ESTRO 2025 - Abstract Book

S2477

Physics - Autosegmentation

ESTRO 2025

References: [1] Chen L, Platzer P, Reschl C, Schafasand M, Nachankar A, Lukas Hajdusich C, Kuess P, Stock M, Habraken S, Carlino A. Validation of a deep-learning segmentation model for adult and pediatric head and neck radiotherapy in different patient positions. Phys Imaging Radiat Oncol. 2023 Dec 27;29:100527. doi: 10.1016/j.phro.2023.100527.

2499

Poster Discussion Evaluating cardiac substructure exposure and auto-contouring feasibility in paediatric patients treated with proton beam therapy Love Goyal 1,2 , Kathyryn Banfill 3 , Samuel Ingram 4,2 , Henry Mandeville 5 , Matthew Lowe 4,2 , Eliana Vasquez Osorio 2 , Shermaine Pan 4 , Yuwei Wang 1 , Ed Smith 4 , Marianne Aznar 2 1 Proton clinical outcomes unit, The Christie NHS Foundation Trust, Manchester, United Kingdom. 2 Faculty of biology medicine and health, Division of cancer sciences cancer sciences, The University of Manchester, Manchester, United Kingdom. 3 Department of clinical oncology, The Christie NHS Foundation Trust, Manchester, United Kingdom. 4 Proton beam therapy, The Christie NHS Foundation Trust, Manchester, United Kingdom. 5 Department of clinical oncology, The Royal Marsden NHS Foundation Trust, London, United Kingdom Purpose/Objective: Paediatric patients treated with thoracic radiotherapy are at risk of radiation-induced cardiovascular disease but published dose-response relationships are based on older radiotherapy techniques (1). This study aims to quantify the exposure to cardiac substructures in children treated today with Proton Beam Therapy (PBT) and assess performance of a commercially-available deep learning-based auto-contouring solution (trained on adult data) in this population. Material/Methods: We retrospectively identified paediatric and young adults patients receiving PBT to the thorax. All patients were treated with pencil beam scanning proton therapy (Eclipse v18.0). Heart and cardiac substructures (Figure 1) were retrospectively delineated on contrast-enhanced planning CT scans by a clinical oncologist (and peer-reviewed by two independent oncologists) using a published cardiac atlas (2). We could not reach a consensus on mitral and tricuspid valves and AV node; hence these were excluded from the analysis. AI-based contours were generated on the same CT using Limbus AI (v1.8.0). AI performance was quantitatively assessed by comparing with the manual contours using the Dice similarity coefficient, Hausdorff distance and distance to agreement. Mean and maximum doses to all structures and contours were recorded. Mann-Whitney U test and Paired T-test were used to assess statistical significance in dose to these structures between manual and AI contours. Results: Twenty patients (8 females, 12 males) were included in the analysis. The median (range) age at the start of PBT was 13.5 years old (2-18). Seven patients were treated with combined photon and protons and thirteen with protons alone. Median (range) prescription dose was 52.2 Gy (19.8, 70.2) in 29 (11- 39) fractions. Several structures of interest (e.g., left circumflex and right coronary artery) were not available in the AI model, or contoured as composite structures and hence excluded from the comparison. Failures of the AI model, defined as “no AI contour produced” or “AI contour produced with no overlap of the clinical structure of interest”, were observed in 3 patients aged 2, 4 and 6 years old (Figure 1). Differences in dose estimates from the AI model were considerable (Figure 2), reaching 3.5 Gy for mean heart dose and mean left ventricle dose and 7.5 Gy for mean left atrium dose.