ESTRO 2024 - Abstract Book

S3043

Physics - Autosegmentation

ESTRO 2024

Skłodowska-Curie National Research Institute of Oncology, 5. IIIrd Radiotherapy and Chemotherapy Department, Gliwice, Poland. 6 Maria Skłodowska-Curie National Research Institute of Oncology, Radiotherapy, Gliwice, Poland. 7 University Hospital of Zurich, Radiation Oncology, Zurich, Switzerland. 8 Medical University of Gdańsk, Oncology and Radiotherapy, Gdańsk, Poland

Purpose/Objective:

Stereotactic arrhythmia radio-ablation (STAR) is increasingly used to treat ventricular tachycardia (VT), especially recurrent VT after catheter ablations. The most common STAR treatment protocol involves delivering 1x25 Gy to the arrhythmia substrate, with maximum doses reaching up to 40 Gy. Applying such high dose levels in the heart raises concerns regarding potential treatment-induced cardiotoxicity. Currently, only whole heart dose constraints originating from thoracic radiotherapy are used while cardiac substructure (CS) dose constraints are uncommon. A substantial obstacle to understanding CS dose-effect relationships is the lack of consistent, high-quality delineations. CS are not routinely contoured in clinical practice, as manual contouring is labour-intensive and offers no immediate clinical benefit. The CS contouring challenges are compounded in STAR patients who typically suffer from structural heart disease and have implantable cardioverter defibrillator (ICD) with cardiac leads that induce artifacts on CT.

In this study, we develop and evaluate an auto-contouring tool specifically geared towards CS and major vessels delineations in VT patients using a deep learning model trained on a high-quality, multi-institutional data set.

Material/Methods:

Six experienced centres within the STOPSTORM.eu consortium provided delineations of all relevant cardiac and vessel structures in institutionally-treated VT patient cases following established guidelines [1]. These delineations included 16 structures: four major vessels (aorta, inferior/superior vena cava, and pulmonary artery), four valves (aortic, pulmonic, mitral, and tricuspid), four coronary artery (CA) segments (left main, left anterior descending, left circumflex, and right coronary artery) and the four chambers. Twenty-eight contour sets were delineated based on 27 planning CTs and 14 contrast-enhanced CTs of VT patients. In addition, CT planning scans and contours (limited to the four major vessels, the aortic valve and four chambers) were available for a single-institutional cohort of 70 lung cancer patients. Four different auto-contouring networks were trained (all nnUNets [2]) for 1000 epochs using 5-fold cross-validation: model A included only the lung cancer patient data, model B included the lung cancer patient data plus 10 representative VT patients, and model C was based on all available data. In addition, model C* was trained on VT patient data only to contour the smallest CS (valves, CAs). The performance of these models was evaluated based on Dice score (DSC) and 95% Hausdorff distance (HD95).

Model performance was evaluated on three independent VT patient test cases. The test patients were previously delineated by 20-24 STOPSTORM.eu consortium centres.

All VT patients in the training and test set had ICDs. Both the training (Models B, C and C*) and the test sets had one patient with a left ventricular assistant device (LVAD). The LVAD caused prominent image artefact which were much more severe compared to the ICD leads.

Results: