Dual Consistency Enabled Weakly and Semi-Supervised Optic Disc and Cup Segmentation With Dual Adaptive Graph...

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Accurate optic disc (OD) and optic cup (OC) segmentation is crucial for glaucoma diagnosis via vertical cup-to-disc ratio (vCDR) estimation.
• Current methods using fully supervised deep learning require large, costly annotated datasets.

Purpose of the Study:

• To develop a weakly and semi-supervised graph-based network for OD and OC segmentation and vCDR estimation.
• To reduce reliance on extensive pixel-level annotations in medical image analysis.
• To leverage geometric properties and domain knowledge for improved glaucoma assessment.

Main Methods:

• Proposed a Dual Adaptive Graph Convolutional Network (DAGCN) to analyze segmentation probability maps (PM) and modified signed distance function representations (mSDF).
• Implemented a dual consistency regularization paradigm for semi-supervised learning, enforcing consistency between PM and mSDF.
• Incorporated domain knowledge of oval OD/OC shapes using a differentiable vCDR estimation layer for weak supervision.

Main Results:

• Achieved superior performance in OD and OC segmentation across six large-scale datasets.
• Demonstrated accurate vCDR estimation without additional manual annotations.
• The proposed weakly and semi-supervised approach significantly outperforms existing methods.

Conclusions:

• The novel graph-based network effectively segments ocular structures and estimates vCDR with reduced annotation burden.
• This approach offers a more efficient and cost-effective solution for glaucoma screening and assessment.
• The method holds promise for advancing automated analysis of fundus images in clinical settings.