Belgium. 2 Department of Radiation Oncology, Iridium Netwerk, Antwerp, Belgium Purpose/Objective: Cone-Beam CT (CBCT)-based online Adaptive Radiotherapy (oART) is a promising approach for
Table 1 – Mean, std, acceptance thresholds and values of the two test cases for the different metrics
Made with FlippingBook - Share PDF online