A Unified and Interpretable Benchmarking Framework for Brain Tumor Classification Using Classical and Deep Learning Models.

Accurate detection of brain tumors from magnetic resonance imaging (MRI) remains a clinically critical yet computationally challenging task due to high-dimensional image complexity and sensitivity to diagnostic errors. This study presents a unified hybrid diagnostic framework that systematically integrates classical machine learning algorithms and deep neural architectures within a standardized and reproducible benchmarking environment. Unlike model-centric investigations that evaluate isolated classifiers, the proposed framework establishes a controlled cross-paradigm experimental ecosystem in which regression-based, probabilistic, distance-driven, shallow neural, and convolutional models operate under identical preprocessing, validation, and testing protocols. Beyond comparative performance analysis, interpretability is incorporated through Gradient-weighted Class Activation Mapping (Grad-CAM), enabling visualization of spatial attention patterns underlying convolutional predictions. A multi-metric evaluation strategy including accuracy, precision, recall, and F1-score provides comprehensive assessment of diagnostic reliability. Experimental results demonstrate a consistent performance hierarchy, with convolutional neural networks achieving superior discriminative capability and improved tumor sensitivity relative to classical approaches. By combining standardized benchmarking, interpretability integration, statistical validation, and deployment-aware evaluation, the proposed framework contributes a reproducible methodological reference for evidence-guided algorithm selection in medical imaging. The study advances transparent and clinically aligned artificial intelligence for MRI-based brain tumor detection.