A Multi-Resolution Hypergraph Transformer for Explainable Retinal Disease Prediction.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of retinal diseases is crucial for preventing vision loss.
• Current diagnostic models often struggle with classifying multiple co-existing retinal diseases.
• Interrelated features of multiple pathologies present a significant challenge in retinal image analysis.

Purpose of the Study:

• To develop a novel framework for multi-label retinal disease classification and lesion characterization.
• To address the limitations of traditional single-disease diagnostic models.
• To improve the accuracy and interpretability of automated retinal disease diagnosis.

Main Methods:

• Proposed a Multi-Resolution Hypergraph Vision Transformer (MR-HGViT) framework.
• Constructed multi-resolution hypergraphs to integrate global structures and local lesion details.
• Employed Dynamic Hypergraph Convolutional Networks (DHGCNs) and Vision Transformers for feature propagation and dependency capture.
• Utilized attention maps and Grad-CAM for model interpretability.

Main Results:

• Achieved state-of-the-art accuracies on three benchmark datasets: IDRiD (94.37%), REFUGE (94.12%), and MuReD (93.78%).
• Demonstrated superior performance in multi-label classification of retinal diseases.
• Provided clinically relevant insights through enhanced model interpretability.

Conclusions:

• The MR-HGViT framework is highly effective for multi-label retinal disease diagnosis and lesion characterization.
• The model offers significant potential for clinical applications in ophthalmology.
• MR-HGViT advances automated diagnostic tools by combining accuracy with interpretability.