S1587
Physics - Autosegmentation
ESTRO 2026
root) including a recommended threshold dose (maximum dose to 1cc <19.5Gy for 55Gy in 20 fraction regimens) for thoracic radiotherapy2. Autocontouring tools, including open-source solutions such as TotalSegmentator³, are increasingly adopted in clinical workflows and may provide accuracy comparable to commercial alternatives. This study evaluates TotalSegmentator’s performance to generate high- quality AI cardiac substructure contours. Material/Methods: A dataset of 19 patients3 with ground truth (GT) contours for cardiac substructures delineated against published cardiac radiotherapy atlases4 was available for AI-contour quality evaluation. AI-contours using TotalSegmentator were generated for available cardiac substructures (whole heart, aorta, atria, ventricles, pulmonary artery, superior vena cava) and compared with GT-contours. Volumetric Dice Similarity Coefficients (VDSC) were calculated for geometric comparison. Dosimetric changes observed when using AI-contours compared to GT were assessed for the right atrium (CAA substructure and therefore the most clinically important currently available structure) for maximum dose to 1cc and mean dose. Paired AI- contour and GT-contour dosimetric values were compared using Wilcoxon’s signed-rank test for paired non-parametric data to determine whether statistical significance was observed for dosimetric differences. Differences were considered clinically significant if they crossed the published dose constraint. Results: Median VDSC values observed ranged from 0.66 to 0.86 (Table 1, Figure 1), with values >0.8 for 4 of 8 structures. Best values were observed for aorta (0.86), with worst performance seen for left ventricle (0.66). Median right atrium VDSC was 0.82.On dosimetric comparison, the right atrium 1cc maximum dose for AI- and GT-contours crossed the published dose constraint in 1 of 19 cases. No statistically significant dosimetric differences was seen between AI-contours and GT-contours for maximum 1cc (p=0.326) or mean dose (p=0.126).
Conclusion: The model was designed to reproduce particular delineation strategy. Auto-segmentation produced target contours in ~1 minute, achieving high DSCs and small volume differences compared to contours created by a physician including detection of sub ‐ centimeter lesions. The model shows a slight tendency to overestimate volumes and offers a useful "second reading" without delaying planning. In cases 6 and 7, the auto-segmentation flagged 3 lesions that were confirmed as previously undetected metastases. Keywords: brain, metastases, nnU-Net Digital Poster 3649 Evaluation of an open-source AI autocontouring tool for cardiac substructure delineation in radiotherapy treatment planning Tom Young 1,2 , Anil Mistry 3 , Conor Hudin 4 , Caroline Sisodia 4 , Michelle Stenson 4 , David Cutter 5 , Dika Vilic 3 , Christopher Thomas 4 , Teresa Guerrero Urbano 1,2 , George Ntentas 4,5 , Shahreen Ahmad 1 1 Department of Clinical Oncology, Guy´s and St Thomas´ NHS Foundation Trust, London, United Kingdom. 2 School of Cancer & Pharmaceutical Sciences, King's College London, London, United Kingdom. 3 Clinical Scientific Computing, Guy´s and St Thomas´ NHS Foundation Trust, London, United Kingdom. 4 Department of Medical Physics and Clinical Engineering, Guy´s and St Thomas´ NHS Foundation Trust, London, United Kingdom. 5 Nuffield Department of Population Health, University of Oxford, Oxford, United Kingdom Purpose/Objective: There is increasing recognition of the importance of reducing radiation dose to cardiac substructures and that this may improve survival outcomes1. However, delineation and dose optimisation of substructures are not yet routine in UK radiotherapy practice. Recent work defined a consensus cardiac avoidance area (CAA) at the heart base (right atrium, proximal portions of left and right coronary arteries, aortic valve
Made with FlippingBook - Share PDF online