ESTRO 2025 - Abstract Book

S2505

Physics - Autosegmentation

ESTRO 2025

3032

Proffered Paper Available solutions for auto-contouring of heart substructures: trap or truth?

Alexandra Moignier 1 , Elise Prangères 1 , François Thillays 2 , Bastien Bernard 3 , Angela Botticella 4 , Robert Finnegan 5 , Sandrine Huger 6 , Anna Karlhede 7 , Thomas Lacornerie 8 , Fredrik Löfman 7 , Jérémy Palisson 9 , Charlotte Robert 10 , Killian Sambourg 10 , Jonas Söderberg 7 , Remus Stoica 11 , Grégory Delpon 1 , Tanguy Perennec 2 , Loïg Vaugier 2 1 Medical Physics Department, Institut de cancérologie de l'Ouest, Saint Herblain, France. 2 Radiation Oncology Department, Institut de cancérologie de l'Ouest, Saint Herblain, France. 3 Siemens Healthineers, Siemens Healthineers, Saint Denis, France. 4 Radiation Oncology Department, Institut Gustave Roussy, Villejuif, France. 5 Medical Physics Department, Northern Sydney Cancer Centre, Sydney, Australia. 6 Medical Physics Department, Institut de cancérologie de Lorraine, Vandoeuvre-lès-Nancy, France. 7 Machine Learning Department, RaySearch Laboratories, Stockholm, Sweden. 8 Medical Physics Department, Centre Oscar Lambret, Lille, France. 9 Radiation Oncology Department, Centre de la baie, Avranches, France. 10 Medical Physics Department, Institut Gustave Roussy, Villejuif, France. 11 Synaptiq, Synaptiq, Cluj-Napoca, Romania Purpose/Objective: Moving beyond mean heart dose for evaluating radiotherapy toxicity is now widely accepted [Jacob et al., 2019; Jones et al., 2024]. However, consensus on dose constraints for cardiac substructures remains lacking. Building evidence for dose limits requires accurate and standardized contours, which are currently unavailable for many heart substructures in clinical practice. This study aimed to compare CT-based auto-contouring of heart substructures from currently available solutions and to analyse their sensitivity to the presence/absence of contrast enhancement (CE). Material/Methods: Twenty CE-CT scans and twenty non-CE-CT (NCE-CT) scans were selected for their sequential acquisition, in inspiratory apnea, during a single session (identical positioning/anatomy). Seven commercial and three open-source solutions for automatic heart substructure delineation (all AI-based except one hybrid) were applied to this dataset. One solution (RayStation, RaySearch Laboratories) was arbitrarily chosen as reference for comparison purposes on CE-CT, using the mean surface distance (MSD) and the 95th percentile of the Hausdorff distance (HD95). To evaluate the impact of NCE on delineation performance, volume ratios between NCE-CT and CE-CT of each heart substructure from each solution were calculated. After harmonizing nomenclature, the structures of interest included: the heart, heart with pulmonary artery, left and right atria and ventricles, myocardium (Ventricle_L_Myo), aorta and pulmonary arteries, pulmonary vein, inferior and superior vena cava, left main/left anterior descending (A_LM_LAD), circumflex (A_Cflx) and right (A_Coronary_R) coronary arteries, and mitral valve. Results: The MSD is below 10 mm except for the A_Aorta of Radformation and the A_Aorta, A_Pulmonary, V_VenaCava_S, A_Cflx, A_Coronary_R of Platipy. According to the HD95, significant differences are observed depending on the heart substructure and the solution (Table 1). For example, the heart base and the left ventricle differ due to varying definitions from the modelling process, while the coronary arteries exhibit location variability. The substructures volumes can substantially differ between NCE-CT and CE-CT (Table 2), with some solutions (e.g. Limbus AI and Radformation) exhibiting their difficulty to manage such differences.