ESTRO 2025 - Abstract Book

S2467

Physics - Autosegmentation

ESTRO 2025

2145

Digital Poster Objective evaluation of automatic target Segmentation for brain and head & neck tumors Mehdi Astaraki 1,2 , Iuliana Toma-Dasu 1,2 1 Medical Radiation Physics, Stockholm University, Stockholm, Sweden. 2 Oncology-Pathology, Karolinska Institutet, Stockholm, Sweden Purpose/Objective: The advances in deep learning (DL) models have resulted in a paradigm shift in autosegmentation approaches for radiation treatment (RT) planning systems. While robust DL models for organ-at-risk segmentation are already integrated into commercial tools, accurate tumor delineation remains challenging for various cancers, often necessitating manual contouring. Numerous DL models have shown promise in segmenting gross tumor volume (GTV) on limited, and private datasets; nevertheless, their generalizability and clinical applicability require rigorous evaluation. International benchmarking challenges provide a crucial platform for such an objective and comparative assessment. This study aims to evaluate the performance of the models we developed for the brain, and head&neck tumor segmentations and report our successful contributions to the MICCAI2024 challenges. 1.The BraTS Adult Glioma Post-treatment(GLI) Challenge[1] on segmenting glioma from four MRI sequences into sub-regions: non-enhancing tumor core(NETC), surrounding non-enhancing FLAIR hyperintensity(SNFH), enhancing tissue(ET), resection cavity(RC), and tumor core(ET+NETC), utilizing 1350 training and 188 validation cases. 2.The BraTS Pediatric Tumor(PED) Challenge[2] aims to segment pediatric glioma into ET, NETC, cystic component(CC), and SNFH from 261 training and 91 validation MRI sequences. 3.The BraTS Meningioma RT(MenRT) Challenge[3] focused on segmenting GTV from T1-contrast MRI of 500 training and 70 validation subjects. 4.The HNTS-MRG Challenge[4] on the segmentation of primary (GTVp) and metastatic lymph node (GTVn) from pre RT and mid-RT T2-weighted MRI of 150 training and 50 testing cases. All datasets underwent preprocessing, including intensity harmonization, bias-field correction, and maximal volume cropping. Several established DL models were examined: SegResNet, MedNeXt, nnU-Net, and U-Mamba. Following training and evaluation, the most robust models were selected and ensembled to enhance generalization. Results: Table 1 shows the performance of the evaluated models on the validation sets. Figure 1 shows examples of the predicted segmentation masks. Table1.Segmentation powers in terms of Dice metric (μ±σ) Material/Methods: The datasets used in this study are detailed below.