SURG-61. Deep Learning-Driven Automated Assessment of Extent of Resection for Survival Prediction in Newly Diagnosed IDH-Wildtype Glioblastoma

Abstract BACKGROUND Extent of resection (EOR) is a well-established prognostic factor in glioblastoma (GBM), yet its measurement in clinical settings is often subjective and time-consuming. We aimed to develop a deep learning-based model for automated post-operative tumor segmentation and to validate its prognostic value in newly diagnosed IDH-wildtype GBM patients. METHODS A total of 195 patients with newly diagnosed IDH-wildtype GBM who underwent surgery at Seoul St. Mary’s Hospital between 2008 and 2023 were retrospectively reviewed. A cross-sequence voxelwise subtraction model using contrast-enhanced and non-enhanced T1-weighted MRI was developed to estimate residual enhancing tumor volume. EOR was categorized according to the RANO resect classification. Survival outcomes were evaluated using Kaplan–Meier analysis and Cox proportional hazards modeling, adjusting for clinical covariates. RESULTS Patients with greater extent of resection demonstrated superior survival outcomes. One-year overall survival (OS) ranged from 79.6% (Class 1/2) to 31.2% (Class 4), and median OS ranged from 21.2 months (Class 1/2) to 9.4 months (Class 4). One-year progression-free survival (PFS) ranged from 53.1% to 18.7%, and median PFS from 13.3 to 6.7 months across EOR classes. The automated EOR classification was significantly associated with both OS and PFS (log-rank p < 0.001). In multivariate Cox regression including age, MGMT methylation, and performance status, automated EOR remained an independent predictor of both OS and PFS (p < 0.01). CONCLUSION This is the first clinical validation of a deep learning-based, fully automated segmentation model for EOR assessment using real-world postoperative imaging. The model provides accurate and clinically relevant stratification of GBM patients, with performance comparable to manual expert assessment. Integration of this tool in routine workflow could facilitate rapid and objective evaluation of surgical outcomes. Further external validation and expansion to T2-based segmentation are ongoing.