EAMF: An Entropy-enhanced Attention-based Ensemble Metric Few-Shot Learning for MRI Image Classification

In MRI image classification, obtaining large-scale labeled datasets can be challenging, particularly for rare diseases or when patient data access is restricted due to privacy concerns. Few-Shot Learning (FSL) aims to classify target classes with only a few available training examples. Metric-based FSL (MFSL) is a widely used approach that focuses on learning a generalizable class-level metric to distinguish between different classes. However, the deep feature embeddings generated by state-of-the-art MFSL networks often do not achieve intra-class and inter-class variations effectively, thus leading to poor class discrimination. To address this limitation, we propose a novel approach, called Entropy-enhanced Attention-based Ensemble Metric FSL (EAMF). In this method, we compute patch-wise image entropy to generate an additional entropy-feature vector, which is then concatenated with the embedding vector of each backbone network in the ensemble. We also introduce an attention mechanism to assign different weights to the embeddings based on their class representativeness, allowing embeddings to be combined in a weighted ensemble. In addition, we evaluate three different distance metrics to determine the most effective one for decision-making. Experimental results on two MRI datasets demonstrate the superiority of EAMF when compared to standalone deep learning models.