Abstract
Brain tumors, caused by rapid and uncontrolled cellular proliferation, pose significant health risks worldwide. Accurate and early classification of tumor types such as Glioblastoma, IDH-wildtype, grade 4, Meningioma, and Brain Metastasis is crucial for improving treatment outcomes. This study combines Radiomic features with advanced deep learning models, including EfficientNet-B0 and ResNet-18, to classify 2D and 3D brain tumor images effectively. The classifiers, particularly MLP and CatBoost, demonstrated robust performance, achieving macro F1-scores of 86% and 84%, respectively. However, the imbalanced dataset posed challenges in classification accuracy for less represented classes, emphasizing the need for comprehensive evaluation using AUC and ROC analyses. Additionally, SHAP analysis revealed that texture heterogeneity features, especially original_gldm_DependenceNonUniformityNormalized, consistently contributed significantly across tumor types. Model uncertainty assessment showed that brain metastasis was the most confidently predicted class in both models, meningioma classification had higher confidence in MLP, and Glioblastoma, IDH-wildtype was predicted more reliably by CatBoost. These findings highlight the complementary strengths of combining Radiomics and deep learning for enhanced brain tumor classification and underscore the value of uncertainty quantification in model evaluation.