Brain tumor segmentation is critical in medical imaging but remains a challenging task due to the limited availability of annotated data. While Deep Neural Networks (DNNs) have demonstrated promising results, they typically require extensive datasets and often struggle to generalize across unseen tumor types. To address these challenges, we propose a novel few-shot learning framework for brain tumor segmentation in Magnetic Resonance Imaging (MRI) using prototype similarity scores. Unlike conventional methods that process entire image sets, our approach focuses on specific slices containing tumor regions, significantly reducing training data requirements. The model employs an iterative training strategy, selecting random tumor slices and pairing them with other slices from the same scan to form a support set. By utilizing metric learning with non-parametric thresholds, the model effectively distinguishes query images from class prototypes. Evaluation on the BraTS 2021 dataset, comprising 360 training and 100 testing MRI scans, highlights the framework’s efficiency. The model achieves an F1-Score of 78.96% and a Mean Intersection over Union (mIoU) of 80.97% in a 5-shot 2-way configuration. In a 10-shot 2-way setup, the performance improves, reaching an F1-Score of 81.75% and an mIoU of 81.59%, demonstrating its ability to segment tumor regions more accurately with additional data. These results emphasize the effectiveness of the proposed few-shot learning framework for accurate and efficient brain tumor segmentation in settings with limited data.