ViResGF-Net: Gated Multi-Scale Hybrid Vision Transformer for Robust Fundus Image Multi-Label Classification

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Global population aging increases the prevalence of visual impairment due to fundus diseases like cataracts and glaucoma.
• Traditional ophthalmic diagnosis relies on subjective visual inspection of fundus images by doctors, leading to diagnostic discrepancies and challenges in identifying coexisting diseases.

Purpose of the Study:

• To develop an automated, accurate, and robust model for multi-class classification of fundus diseases.
• To address the limitations of manual diagnosis, including inter-observer variability and specialization constraints.

Main Methods:

• Proposed ViResGF-Net, a gated multi-scale hybrid vision Transformer model integrating Convolutional Neural Network (CNN) and Vision Transformer (ViT) branches.
• Employed a Feature Pyramid Network (FPN) to enhance local feature extraction within the CNN branch.
• Integrated features from both branches using a Gated Fusion Unit (GFU) before classification with a Multi-Layer Perceptron (MLP) classifier.

Main Results:

• ViResGF-Net achieved high performance metrics: 93.56% accuracy, 92.99% precision, and 92.36% F1 score.
• The model's performance surpassed existing methods in multi-class fundus disease classification.

Conclusions:

• The ViResGF-Net model demonstrates significant potential for improving the accuracy and efficiency of fundus disease diagnosis.
• This hybrid deep learning approach offers a promising solution for automated ophthalmic image analysis, particularly in complex cases involving multiple pathologies.