OCT-SelfNet: a self-supervised framework with multi-source datasets for generalized retinal disease detection.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Limited labeled medical image data due to privacy concerns hinders AI model generalization in diverse patient populations.
• A gap exists between AI advancements and their practical application in medical settings, particularly for localized models.

Purpose of the Study:

• To develop a self-supervised machine learning framework, OCT-SelfNet, for detecting eye diseases from optical coherence tomography (OCT) images.
• To achieve generalized learning in AI for ophthalmology while minimizing the need for extensive labeled datasets.

Main Methods:

• Implemented a two-phase training strategy: self-supervised pre-training on unlabeled data, followed by supervised training.
• Utilized a masked autoencoder with a SwinV2 backbone for robust feature extraction.
• Integrated diverse datasets from multiple sources to enhance disease representation.

Main Results:

• OCT-SelfNet demonstrated superior performance compared to baseline models (ResNet-50, ViT) across various experimental conditions.
• The framework exhibited strong cross-dataset generalization capabilities, outperforming baseline models.
• Ablation studies confirmed significant performance gains from self-supervised pre-training and data fusion.

Conclusions:

• The OCT-SelfNet framework shows significant promise for clinical deployment in eye disease detection using OCT images.
• The proposed two-phase training approach and masked autoencoder architecture effectively enhance domain adaptation and generalization.
• This work bridges the gap between AI research and clinical application in ophthalmology.