Background: Glioblastoma (GBM) is the most aggressive primary brain tumor, and the Ki-67 proliferative index (PI) serves as a key biomarker for cellular proliferation. Current assessment of Ki-67 relies on postoperative immunohistochemistry, which is time-consuming and invasive. This study aimed to develop and validate a radiomics-based machine learning framework for the preoperative prediction of Ki-67 expression in GBM using multiparametric magnetic resonance imaging (MRI).

Methods: A total of 580 patients with isocitrate dehydrogenase–wildtype GBM who underwent maximal safe resection were retrospectively enrolled, comprising 300 in the training cohort, 127 in the internal validation cohort, and 60 in the external validation cohort. Volumes of interest were segmented on preoperative contrast-enhanced T1-weighted and FLAIR images using a deep learning–based framework into enhancing tumor, necrotic non-enhancing core, and peritumoral edema. From these, 4,227 radiomic features were extracted. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression, and multiple classifiers were tested. A two-class end-to-end framework integrating radiomics and clinical variables was developed, with model performance evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, and decision curve analysis (DCA).

Results: The LASSO combined with quadratic discriminant analysis (QDA) achieved the best predictive performance, with AUCs of 0.92, 0.88, and 0.86 in the training, internal, and external validation cohorts, respectively. The model maintained classification accuracy above 0.85 across all cohorts. Multivariable logistic regression identified age, Karnofsky Performance Status, and edema-based radiomic signature as independent predictors of high Ki-67 expression (all p < 0.05). A nomogram integrating these predictors demonstrated favorable calibration and clinical utility, while DCA demonstrated favorable clinical applicability.

Conclusion: This study presents and externally validates a robust radiomics-based machine learning model for preoperative prediction of Ki-67 expression in GBM. By combining deep learning–based segmentation, radiomics, and clinical variables, the LASSO + QDA model provides a noninvasive, individualized tool for proliferation assessment, with potential to guide surgical planning, risk stratification, and personalized treatment strategies.