Test-Time Augmentation-Enhanced Knowledge Distillation for Brain Tumor MRI Classification

Abstract

Deep learning models for brain tumor MRI classification often face a trade-off between accuracy and efficiency. In this paper, we propose a Test-Time Augmentation (TTA)-enhanced knowledge distillation framework to improve the stability of teacher predictions. By aggregating multiple augmented views, TTA reduces prediction variance and provides more reliable supervision signals for distillation. The proposed approach transfers knowledge from a CNN ensemble to lightweight Hybrid Vision Transformers, achieving up to 45 × model compression while maintaining competitive performance. These results demonstrate the effectiveness of TTA-enhanced distillation for efficient medical imaging applications.