Explainable Transformer-Based Framework for Glaucoma Detection from Fundus Images Using Multi-Backbone Segmentation...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma, a leading cause of irreversible blindness, is increasing globally, exacerbated by a shortage of specialists.
• Early diagnosis and intervention are crucial to prevent vision loss.

Purpose of the Study:

• To develop an explainable, end-to-end AI pipeline for automated glaucoma diagnosis from fundus images.
• To evaluate Vision Transformers (ViTs) against traditional Convolutional Neural Network (CNN) models for glaucoma diagnosis.

Main Methods:

• Utilized YOLOv11 for optic disc detection and U-Net/MaskFormer for optic disc/cup segmentation across three datasets (REFUGE, ORIGA, G1020).
• Employed ResNet50, VGG16, MobileNetV2, and Swin-Base backbones for segmentation.
• Classified glaucoma based on the vertical cup-to-disc ratio (vCDR) using explainable AI principles.

Main Results:

• MaskFormer achieved superior segmentation performance (IoU/DSC scores >88%).
• The classification model reached 84.03% accuracy and 84.56% F1-score.
• Vision Transformers demonstrated high segmentation performance with explainable, biomarker-driven diagnosis.

Conclusions:

• The proposed framework offers a transparent and trustworthy solution for glaucoma screening by leveraging interpretable features like vCDR.
• Explainable AI and Vision Transformers present a promising direction for accurate and reliable automated glaucoma diagnosis.