CA-ViT: Contour-Guided and Augmented Vision Transformers to Enhance Glaucoma Classification Using Fundus Images

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness globally, often detected late due to asymptomatic onset.
• Current deep learning methods for glaucoma classification face challenges with limited data, feature variations, and image quality.
• Vision transformers (ViTs) show promise but struggle with subtle differences in fundus images.

Purpose of the Study:

• To develop an advanced deep learning model for enhanced glaucoma classification using fundus images.
• To address limitations in existing methods, including data scarcity and feature variability.
• To improve early detection rates for glaucoma, preventing vision impairment.

Main Methods:

• Introduced the contour-guided and augmented vision transformer (CA-ViT) model.
• Utilized a Conditional Variational Generative Adversarial Network (CVGAN) for data augmentation and enhancement.
• Integrated a contour-guided approach focusing on optic disc and cup regions for feature extraction.
• Employed feature alignment with weighted cross-entropy loss for the ViT backbone.

Main Results:

• The CA-ViT model achieved high performance metrics: 93.0% precision, 93.08% recall, 92.9% F1 score, and 93.0% accuracy.
• Demonstrated significant improvement over existing glaucoma classification methods.
• Validated on the Standardized Multi-Channel Dataset for Glaucoma (SMDG).

Conclusions:

• The combination of CVGAN-based augmentation and contour guidance effectively enhances glaucoma classification.
• The proposed CA-ViT model offers a robust solution for early and accurate glaucoma detection.
• This approach holds potential for reducing vision impairment caused by glaucoma.