Development and validation of a preoperative magnetic resonance imaging-based and machine learning model for the noninvasive differentiation of intracranial glioblastoma, primary central nervous system lymphoma and brain metastases: a retrospective analysis.

Background: Accurate preoperative identification of intracranial glioblastoma (GB), primary central nervous system lymphoma (PCNSL), and brain metastases (BM) is crucial for determining the appropriate treatment strategy. Purpose: We aimed to develop and validate the utility of preoperative magnetic resonance imaging-based radiomics and machine learning models for the noninvasive identification them. STUDY TYPE: Retrospective. POPULATION: We included 202 patients, including 71 GB, 59 PCNSL, and 72 BM, randomly divided into a training cohort (n =141) and a validation cohort (n = 61).FIELD STRENGTH/SEQUENCE: Axial T2-weighted fast spin-echo sequence (T2WI) and contrast-enhanced T1-weighted spin-echo sequence (CE-T1WI) using 1.5-T and 3.0-T scanners. ASSESSMENT: We extracted radiomics features from the T2 sequence and CE-T1 sequence separately. Then, we applied the F-test and recursive feature elimination (RFE) to reduce the dimensionality for both individual sequences and the combined sequence CE-T1 combined with T2.The support vector machine (SVM), k-nearest neighbor (KNN), and naive Bayes classifier (NBC) were used in model development. STATISTICAL TESTS: Chi-square test, one-way analysis of variance, and Kruskal-Wallis test were performed. The P values <0.05 were considered statistically significant. Performance was evaluated using AUC, sensitivity, specificity, and accuracy metrics. Result: The SVM model exhibited superior diagnostic performance with macro-average AUC values of 0.91 for CE-T1 alone, 0.86 for T2 alone, and 0.93 for combined CE-T1 and T2 sequences. And the combined sequence model demonstrated the best overall accuracy, sensitivity, and F1 score, with an accuracy of 0.77, outperforming both KNN and NBC models. Conclusion: The SVM-based MRI radiomics model effectively distinguishes between GB, PCNSL, and BM. Combining CE-T1 and T2 sequences significantly enhances classification performance, providing a robust, noninvasive diagnostic tool that could assist in treatment planning and improve patient outcomes.