Evaluation of compartmentalized automatic segmentation for definition of the GTV in glioblastoma radiotherapy.

Abstract

Manual delineation of target volumes in glioblastoma (GBM) radiotherapy (RT) is time-consuming and variable. This study evaluates the clinical applicability of a preliminary deep learning model (Neosoma Glioma) for automating gross tumor volume (GTV) segmentation in postoperative GBM per ESTRO-EANO guidelines. We retrospectively analyzed 100 GBM cases treated at Inselspital University Hospital, Bern (2016-2020) with standardized multi-modal MRI. Auto-segmented GTVs were compared to expert-defined contours using geometric metrics. Radiation oncologists reviewed and adjusted the best-performing configuration. Time savings, geometric similarity, and dosimetric impact were assessed. Optimal auto-segmentation (resection cavity plus enhancing tumor with 1 mm margin) achieved a mean Dice similarity coefficient of 0.79 (SD = 0.14) vs. ground truth. Manual adjustment took 5.9 (SD = 4.6) minutes vs. 12.3 (SD = 6.8) minutes for manual contouring (>50 % time reduction). The mean Dice between auto-segmented and adjusted GTVs was 0.84 (SD = 0.18). Dosimetric evaluation showed plans from adjusted auto-segmentations were equivalent to those based on consensus contours, with no clinically relevant differences in target coverage or organ-at-risk sparing. The Neosoma Glioma model generates clinically useful postoperative GTV segmentations, with geometric performance comparable to expert variability and dosimetric equivalence to consensus contours. It reduces contouring time by over 50%, enabling faster RT workflows. Its consistency across diverse GBM presentations supports its practical value. AI-based segmentation can help standardize GBM target definition when integrated into RT planning with proper quality assurance.