ESTRO 2025 - Abstract Book

S2521

Physics - Autosegmentation

ESTRO 2025

physician references. The image quality parameters cohort (20 patients: H&N, pelvis, brain, thorax) evaluated 12 reconstructions per patient using CT-scan commercial solutions (Philips): Standard, iDose, and IMR algorithms, with noise and spatial resolution (SR) degradations simulated via Python scripts (v3.11.5). AI performance was assessed using DSC and HD95%. Statistical significance was determined via the Wilcoxon signed-rank test. Results: Software Updates Cohort: DSC increased with successive software versions for re-trained structures (mean DSC ≥0.75 for all contours). Breast contour DSC was reduced by 1% between v1.5 and v1.8B3 (p>0.05). Mean HD95% values were <3mm, <12mm, and <15mm for H&N, pelvis, and thorax contours, respectively, across all versions (p>0.05). In image quality parameters cohort , reconstruction algorithms (Standard, iDose, IMR) showed no significant differences in performance. The DSC (HD95%) was ≥0.9 (≤4mm) for 97% (90%) of contours with iDose and 93% (81%) with IMR (p<0.001). Noise introduction led to a performance degradation: DSC≥0.9 was observed for 89%, 58%, 34%, and 24% of contours at 2%, 5%, 10%, and 20% noise levels, respectively (p<0.001). HD95% ≤4mm was observed for 85%, 50%, 26%, and 13% of structures under these conditions (p<0.001). An SR degradation of 1.2 resulted in DSC≥0.9 for 87% and HD95%≤4mm for 83% of structures (p<0.001). At a 1.4 degradation level, these values dropped to 70% (p<0.001). Conclusion: Software updates consistently enhanced AI contouring accuracy for re-trained structures without compromising non-retrained ones. AI-based contouring demonstrated resilience to minor reconstruction algorithm variations but was significantly impacted by increased noise and SR degradation. These findings highlight the need for quality control of AI-generated contours following software updates and vigilance in daily practice. Further research should explore additional anatomical regions in larger patients’ cohorts. Proffered Paper An automated method to assess the growth-rates of vertebrae in children: A framework to evaluate growth retardation in Paediatric radiation therapy Kartik Kumar 1,2 , Adam Yeo 1,2,3 , Tomas Kron 1,2,3 , Rick Franich 1,2 1 School of Science, RMIT University, Melbourne, Australia. 2 Physical Sciences Department, Peter MacCallum Cancer Centre, Melbourne, Australia. 3 Sir Peter MacCallum Department of Oncology, University of Melbourne, Melbourne, Australia Purpose/Objective: In paediatric radiotherapy, craniospinal irradiation is associated with adverse late effects such as growth retardation, scoliosis, and kyphosis [1]. Traditional methods for assessing growth retardation, using patient height measurements, do not adequately reflect the effects on individual vertebral bodies. In this study, an automated framework is developed to assess the growth-rates of vertebral heights on CT datasets. Material/Methods: The framework comprised two parts: a) vertebrae contouring and b) landmark extraction (Figure 1). For contouring, an extended nnU-Net model was trained using 250 CT images (170 adult and 80 paediatric) to segment vertebrae [2]. The contours were used to extract six landmarks on the sagittal central slice of each vertebra using a Python based script (Figure 1). The model was evaluated on 50 paediatric test cases using the Dice similarity coefficient (DSC) between model contours and manual contours of vertebrae. The landmarks were manually inspected and Keywords: Autosegmentation, AI, Image Quality 3358