P03.10.A AUTOMATIC SEGMENTATION OF CLINICAL TARGET VOLUME IN GLIOBLASTOMA FOR RADIOTHERAPY PLANNING

Abstract BACKGROUND While robust methods exist for segmenting the gross tumor volume (GTV) in glioblastoma MRI, clinical target volume (CTV) delineation remains challenging and largely manual. The ESTRO-EANO guidelines recommend a 15mm isotropic GTV expansion, constrained by anatomical boundaries, to define the CTV, based on clinical outcomes and reduced toxicity compared to older 20mm margins. Although biophysical models offer potential for glioma margin definition, their ill-posed nature hinders clinical adoption. This study investigates the feasibility of automating CTV segmentation in multi-parametric MRIs using a deep learning (DL) pipeline aligned with updated guidelines. MATERIAL AND METHODS We utilized the BraTS 2024 dataset to generate CTV pseudo-reference masks. Binary GTV masks were derived from combined enhancing tissue, non-enhancing core, and resected cavity labels. A rule-based approach computed Euclidean distance maps from the GTV surface, constrained by automatically segmented anatomical structures (cerebral hemispheres, brainstem, falx, cerebellum, chiasm) using a SynthSeg-based model. CTV masks were defined by thresholding the constrained maps at 15mm, excluding optic nerves, eyes, and lens in skull-stripped MRIs. The accuracy of this rule-based generation was assessed on the clinical Burdenko dataset (180 subjects). Supervised DL models (nnU-Net ResENC and MedNeXt) were trained on 1350 BraTS pseudo-reference masks to explore automated segmentation, circumventing the rule-based steps. The trained models were then evaluated on the clinical Burdenko data. RESULTS The unsupervised rule-based pipeline achieved a Dice score of 0.82±0.08 on the clinical Burdenko dataset. Supervised DL models trained on pseudo-references yielded a comparable Dice score of 0.81±0.08 on the clinical Burdenko dataset, indicating performance similar to the rule-based extension while respecting surrounding tissues. High agreement (0.95±0.04 Dice score) was observed between DL predictions and pseudo-references on the BraTS data. The performance difference between BraTS and Burdenko datasets is attributable to the synthetic nature of BraTS pseudo-labels compared to real clinical labels in Burdenko. CONCLUSION The promising results achieved in our preliminary experiments highlight the potential of automatic techniques for the segmentation of glioma beyond the visible GTVs. Employing larger-scale labeled datasets for guiding the segmentation and further evaluation of clinical datasets will be held in our future studies.