Cross-attention multi-branch network for fundus diseases classification using SLO images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate fundus disease classification is crucial for human health.
• Current methods using single-angle fundus images lack comprehensive pathological information.

Purpose of the Study:

• To develop a novel deep learning method for fundus disease classification.
• To leverage ultra-wide field scanning laser ophthalmoscopy (SLO) images for enhanced pathological detection.

Main Methods:

• A multi-branch deep learning model incorporating ResNet-34 backbone.
• Utilized atrous spatial pyramid pooling (ASPP) for multi-scale feature extraction.
• Integrated depth-wise and cross-attention modules for enhanced feature representation.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art techniques.
• Achieved promising classification results on collected and public SLO image datasets.
• Effectively fused multi-scale and attention-based features for improved accuracy.

Conclusions:

• The novel deep learning model effectively classifies fundus diseases using ultra-wide field SLO images.
• The integration of advanced deep learning modules enhances the detection of subtle pathological features.
• This approach offers a significant advancement in automated fundus disease diagnosis.