Adaptive Gating and Focal Debiasing for robust few-shot retinal disease classification

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Few-shot learning models for retinal disease classification face challenges with static feature fusion and poor performance on minority classes.
• Existing hybrid attention models struggle to adaptively integrate global and local visual information, limiting their effectiveness in data-scarce scenarios.

Purpose of the Study:

• To introduce the Adaptive Gating and Focal Debiasing (AGFD) method to improve automated retinal disease classification in few-shot learning settings.
• To address limitations of static feature fusion and enhance performance on underrepresented classes in retinal image datasets.

Main Methods:

• Developed a Dynamic Attention Gating (DAG) module within the AGFD framework to learn input-specific weights for global and local attention branches.
• Replaced the standard cross-entropy loss function with focal loss to emphasize learning from hard-to-classify, underrepresented cases.
• Evaluated the AGFD model on the ODIR-5K dataset using episodic evaluation across various few-shot settings (e.g., 5-shot, 10-shot).

Main Results:

• AGFD demonstrated significant improvements over a strong hybrid-attention baseline, achieving 78.7% accuracy and 76.9% macro-F1 at 5-shot, and 83.2% accuracy at 10-shot.
• Minority classes, including Glaucoma, Cataract, and Hypertension, experienced substantial gains, with F1 scores increasing by 11-15 percentage points.
• Analysis of the gating mechanism revealed adaptive weighting: global features were prioritized for widespread conditions, while local features were favored for lesion-driven diseases.

Conclusions:

• The combination of adaptive feature fusion and a debiased objective function (AGFD) enhances both overall accuracy and reliability for underrepresented classes in retinal disease classification.
• The proposed method represents a significant step towards developing more clinically useful automated screening tools for retinal conditions, particularly in resource-limited settings.
• Dynamic attention gating effectively balances global and local feature importance based on disease characteristics, leading to more robust classification performance.