Few-Shot Class-Incremental Learning for Retinal Disease Recognition.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models require continuous learning for evolving retinal diseases.
• Acquiring large annotated datasets for retinal disease diagnosis is challenging.
• Few-Shot Class-Incremental Learning (FSCIL) enables models to learn new classes with limited data while retaining prior knowledge.

Purpose of the Study:

• To introduce Re-FSCIL, a novel framework for Few-Shot Class-Incremental Retinal Disease Recognition (FSCIRDR).
• To address the limitations of current deep learning models in diagnosing evolving retinal diseases with limited samples.
• To establish new benchmarks for FSCIRDR research.

Main Methods:

• Integration of the RETFound model with a fine-grained module.
• Implementation of a forward-compatible training strategy for improved adaptability.
• Application of supervised contrastive learning for enhanced feature discrimination and feature fusion for robust representation.

Main Results:

• Re-FSCIL achieves State-of-the-art (SOTA) performance on the newly created FSCIRDR benchmarks (RFMiD38 and JSIEC39).
• The proposed framework significantly surpasses existing FSCIL methods.
• Demonstrated effectiveness in Few-Shot Class-Incremental Retinal Disease Recognition.

Conclusions:

• Re-FSCIL provides a robust and adaptable solution for Few-Shot Class-Incremental Retinal Disease Recognition.
• The framework shows significant potential for improving DL-based diagnostic systems for retinal diseases.
• The established benchmarks will facilitate future research in FSCIRDR.