Prototype-guided semantic transferability for reliable knowledge distillation in brain tumor MRI classification

Abstract Brain tumor classification from magnetic resonance imaging (MRI) remains challenging due to subtle inter-class differences, imaging variability, and uncertainty in model predictions. Although deep learning models achieve strong performance, their complexity limits deployment in resource-constrained settings, making knowledge distillation a practical approach for transferring knowledge to lightweight student models. However, conventional distillation methods typically assume uniform teacher reliability and often overlook semantic structure and prediction reliability. To address this limitation, we propose a prototype-guided semantic transferability framework that selectively transfers knowledge based on predictive confidence and alignment with class prototypes in an embedding space. The method introduces a prototype memory to capture class-level structure and formulates a transferability score to adaptively weight the distillation process. Experiments demonstrate consistent improvements in classification performance, calibration, robustness, and cross-dataset generalization while preserving the efficiency of the student model. Ablation studies, statistical analysis, and feature space visualizations further support the effectiveness of the proposed components. Overall, the proposed framework aims to enhance knowledge distillation by incorporating semantic structure and uncertainty, leading to more reliable and discriminative representations without additional inference overhead.