ESTRO 2025 - Abstract Book

S2472

Physics - Autosegmentation

ESTRO 2025

Keywords: Autosegmentation, H&N vessels, deep learning

References: [1] Çiçek, Ö.; Abdulkadir, A.; Lienkamp, S.S.; Brox, T.; Ronneberger, O. 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, Athens, Greece, 17–21 October 2016; pp. 424–432.

2442

Digital Poster Toward automatic delineation of Clinical Target Volume in glioblastoma Mehdi Astaraki 1,2 , Iuliana Toma-Dasu 1,2 1 Medical Radiation Physics, Stockholm University, Stockholm, Sweden. 2 Department of Oncology and Pathology, Karolinska Institutet, Stockholm, Sweden Purpose/Objective: Deep learning (DL) models have enabled robust and accurate automatic segmentation of gross tumor volume (GTV) of glioblastoma in MRIs. However, clinical target volume (CTV) delineation, accounting for microscopic tumor spread beyond the GTV, remains challenging and predominantly reliant on manual efforts. To standardize this process, the ESTRO-EANO guidelines [1] recommend a 1.5cm volumetric expansion of the GTV constrained by anatomical boundaries, reflecting observations of comparable clinical outcomes with reduced toxicity compared to previous 2cm margins. This study investigates the feasibility of automating CTV segmentation in MRIs by developing a pipeline conforming to the updated guidelines. Material/Methods: Utilizing the BraTS 2024 Adult Glioma Post-operative dataset [2], binary GTV masks were generated by combining enhancing tissues, non-enhancing tumor core, and resected cavity labels. Euclidean distance maps were computed toward outward normals to represent the extension from the GTV surface, constrained by six segmented anatomical structures within skull-stipped volumes using an in-house model and SynthSeg [3]. CTV masks were then defined by binarizing the constrained extension maps at 1.5cm (new-guideline) and 2cm (previous-guideline) thresholds. The accuracy of the proposed rule-based CTV approach was evaluated on the preprocessed Burdenko dataset (180 subjects) [4]. Furthermore, CTV masks generated from 1350 BraTS cases served as pseudo-references for supervised training of a DL pipeline, aiming to circumvent the stepwise rule-based procedure. The well-trained model was then tested on the Burdenko dataset. Figure 1 illustrates the predicted CTV masks against expert references for Burdenko cases.