Joint Optic Disc and Cup Segmentation Based on Multi-Label Deep Network and Polar Transformation

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible vision loss.
• Accurate segmentation of the optic disc (OD) and optic cup (OC) is crucial for glaucoma diagnosis via the cup to disc ratio (CDR).
• Existing methods often segment OD and OC separately and rely on manual feature engineering.

Purpose of the Study:

• To propose a novel deep learning architecture, M-Net, for joint, one-stage, multi-label segmentation of OD and OC.
• To improve the accuracy and automation of glaucoma screening through enhanced fundus image analysis.
• To evaluate M-Net's performance against existing methods on public datasets.

Main Methods:

• Developed M-Net, a deep learning model featuring a multi-scale input layer, U-shaped convolutional network, and side-output layer.
• Implemented a multi-label loss function for simultaneous OD and OC segmentation.
• Incorporated polar transformation to enhance image representation.

Main Results:

• M-Net achieved state-of-the-art OD and OC segmentation performance on the ORIGA dataset.
• The system demonstrated satisfactory glaucoma screening performance with accurate CDR calculation on both ORIGA and SCES datasets.
• Joint segmentation approach outperformed separate segmentation methods.

Conclusions:

• M-Net offers an effective and automated solution for joint OD and OC segmentation.
• The proposed method significantly advances the potential for early and accurate glaucoma detection.
• Deep learning, particularly M-Net, shows great promise in improving ophthalmic diagnostic tools.