Contrastive learning-based pretraining improves representation and transferability of diabetic retinopathy...

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss globally.
• Early diagnosis of DR is crucial for effective patient management.
• Existing machine learning models often require large datasets for robust performance.

Purpose of the Study:

• To develop a robust deep learning model for diabetic retinopathy classification using self-supervised contrastive learning.
• To enhance model generalizability and performance with smaller labeled datasets.
• To reduce the annotation burden on clinicians through improved data efficiency.

Main Methods:

• Developed a self-supervised contrastive learning (CL) pipeline for DR classification.
• Integrated neural style transfer (NST) augmentation into the CL pipeline.
• Trained and validated the model on the EyePACS dataset and tested on independent clinical datasets (UIC).
• Compared performance against baseline models, including evaluations with reduced labeled training data (10%).

Main Results:

• The CL-pretrained FundusNet model achieved higher AUC (0.91) on independent clinical data compared to baseline models (0.80, 0.83).
• Even with 10% labeled data, the FundusNet model maintained superior performance (AUC 0.81) versus baselines (0.58, 0.63).
• The model demonstrated strong generalizability, performing well on data from a different source (EyePACS to UIC).

Conclusions:

• Self-supervised contrastive learning with NST significantly improves deep learning classification for diabetic retinopathy.
• The proposed method enhances model generalizability and enables effective training with small, annotated datasets.
• This approach offers a promising solution for accurate and efficient DR detection, reducing clinical workload.