Toward Noise-Resilient Few-Shot Learning: Optimizing Prototypes for Glioblastoma Classification

Few-shot learning (FSL) with prototype-based methods offers an appealing solution for medical image classification, particularly for rare conditions with limited data. However, traditional prototypical networks struggle with data heterogeneity and bias, often focusing on irrelevant patterns rather than true discriminative features. We propose a novel metric-based optimization algorithm that enhances prototype selection in few-shot learning, evaluated on the challenging case of Glioblastoma (GBM) classification. Our method addresses the inherent heterogeneity of medical imaging data by intelligently selecting support sets that capture genuine pathological features while mitigating orientation and imaging technique biases. Experimental results demonstrate that our approach consistently outperforms baseline methods across varying noise conditions, achieving average improvements of 5.89% and 4.66% in classification accuracy over random sampling when applied to prototypical networks and IMP, respectively. The algorithm maintains computational efficiency while significantly enhancing robustness to data bias and noise, making it particularly valuable for medical imaging applications where data quality and representation vary substantially.

Few-Shot Learning (FSL), inspired by the adaptability of human cognition, is a widely used methodology for scenarios with limited data samples. Prototype-based methods are a classical and popular subset of FSL techniques. However, these methods can be particularly vulnerable to the selection of samples used to form prototypes, especially when unintentional noise is introduced.

To address this limitation, we present a framework designed to optimize the sampling process, making it more robust to noise. We experiment on classification of Glioblastoma, an aggressive and heterogeneous brain tumor. We model the unexpected variability introduced by different orientations, imaging techniques, resolutions, and other factors.

Our method consistently enhances the performance of both the original Prototypical Network and the Infinite Prototypical Network, demonstrating improved resilience and reliability across diverse conditions.

To motivate our approach, consider the example in the figure below. Traditional Prototypical Networks are sensitive to noisy and heterogeneous data, such as variations in collected brain orientations (e.g., coronal, sagittal, and axial orientations). This often leads to biased prototypes that fail to capture the true distinguishing features of classes, resulting in misclassification.

To address these challenges, we propose a Noise-Resilient Sampling Framework. Our method optimizes prototype formation through an iterative process that selects representative support samples, reduces noise impact, and improves resilience to data heterogeneity. Key steps include support set refinement, prototype scoring with a validation set, and iterative updates to enhance classification performance.

By iteratively refining the support set and using validation scores, our framework ensures prototypes are robust to noise and bias. This process enhances the performance of both Prototypical Networks and Infinite Prototypical Networks (IMP), consistently improving classification accuracy across heterogeneous datasets.