Radiomics-based differentiation between glioblastoma and primary central nervous system lymphoma: CT vs MRI

BACKGROUND: To systematically evaluate and compare the diagnostic efficacy of radiomics models derived from noncontrast CT (NCCT) versus multiparametric MRI in differentiating glioblastoma (GBM) from primary central nervous system lymphoma (PCNSL). METHODS: In this retrospective, multicenter study, 543 patients with pathologically confirmed GBM (n = 401) or PCNSL (n = 142) were divided into 3 cohorts. 1084 quantitative features were extracted from contrast-enhancing (CE) and non-enhancing (NE) regions across NCCT and five MRI sequences (T2WI, T1WI, ADC, FLAIR, and CE-T1WI). Feature selection employed ANOVA, Kruskal-Wallis test, and recursive feature elimination, followed by nested cross-validation (5-fold outer, 3-fold inner) to construct four machine learning classifiers: support vector machine, linear discriminant analysis, logistic regression, and decision tree. Model performance was rigorously assessed through AUC, accuracy, sensitivity, specificity with bootstrap-derived 95% confidence intervals. The Shapley Additive Explanation (SHAP) analysis was employed to explore the interpretability of models. RESULTS: The CE-T1WI radiomics model demonstrated superior diagnostic capability, with its AUCs of train/internal test/external test in CE regions and NE regions were 0.962/0.963/0.907 and 0.966/0.892/0.867, respectively. Notably, the CT-based model was not significantly different from other MRI models except for CE-T1WI model. The AUCs of train/internal test/external test for CT model in CE and NE regions were 0.941/0.906/0.822 and 0.902/0.891 /0.782, respectively. CONCLUSIONS: Both NCCT and multiparametric MRI are valuable in identifying GBM and PCNSL. The CE-T1WI radiomics model has the best diagnostic efficacy.