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

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Large-scale labeled datasets are crucial but challenging to obtain for MRI image classification, especially for rare diseases or due to privacy concerns.
• Few-Shot Learning (FSL) addresses data scarcity by enabling classification with minimal training examples.
• Metric-based FSL (MFSL) relies on learning discriminative metrics but struggles with effective intra-class and inter-class feature variations.

Purpose of the Study:

• To propose a novel Few-Shot Learning (FSL) method, Entropy-enhanced Attention-based Ensemble Metric FSL (EAMF), to enhance class discrimination in MRI image classification.
• To improve the effectiveness of deep feature embeddings in distinguishing between classes despite limited data.
• To evaluate the performance of the proposed EAMF method against existing deep learning models.

Main Methods:

• Developed EAMF, incorporating patch-wise image entropy to create an additional entropy-feature vector.
• Concatenated entropy-feature vectors with backbone network embeddings for an ensemble approach.
• Implemented an attention mechanism to weight embeddings based on class representativeness and evaluated three distance metrics.

Main Results:

• The proposed EAMF method demonstrated superior performance in MRI image classification compared to standalone deep learning models.
• The integration of entropy-features and the attention-based ensemble significantly improved class discrimination.
• Experimental results on two MRI datasets validated the effectiveness of the EAMF approach.

Conclusions:

• EAMF effectively addresses the limitations of traditional MFSL by enhancing feature embeddings for better class discrimination.
• The novel approach shows significant promise for improving MRI image classification in data-scarce scenarios.
• EAMF offers a robust solution for medical image analysis where data availability is a constraint.