Enhanced Diabetic Retinopathy Detection: An Explainable Semi-Supervised Approach Using Contrastive Learning.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a major cause of blindness globally, necessitating early detection for effective treatment.
• Automated medical image analysis is crucial for timely DR diagnosis, but faces challenges with limited labeled, imbalanced, and unlabeled data.
• Existing semi-supervised learning methods for DR detection struggle with unreliable pseudo-labeling, data exclusion, and dataset biases.

Purpose of the Study:

• To develop a novel semi-supervised learning framework for enhanced diabetic retinopathy detection.
• To address limitations of current methods by effectively utilizing unlabeled and low-confidence data.
• To improve the transparency and interpretability of DR diagnostic models.

Main Methods:

• A novel semi-supervised learning framework combining similarity and contrastive learning for DR detection.
• Utilizing class prototypes and an ensemble of classifiers for reliable pseudo-label generation.
• Integrating unreliable samples via contrastive learning and employing GradCAM for model interpretability.

Main Results:

• The proposed framework demonstrated improved performance over existing semi-supervised learning methods on the Kaggle DR dataset.
• Effectively leveraged unreliable samples, enhancing feature extraction and overall model performance.
• Provided enhanced transparency and interpretability through GradCAM integration.

Conclusions:

• The novel semi-supervised learning framework offers a promising solution for advancing diabetic retinopathy diagnosis.
• The method's ability to utilize unreliable data signifies a potential breakthrough in automated DR detection.
• The integration of explainable AI enhances clinical trust and adoption of AI-driven diagnostic tools.