Multi-Label Fundus Image Classification Using Attention Mechanisms and Feature Fusion.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Fundus diseases pose a significant risk of irreversible vision loss.
• Accurate diagnosis is crucial for timely treatment.
• Existing deep learning models struggle with multi-labeled fundus images, showing low diagnostic accuracy.

Purpose of the Study:

• To develop a robust deep learning model for accurate multi-label classification of fundus diseases using binocular images.
• To enhance diagnostic accuracy in cases with multiple concurrent fundus pathologies.

Main Methods:

• A novel neural network algorithm incorporating attention mechanisms and feature fusion was developed.
• The model utilizes ResNet50 with attention to extract detailed lesion features from binocular fundus images.
• Feature fusion integrates global features, followed by Softmax for multi-label classification.

Main Results:

• The model achieved high performance on the ODIR dataset.
• Key metrics included accuracy (94.23%), precision (99.09%), recall (99.23%), and F1-score (99.16%).
• Ablation experiments validated the effectiveness of the proposed architecture.

Conclusions:

• The developed attention-based, feature-fusion model demonstrates superior performance in multi-label fundus image classification.
• This approach offers a promising tool for improving the diagnosis of complex fundus diseases.
• The findings contribute to advancing AI applications in ophthalmology for better vision preservation.