Direct Cup-to-Disc Ratio Estimation for Glaucoma Screening via Semi-Supervised Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible vision loss.
• The Cup-to-Disc Ratio (CDR) is a critical indicator for glaucoma screening and diagnosis.
• Accurate CDR measurement is challenging due to optic disc/cup overlap, hindering automated methods.

Purpose of the Study:

• To develop a direct CDR estimation method that bypasses optic disc/cup segmentation.
• To improve the accuracy and robustness of automated CDR calculation for glaucoma screening.
• To leverage deep learning for a more efficient and effective glaucoma diagnostic tool.

Main Methods:

• A semi-supervised learning scheme was designed for direct CDR estimation, treating it as a regression problem.
• A two-stage cascaded approach was employed: unsupervised feature representation (MFPPNet) followed by random forest regression.
• The method directly regresses CDR from optic nerve head features, avoiding intermediate segmentation steps.

Main Results:

• The proposed method achieved a low average CDR error of 0.0563 and a high correlation (0.726) with expert measurements.
• Glaucoma screening using the estimated CDR values yielded an area under the curve of 0.905 on a dataset of 421 fundus images.
• The method demonstrates state-of-the-art CDR estimation and satisfactory performance in glaucoma screening.

Conclusions:

• Direct CDR estimation via deep learning offers a robust alternative to segmentation-based methods.
• The proposed approach shows significant potential for improving early glaucoma detection and patient outcomes.
• This method provides a valuable tool for both clinical screening and automated analysis of fundus images.