Rotation-Oriented Collaborative Self-Supervised Learning for Retinal Disease Diagnosis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automatic diagnosis of ophthalmic diseases from fundus images is crucial but hindered by large data and annotation costs.
• Developing effective automatic diagnostic solutions requires robust feature learning from medical images.

Purpose of the Study:

• To present a novel self-supervised learning framework for retinal disease diagnosis, minimizing annotation efforts.
• To leverage unlabeled fundus images for learning discriminative visual features.

Main Methods:

• A rotation-oriented collaborative method was developed to explore both rotation-related and rotation-invariant features.
• The framework learns visual features from unlabeled fundus images for disease classification.

Main Results:

• The proposed method achieved superior performance compared to other self-supervised feature learning techniques, showing a 4.2% increase in Area Under the Curve (AUC).
• The method surpassed the supervised baseline for pathologic myopia (PM) and neared the supervised baseline for age-related macular degeneration (AMD) with ample unlabeled data.

Conclusions:

• The novel self-supervised learning framework effectively reduces annotation requirements for retinal disease diagnosis.
• The method demonstrates significant potential for clinical application by improving diagnostic accuracy using unlabeled fundus images.