Combining convolutional neural networks and self-attention for fundus diseases identification

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of fundus diseases is crucial for effective treatment.
• Fundus photography is a key screening tool, but classifying multiple diseases simultaneously remains challenging.
• Current deep learning models, primarily CNNs, have limitations in global modeling for complex fundus image analysis.

Purpose of the Study:

• To develop an advanced deep learning model for the simultaneous classification of multiple fundus diseases.
• To improve the accuracy and global modeling capabilities in fundus image analysis.
• To introduce a novel architecture, MBSaNet, integrating CNNs and self-attention mechanisms.

Main Methods:

• Proposed a multistage fundus image classification model (MBSaNet) combining Convolutional Neural Network (CNN) and Self-Attention (SA) mechanisms.
• Developed a multiscale feature fusion stem for enhanced low-level feature extraction.
• Utilized the ODIR-5k dataset for training and testing the model's performance.

Main Results:

• MBSaNet achieved state-of-the-art performance on the ODIR-5k dataset.
• The model demonstrated superior accuracy with fewer parameters compared to existing methods.
• The proposed architecture effectively captured both local and global features for robust classification.

Conclusions:

• MBSaNet offers a significant advancement in the automated diagnosis of multiple fundus diseases.
• The model's effectiveness was validated across diverse diseases and image acquisition conditions.
• MBSaNet shows strong applicability for real-world clinical screening and diagnosis support.