Abstract
BACKGROUND AND AIMS: Accurate classification of brain tumors is vital for effective treatment planning. Manual assessment of magnetic resonance imaging (MRI) scans is often subjective and time-consuming. This exploratory study proposes a machine learning approach integrating radiomic features from contrast-enhanced T1-weighted MRI scans to improve classification accuracy for glioblastomas (GBM), meningiomas, brain metastases, and normal controls (NC). METHODS: A total of 154 patients (74 GBM, 62 meningioma, 18 brain metastases) and 170 NCs were included in the study. MRI scans underwent standard preprocessing. To address class imbalance, data augmentation was applied to the GBM, meningioma, and brain metastases groups, increasing their sample sizes to 148, 124, and 180 cases, respectively. Radiomic features were extracted using gray-level co-occurrence matrix (GLCM) analysis and Curvelet-based transformations. Feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO), followed by multicollinearity assessment using the Spearman correlation test. Principal component analysis (PCA) was also performed. Nine machine learning classifiers, including Random Forest (RF), K-Nearest Neighbors (KNN), Categorical Boosting (CatBoost), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Logistic Regression (LR), Naive Bayes, AdaBoost, and LightGBM, were trained and evaluated. RESULTS: Random Forest and CatBoost achieved the highest performance, with 95.2% accuracy using LASSO-selected features, outperforming other classifiers. Combining GLCM and Curvelet features improved accuracy over either set alone. PCA-based dimensionality reduction (99% variance) also performed well, but slightly below LASSO. CONCLUSION: Integrating Curvelet and GLCM-based radiomics with optimized models improves diagnostic accuracy, offering a promising tool to assist radiologists in tumor classification.