A generalizable eye disease detection method based on Zero-Shot Learning.

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Artificial Intelligence

Background:

• Deep learning in medical imaging is hindered by the need for large, expert-annotated datasets, especially in ophthalmology for early disease detection like mild Diabetic Retinopathy (DR1).
• Scarcity of annotations for subtle DR1 lesions limits the effectiveness of traditional supervised learning methods.

Purpose of the Study:

• To develop a generalizable eye disease detection framework using Zero-shot Learning (ZSL) that simulates clinical reasoning.
• To enable unsupervised detection of DR1 without requiring any labeled DR1 cases.

Main Methods:

• Utilized the LCFP-14M dataset, a novel large-scale fundus image resource.
• Employed a Siamese network to identify disease correlations and transfer knowledge by segmenting DR1-specific lesions from a correlated source disease.
• Implemented a ResNet-Agglomerative clustering pipeline for unsupervised DR1 detection.

Main Results:

• The proposed ZSL framework achieved effective DR1 detection without annotated DR1 data.
• Performance metrics include accuracy (0.8337), precision (0.8700), recall (0.7456), F1 score (0.8030), and ROC-AUC (0.9226).
• The framework outperformed most supervised baselines on external test datasets.

Conclusions:

• Zero-shot Learning (ZSL) can effectively simulate clinical diagnostic logic for eye diseases.
• The approach demonstrates generalization capabilities for detecting unseen eye diseases.
• This offers a promising avenue for automated screening in scenarios with limited labeled data.