A difficulty-aware and task-augmentation method based on meta-learning model for few-shot diabetic retinopathy...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate diabetic retinopathy (DR) classification is crucial for early diagnosis and treatment.
• Limited annotated DR data challenges current deep learning models.
• Few-shot learning approaches are needed for DR classification.

Purpose of the Study:

• To propose a difficulty-aware and task-augmentation method based on meta-learning (DaTa-ML) for few-shot DR classification.
• To address the challenge of limited annotated DR data in fundus images.
• To improve the efficiency and accuracy of DR classification models.

Main Methods:

• Implemented a difficulty-aware (Da) method to dynamically adjust cross-entropy loss, prioritizing challenging tasks.
• Utilized a task-augmentation (Ta) method to increase meta-training tasks via image rotation, enhancing feature extraction.
• Optimized meta-training task sampling and initialization parameters for improved meta-generalization.

Main Results:

• The DaTa-ML model achieved 89.6% accuracy on the APTOS 2019 dataset using only 1% of training data (5-way, 20-shot) and a single update step.
• Demonstrated a 1.7% improvement over transfer learning (ResNet50 pre-trained on ImageNet) and a 16.8% improvement over scratch-built models.
• Achieved these results with only 0.47% of ResNet50 parameters, indicating high efficiency.

Conclusions:

• The DaTa-ML model offers an efficient solution for DR classification with minimal annotated data.
• The model shows significant advantages over state-of-the-art methods in few-shot learning scenarios.
• DaTa-ML can assist ophthalmologists in determining DR severity, improving patient care.