CLRD: Collaborative Learning for Retinopathy Detection Using Fundus Images.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Retinopathy is a leading cause of vision loss, necessitating early detection through fundus imaging.
• Limited fundus image availability and imbalanced datasets hinder the development of effective diagnostic algorithms.
• Advanced algorithms are crucial for improving the precision and efficiency of retinopathy diagnosis.

Purpose of the Study:

• To present a novel online knowledge distillation framework (CLRD) for enhanced retinopathy detection.
• To improve diagnostic performance by addressing challenges of limited data and imbalanced datasets in fundus imaging.
• To develop a collaborative learning approach combining diverse AI models for robust retinopathy identification.

Main Methods:

• Developed a collaborative online knowledge distillation (CLRD) framework for retinopathy detection.
• Integrated student models with varied scales and architectures, including Transformer-based BEiT and CNN-based ConvNeXt.
• Implemented fundus image-specific distortion information transfer to enhance model invariance.

Main Results:

• The CLRD framework achieved high diagnostic accuracies: BEiT (98.77%) and ConvNeXt (96.88%).
• CLRD demonstrated significant improvements over advanced visual models, with higher accuracy, precision, recall, specificity, and F1 scores.
• The framework effectively minimized generalization errors while preserving individual student model predictions.

Conclusions:

• The CLRD framework offers a promising approach for accurate and efficient retinopathy detection using fundus images.
• Collaborative learning and knowledge distillation can overcome data limitations in medical image analysis.
• This study provides a foundation for further research in AI-driven diagnostic tools for visual impairments.