Multi-Learner Based Deep Meta-Learning for Few-Shot Medical Image Classification.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Few-shot learning (FSL) shows promise in medical image analysis due to the high cost of creating large, high-quality datasets.
• Existing FSL methods, primarily developed for natural images, require further adaptation and evaluation for medical imaging applications.
• Meta-learning offers a robust framework to address the challenges inherent in few-shot settings.

Purpose of the Study:

• To develop a novel multi-learner based few-shot learning (FSL) method tailored for multiple medical image classification tasks.
• To integrate meta-learning with transfer-learning and metric-learning to enhance model adaptability for unseen medical imaging tasks.
• To improve the efficiency and generalization capabilities of FSL models in medical scenarios.

Main Methods:

• A multi-learner model comprising an auto-encoder, metric-learner, and task-learner was designed.
• Transfer-learning was employed to train all learners on base classes, followed by meta-learning to fine-tune learners on novel tasks.
• Real-time data augmentation and a dynamic Gaussian disturbance soft label (GDSL) scheme were implemented for enhanced generalization.

Main Results:

• The proposed method was evaluated on three challenging, newly-built medical benchmarks: BLOOD, PATH, and CHEST, for three-class few-shot classification.
• The model demonstrated superior performance on both homogeneous and cross-domain medical datasets compared to existing related works.
• The combined approach of meta-learning, transfer-learning, and metric-learning proved effective for few-shot medical image classification.

Conclusions:

• The developed multi-learner FSL method effectively addresses the challenges of limited data in medical image analysis.
• The integration of meta-learning, transfer-learning, and metric-learning strategies significantly improves model performance and adaptability.
• The proposed method offers a promising solution for advancing few-shot learning applications in the medical domain.