2193 - Radiomics Analysis of the Tumor Bed on Computed Tomography for Survival Prediction in Glioblastoma

Purpose/Objective(s): Glioblastoma is a highly aggressive tumor with marked biological heterogeneity and poor prognosis. Early identification of patients at risk of rapid progression remains an unmet clinical need. The study hypothesis was that quantitative radiomic features extracted from postoperative tumor bed CT images are associated with overall survival and enable accurate short-term survival prediction. The primary objective was to develop and validate radiomics-based models to predict survival at 3, 6, and 12 months.

Materials/Methods: A retrospective cohort of 26 glioblastoma patients treated with radiotherapy was analyzed. A total of 107 radiomic features were extracted from tumor bed CT images using PyRadiomics, including shape, first-order, and texture features. Clinical variables included IDH mutational status and radiotherapy dose. Primary endpoints were survival beyond 3, 6, and 12 months. Logistic regression models with L1/L2 regularization were trained for each endpoint and internally validated using leave-one-out cross-validation. Model performance was assessed using AUC, sensitivity, specificity, and calibration (Brier score). Odds ratios (OR) were reported for relevant predictors. A Cox proportional hazards model was additionally used for continuous survival analysis, reporting hazard ratios (HR) with 95% confidence intervals (CI).

Results: The 3-month survival model demonstrated excellent discrimination with an AUC of 0.935 (95% CI: 0.814–1.000), sensitivity 0.787 (95% CI: 0.571–1.000), and specificity 0.896 (95% CI: 0.667–1.000), with excellent calibration (Brier score 0.108). Key predictors included shape elongation (OR 0.06), first-order mean intensity (OR 11.70), and GLCM sum entropy (OR 9.74). Six-month prediction showed moderate performance (AUC 0.712; 95% CI: 0.495–0.914). Twelve-month prediction was limited by class imbalance. In Cox analysis, glcm_Idn was associated with reduced mortality risk (HR 0.42; 95% CI: 0.21–0.81; p = 0.010), while firstorder_90Percentile (HR 1.70; 95% CI: 1.04–2.78; p = 0.034) and ngtdm_Contrast (HR 1.59; 95% CI: 1.04–2.43; p = 0.034) were associated with increased risk. The Cox model achieved a concordance index of 0.771.

Conclusion: CT-based radiomics of the postoperative tumor bed provides meaningful prognostic information in glioblastoma, particularly for short-term survival prediction. The strong performance of the 3-month model supports its potential use for early risk stratification. External validation in larger cohorts is required prior to clinical implementation.