An Adaptive Hybrid Feature Fusion Framework for Interpretable Brain Tumor Classification in MRI

Abstract

Accurate brain tumor classification from magnetic resonance imaging requires models capable of capturing both structural complexity and subtle radiomic heterogeneity within lesion regions. Handcrafted radiomic approaches offer interpretability but lack hierarchical spatial abstraction, whereas deep convolutional networks, despite strong predictive capability, may overlook complementary statistical descriptors. This study proposes an Adaptive Hybrid Feature Fusion framework that integrates convolutional embeddings and handcrafted radiomic features through an instance-aware dynamic weighting mechanism. Radiomic descriptors derived from Gray-Level Co-occurrence Matrix statistics and histogram-based measures are combined with deep convolutional representations via a learned adaptive coefficient that regulates feature dominance prior to classification. The framework is evaluated on a publicly available multi-class MRI dataset using a standardized validation protocol. Experimental results demonstrate that adaptive fusion achieves 91.42% classification accuracy, outperforming standalone convolutional models by 2.85%, with statistically significant improvement (p = 0.012) and reduced cross-validation variance. Interpretability is incorporated through Grad-CAM to verify spatial alignment between predictive attention and tumor regions. The proposed approach advances hybrid medical image analysis by introducing statistically validated, stability-oriented, and deployment-aware adaptive feature-level integration, providing a reproducible foundation for interpretable and robust neuro-diagnostic intelligence.