Abstract
<title>Abstract</title> <p>Deep learning–based automated segmentation has the potential to improve radiotherapy planning workflows for glioblastoma; however, safe clinical integration requires not only high accuracy but also reliable performance across heterogeneous postoperative tumor regions. This study presents a clinically oriented reliability and failure-risk evaluation of nnU-Net v2 for automated multi-class segmentation of postoperative glioblastoma using multi-modal MRI. A total of 240 cases from the BraTS 2024 postoperative dataset were analyzed using four MRI modalities (T1, T1-Gd, T2, and FLAIR). Segmentation performance was assessed using a comprehensive multi-metric framework including Dice similarity coefficient, 95th percentile Hausdorff distance (HD95), mean surface distance (MSD), precision, and sensitivity. In addition to global accuracy, a clinically motivated failure-threshold analysis was performed to identify subregion-specific reliability patterns relevant to radiotherapy planning. The model demonstrated high and consistent performance for non-enhancing tumor regions, achieving Dice scores of 0.91 ± 0.05 for the non-enhancing tumor core and 0.90 ± 0.06 for surrounding FLAIR hyperintensity. In contrast, enhancing tumor and resection cavity segmentation showed reduced accuracy and greater variability, with Dice scores of 0.66 ± 0.12 and 0.57 ± 0.15, respectively. Failure-threshold analysis revealed substantial reliability limitations in enhancing tumor detection and cavity delineation, highlighting radiotherapy-relevant segmentation risks. These findings indicate that automated segmentation with nnU-Net can reliably delineate non-enhancing postoperative tumor regions, whereas delineating enhancing tumor and resection cavity contours requires careful expert verification. This study emphasizes the importance of failure-aware, multi-metric evaluation frameworks for the safe clinical integration of deep learning–based segmentation tools into postoperative glioblastoma radiotherapy workflows.</p>