Post-operative CT radiomics for survival prediction in glioblastoma: a single-institution retrospective study

Background: Glioblastoma (GBM) remains the most aggressive primary brain tumor with a dismal prognosis. Radiomics has emerged as a promising approach for survival prediction, although most studies focus on magnetic resonance imaging (MRI). We investigated whether post-operative radiotherapy-planning computed tomography (CT) radiomics can predict survival in GBM patients. Materials and methods: We retrospectively analysed 26 patients with GBM who underwent surgery and completed standard chemoradiotherapy, including radiotherapy (60 Gy in 30 fractions) with temozolomide. Radiomic features (n = 107) were extracted from non-contrast-enhanced post-operative CT scans of the tumor bed. Logistic regression models with elastic-net regularisation were developed to predict survival at 3, 6, and 12 months using leave-one-out cross-validation. A Cox proportional hazards model was built for continuous survival prediction. Results: The 3-month survival model achieved an area under the curve (AUC) of 0.935 [95% confidence interval (CI): 0.81–1.00] with 78.7% sensitivity and 89.6% specificity. The 6-month model showed moderate performance (AUC = 0.712), while the 12-month model had limited discriminative ability (AUC = 0.568). The Cox radiomics model achieved a concordance index of 0.771. Median tumor-bed volume was 64.0 cm 3 [interquartile range (IQR): 25.6–99.5; range 8.5–211.8]. Conclusions: Post-operative CT radiomics demonstrated promising performance for short-term survival prediction in this single-institution cohort. Because of the small sample size, the high AUC of the 3-month model should be interpreted as exploratory and may partly reflect cohort-specific overfitting. External validation in larger cohorts is warranted.