Few-shot learning for inference in medical imaging with subspace feature representations.

Area of Science:

• Medical image analysis
• Machine learning
• Computer vision

Background:

• Deep learning advances in visual recognition contrast with challenges in medical imaging due to limited data.
• Few-shot learning, using pre-trained models as feature extractors, is a strategy for small medical image datasets.
• High complexity and variability in medical images hinder few-shot learning's ability to capture essential features.

Purpose of the Study:

• To address the limitations of few-shot learning in medical image analysis by exploring alternative dimensionality reduction techniques.
• To investigate the effectiveness of discriminant analysis (DA) and non-negative matrix factorization (NMF) as alternatives to singular value decomposition (SVD).

Main Methods:

• Explored dimensionality reduction techniques, specifically discriminant analysis (DA) and non-negative matrix factorization (NMF), as alternatives to principal component analysis/singular value decomposition (PCA/SVD).
• Evaluated methods on 14 diverse medical image datasets covering 11 distinct disease types.
• Focused on feature spaces where dimensionality is comparable to or exceeds the number of available images.

Main Results:

• Discriminant subspaces demonstrated significant improvements over SVD-based subspaces and the original feature space at low dimensions.
• Non-negative matrix factorization (NMF) proved to be a competitive alternative to SVD in modest dimensional settings.
• The proposed methods enhance feature extraction for few-shot learning in complex medical imaging scenarios.

Conclusions:

• Discriminant analysis and NMF are effective alternatives to SVD for dimensionality reduction in few-shot medical image learning.
• These methods improve the performance of classic pattern recognition techniques in high-dimensional, low-data regimes.
• The study provides a practical implementation for advancing medical image analysis with limited datasets.