Hybrid ViT-Prototypical Reptile Learning for Few-Shot Brain Tumor MRI Classification

Accurate brain tumor classification from magnetic resonance imaging (MRI) is vital for clinical diagnosis, yet the scarcity of labeled data often hinders the deployment of deep learning models. To address this challenge, we propose the hybrid Vision Transformer (ViT)-Reptile framework, which integrates a partially fine-tuned ViT-B/16 for feature extraction with a Prototypical Network (ProtoNet) for metric-based classification, enhanced by the Reptile meta-learning algorithm for rapid parameter adaptation at test time. We evaluated the framework on a public brain tumor MRI dataset containing 3,264 T1-weighted images across four categories (glioma, meningioma, pituitary tumor, and healthy brain) under a 4-way K-shot episodic protocol with K ∈ {1, 3, 5}. Our model consistently outperforms baselines including ResNet-50, DenseNet-121, and a non-adaptive ViT-based ProtoNet. Specifically, at K = 5, the framework reaches 87.43 ± 0.28% accuracy with a macro F1-score of 0.8700 ± 0.0029. Performance confirms that Reptile’s test-time refinement leads to cumulative gains, especially in the challenging 1-shot setting. Visualization of the embedding space further reveals that the model produces distinct feature clusters for all four tumor categories, with separability improving as K increases. These findings suggest that combining ViT with first-order meta-learning is an effective and efficient way to handle data scarcity in medical imaging.