Abstract

Glioma grading from medical imaging remains a complex pattern recognition task due to the intrinsic heterogeneity of brain tumors and the high dimensionality of radiomic features. In this study, we present a systematic benchmark of thirteen supervised classification models, including SVM, Random Forest, and XGBoost for automatic glioma grading based on radiomic patterns extracted from multimodal MRI. The dataset comprises 36 preoperative glioma cases (58.3% low-grade, 41.7% high-grade), with features derived from structural T1-weighted images and diffusion tensor imaging maps, including fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RD), and axial diffusivity (AD). For each modality, we applied a modality-specific feature selection and classification pipeline combining synthetic minority over-sampling technique for class imbalance correction, feature scaling, and stratified 5-fold cross-validation with grid search for hyperparameter optimization. The performance was evaluated using the accuracy score across different feature subset sizes. Results indicate that ensemble-based models, particularly those trained on T1, MD, and AD features, yield the highest classification performance. Nevertheless, T1-based model achieved similar performace with fewer features and a simpler architecture (K-Nearest Neighbors), suggesting a more stable and generalizable classification pipeline. To enhance model transparency, we implemented local interpretable model-agnostic explanations, identifying modalityspecific radiomic features that most influence predictions. This work presents a reproducible and interpretable pattern recognition framework for MRI-based glioma grading, demonstrating the value of combining multimodal imaging with explainable AI techniques.