Interpretable Brain Tumor MRI Classification: Comparing GLCM-Based Classical Models and CNNs

This work presents a direct, controlled comparison between handcrafted GLCM-based texture features (with Logistic Regression and Random Forest) and an end-to-end CNN for four-class brain tumor MRI classification, using identical data splits and preprocessing. The CNN achieves 92.00% accuracy and a 91.54% macro-F1 score, substantially outperforming both classical baselines. Grad-CAM is further applied to interpret CNN predictions, with specific focus on the glioma–meningioma confusion pair, offering qualitative insight into the model’s spatial decision-making.