3057 - Multicenter Comparison of Three Commercial Deep Learning-Based Cardiac Auto-Segmentation Algorithms Using the French RTEP7 Cohort
Presenter(s)
A. Licha1, A. Beddok2, C. Bartau3, L. Haas4, S. Klotz5, R. Modzelewski6, A. Moignier7, D. Richard6, L. Vaugier8, S. Thureau9, and P. Giraud10,11; 1Institute Jean Godinot, Reims, France, 2Department of Radiation Oncology, Institut Godinot, Reims, France, 3Aquilab – OncoPlace, Paris, France, 4Centre Henri Becquerel, Le Pre St Gervais 93310, France, 5Hôpital Européen Georges Pompidou, PARIS, France, 6Centre Henri Becquerel, Rouen, France, 7Institut de Cancérologie de l'Ouest René Gauducheau, Nantes, France, 8Institut de Cancérologie de l'Ouest René Gauducheau, France, Saint Herblain, France, 9centre henri becquerel, rouen, France, 10Hôpital Européen Georges Pompidou, Université Paris Cité, Service d'Oncologie Radiothérapie, Paris, France, 11Paris Cité University, Paris, France
Purpose/Objective(s):
Radiation-induced cardiac toxicity is a growing concern in thoracic radiotherapy, with increasing evidence linking dose to specific cardiac substructures to adverse outcomes. The French prospective multicenter RTEP7 trial (Lancet Oncol. 2024;25:1176–87), provides a randomized Phase II protocol-harmonized thoracic radiotherapy cohort across 19 institutions, enabling standardized cardiac dose analyses and controlled comparison of auto-segmentation algorithms. We compared the geometric performance of three commercially available deep learning-based cardiac auto-segmentation tools: RayStation, MVision AI, and Limbus AI. RayStation was used as the reference, supported by prior multicenter expert-reviewed validation (Moignier A et al, Submitted 2025).Materials/Methods:
Patients from the prospective RTEP7 cohort (n=158) were retrospectively analyzed. After predefined exclusions and dataset reconciliation, 129 patients were retained for the final comparative analysis. Cardiac substructures were automatically segmented using RayStation, MVision AI, and Limbus AI. Structure nomenclature was standardized according to TG-263. When a given anatomical structure was provided as multiple subcomponents in some solutions, Boolean unions were performed to create comparable composite structures using ARTWEB within ONCOPLACE (Aquilab). Geometric agreement was evaluated using Dice Similarity Coefficient (DSC), with pairwise comparisons using RayStation as reference across 11 harmonized substructures available in all three solutions.Results:
Across the 11 predefined harmonized cardiac substructures, median Dice Similarity Coefficient (DSC) was 0.81 (IQR 0.61-0.90) for RayStation vs MVision, 0.80 (IQR 0.56–0.88) for RayStation vs Limbus, and 0.86 (IQR 0.80-0.92) for Limbus vs MVision. Agreement versus RayStation was high for major structures, including the whole heart (MVision: 0.91 [0.90-0.93]; Limbus: 0.89 [0.87-0.91]) and the left ventricle (MVision: 0.92 [0.91-0.94]; Limbus: 0.92 [0.91-0.93]). In contrast, DSC was markedly lower for substructures that are either definition-dependent (inferior vena cava: MVision 0.34 [0.26-0.41]; Limbus 0.33 [0.24-0.39]) or small/complex (left anterior descending coronary artery: MVision 0.53 [0.42-0.59]; Limbus 0.47 [0.34–0.54]). Notably, agreement between MVision and Limbus was higher for the inferior vena cava (median DSC 0.84 [0.79-0.87]), consistent with systematic differences in definition/extent across solutions for selected vessels.Conclusion:
Commercial deep learning-based cardiac auto-segmentation tools showed high agreement for major cardiac structures but meaningful discrepancies for smaller, definition-dependent substructures. These differences likely reflect anatomical definitions and may affect dose extraction. Based on prior multicenter expert-reviewed validation, RayStation was retained as the reference framework for subsequent RTEP7 dose-response analyses.