A Lightweight Vision Transformer and Retinal Biomarker Fusion Framework for Early Glaucoma Detection: Toward Improved...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness, often detected late due to asymptomatic progression.
• Current automated methods struggle to identify subtle, early structural changes in retinal images.
• Early detection is crucial for preventing vision loss in glaucoma patients.

Purpose of the Study:

• To develop an automated framework for earlier and more consistent glaucoma screening using deep learning.
• To integrate optic nerve structure analysis with retinal biomarkers for improved glaucoma detection.
• To enhance the accuracy of automated glaucoma risk assessment.

Main Methods:

• A deep learning pipeline combining segmentation and classification on fundus images (ORIGA, REFUGE2 datasets).
• Image preprocessing included resizing, denoising, contrast enhancement, and optic nerve region cropping.
• A custom CNN segmented optic disc/cup for biomarker extraction (e.g., cup-to-disc ratio), complemented by a Vision Transformer for global retinal patterns.

Main Results:

• High segmentation accuracy for optic disc (Dice 0.9082) and optic cup (Dice 0.9994).
• Excellent overall segmentation performance with mean Dice of 0.9538 and mean IoU of 0.9169.
• Strong classification performance with high F1-score, recall, accuracy, and precision for both normal and glaucomatous cases.

Conclusions:

• The framework effectively combines established clinical biomarkers with advanced deep learning for glaucoma risk assessment.
• This approach shows promising performance for automated glaucoma detection using retinal fundus images.
• The integration of interpretable biomarkers and global feature learning enhances diagnostic capabilities.