Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Brighteye: Glaucoma Screening with Color Fundus Photographs based on Vision Transformer

Hui Lin, Charilaos Apostolidis, Aggelos K. Katsaggelos

TL;DR

Brighteye introduces a ViT-based framework for glaucoma screening from color fundus photographs that first localizes the optic disc with YOLOv8 to crop a clinically relevant ROI, then applies patch-based feature aggregation within a Vision Transformer to perform glaucoma detection and glaucomatous feature classification. The approach achieves a highly accurate optic disc detector (AUC ≈ 0.995) and improves downstream task performance when OD cropping and background removal are applied, attaining a sensitivity of 85.70% at 95% specificity and a normalized Hamming distance of 0.1250. With 11 binary classification tasks shared across a single architecture, Brighteye demonstrates robust, interpretable feature-based screening under challenging imaging conditions and ranks 5th in the JustRAIGS development phase among 226 entries. This work provides a practical, scalable framework for automated glaucoma referral and glaucomatous feature analysis in fundus photography, with potential for integration into clinical screening pipelines.

Abstract

Differences in image quality, lighting conditions, and patient demographics pose challenges to automated glaucoma detection from color fundus photography. Brighteye, a method based on Vision Transformer, is proposed for glaucoma detection and glaucomatous feature classification. Brighteye learns long-range relationships among pixels within large fundus images using a self-attention mechanism. Prior to being input into Brighteye, the optic disc is localized using YOLOv8, and the region of interest (ROI) around the disc center is cropped to ensure alignment with clinical practice. Optic disc detection improves the sensitivity at 95% specificity from 79.20% to 85.70% for glaucoma detection and the Hamming distance from 0.2470 to 0.1250 for glaucomatous feature classification. In the developmental stage of the Justified Referral in AI Glaucoma Screening (JustRAIGS) challenge, the overall outcome secured the fifth position out of 226 entries.

Brighteye: Glaucoma Screening with Color Fundus Photographs based on Vision Transformer

TL;DR

Brighteye introduces a ViT-based framework for glaucoma screening from color fundus photographs that first localizes the optic disc with YOLOv8 to crop a clinically relevant ROI, then applies patch-based feature aggregation within a Vision Transformer to perform glaucoma detection and glaucomatous feature classification. The approach achieves a highly accurate optic disc detector (AUC ≈ 0.995) and improves downstream task performance when OD cropping and background removal are applied, attaining a sensitivity of 85.70% at 95% specificity and a normalized Hamming distance of 0.1250. With 11 binary classification tasks shared across a single architecture, Brighteye demonstrates robust, interpretable feature-based screening under challenging imaging conditions and ranks 5th in the JustRAIGS development phase among 226 entries. This work provides a practical, scalable framework for automated glaucoma referral and glaucomatous feature analysis in fundus photography, with potential for integration into clinical screening pipelines.

Abstract

Differences in image quality, lighting conditions, and patient demographics pose challenges to automated glaucoma detection from color fundus photography. Brighteye, a method based on Vision Transformer, is proposed for glaucoma detection and glaucomatous feature classification. Brighteye learns long-range relationships among pixels within large fundus images using a self-attention mechanism. Prior to being input into Brighteye, the optic disc is localized using YOLOv8, and the region of interest (ROI) around the disc center is cropped to ensure alignment with clinical practice. Optic disc detection improves the sensitivity at 95% specificity from 79.20% to 85.70% for glaucoma detection and the Hamming distance from 0.2470 to 0.1250 for glaucomatous feature classification. In the developmental stage of the Justified Referral in AI Glaucoma Screening (JustRAIGS) challenge, the overall outcome secured the fifth position out of 226 entries.
Paper Structure (8 sections, 1 equation, 3 figures, 1 table)

This paper contains 8 sections, 1 equation, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Optic Disc Detection
  4. Classification based on Vision Transformer
  5. Experiments
  6. Results and Discussion
  7. Conclusions
  8. Compliance with Ethical Standards

Figures (3)

  • Figure 1: The proposed framework comprising optic disc detection, glaucoma detection, and feature classification. In the first step, a region of interest (ROI) around the detected optic disc (OD) is cropped. In the second step, the presence of glaucoma is predicted for Task 1's output, and 10 binary classifiers predict the presence of 10 individual features independently for Task 2's output.
  • Figure 2: The architecture of Brighteye, modified from ViT dosovitskiy2021image for binary classification. All 11 classifiers in Fig. \ref{['fig: pipeline']} share this architecture.
  • Figure 3: Top three most extreme false-negative (first three columns) and false-positive (last three columns) cases. The value displayed in the bottom right corner indicates the probability of glaucoma. The bottom left corner in the first three columns shows the glaucomatous features identified by the graders, with the area indicated by the yellow arrows.