Automated vertical cup-to-disc ratio determination from fundus images for glaucoma detection

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of irreversible blindness, often asymptomatic until advanced stages.
• Early detection through optic nerve assessment in fundus images is critical for preventing vision loss.
• Manual evaluation of the vertical cup-to-disc ratio (VCDR) is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate a deep learning (DL) system for automated optic disc and cup detection in fundus images.
• To accurately calculate the VCDR using DL for glaucoma screening.
• To address the challenge of adapting DL models for VCDR estimation across different populations.

Main Methods:

• Utilized the YOLOv7 architecture for optic disc and cup segmentation in fundus images.
• Trained the DL model on multiple public datasets and fine-tuned it on the REFUGE dataset (Chinese population).
• Developed an optimization method for calibrating DL model performance across diverse populations.

Main Results:

• The DL-derived VCDR showed high accuracy, with a Pearson correlation coefficient of 0.91 and MAE of 0.0347 compared to expert assessments.
• The model outperformed existing methods on the REFUGE dataset in terms of Dice similarity and MAE.
• The developed optimization approach successfully adapted the model for a new population.

Conclusions:

• The automated DL system provides a robust and accurate tool for VCDR calculation, significantly reducing manual workload.
• This approach enhances the speed and accuracy of glaucoma detection.
• The system effectively differentiates glaucoma from non-glaucoma cases, serving as a valuable tool for early diagnosis.