Joint optic disc and optic cup segmentation based on boundary prior and adversarial learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma screening relies on optic cup (OC) and optic disc (OD) segmentation from fundus photography.
• Convolutional Neural Networks (CNNs) show promise but struggle with boundary ambiguity, limiting generalization.
• Accurate segmentation is crucial for calculating the cup-to-disc ratio, a key glaucoma indicator.

Purpose of the Study:

• To address the limitations of existing CNNs in segmenting optic disc and optic cup, particularly regarding boundary ambiguity.
• To develop a novel segmentation architecture that improves the generalization and accuracy of OD and OC segmentation.
• To enhance glaucoma screening through precise medical image analysis.

Main Methods:

• Proposed BGA-Net, a novel segmentation architecture incorporating an auxiliary boundary branch.
• Utilized adversarial learning with a generative adversarial network (GAN) for joint multi-label segmentation of OD and OC.
• Focused on encouraging boundary and mask predictions to closely match ground truth data.

Main Results:

• BGA-Net achieved state-of-the-art performance in optic disc and optic cup segmentation across three public datasets (Drishti-GS, RIM-ONE-r3, REFUGE).
• Achieved high Dice scores: 0.975/0.898 (OD/OC) on Drishti-GS, 0.967/0.872 on RIM-ONE-r3, and 0.951/0.866 on REFUGE.
• Demonstrated superior segmentation accuracy compared to existing methods.

Conclusions:

• The BGA-Net system provides superior optic disc and optic cup segmentation results.
• Confirmed the strong correlation between the cup-to-disc ratio, derived from segmentation, and glaucoma.
• Highlights the potential of advanced AI models for improved glaucoma detection and patient care.