Diabetic retinopathy identification based on multi-source-free domain adaptation.

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy (DR) identification relies heavily on deep learning, which requires extensive labeled data.
• Challenges include data labeling difficulties, privacy concerns, and the need for large datasets.
• Existing methods struggle with efficient DR identification from unlabeled data.

Purpose of the Study:

• To develop a source-free domain adaptation (SFDA) method for efficient and effective DR identification using unlabeled data.
• To address limitations of deep learning in DR identification, such as data labeling, privacy, and data volume requirements.

Main Methods:

• A multi-SFDA method was proposed, integrating multiple source models to generate synthetic pseudo-labels for unlabeled target data.
• A softmax-consistency minimization term was employed to reduce intra-class distances and increase inter-class distances between source and target domains.
• Validation was conducted on three public fundus photograph datasets: APTOS2019, DDR, and EyePACS.

Main Results:

• The proposed multi-SFDA model achieved promising F1-scores of 0.8917 for referable DR and 0.9795 for normal/abnormal DR identification.
• The method demonstrated effective DR identification by successfully minimizing intra-class and maximizing inter-class distances between domains.
• Results indicate the model's capability in handling variations across different datasets.

Conclusions:

• The multi-SFDA method offers a practical solution for DR identification, overcoming data labeling and privacy hurdles.
• This approach significantly reduces the dependency on large labeled datasets for deep learning in DR detection.
• The method holds potential for early DR detection and vision preservation in diabetic patients.