Joint optic disk and cup segmentation for glaucoma screening using a region-based deep learning network.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Glaucoma screening relies on accurate measurement of the cup-to-disc ratio (CDR) from retinal fundus images.
• Manual segmentation of the optic disc (OD) and optic cup (OC) is time-consuming and subjective.
• Automated methods are needed for efficient and reliable glaucoma detection.

Purpose of the Study:

• To develop and validate a region-based deep convolutional neural network (R-DCNN) for joint OD and OC segmentation.
• To enable precise cup-to-disc ratio (CDR) measurement for glaucoma screening.
• To assess the performance of the R-DCNN against human experts and on public datasets.

Main Methods:

• A region-based deep convolutional neural network (R-DCNN) was developed for simultaneous OD and OC segmentation.
• The segmentation task was formulated as an object detection problem.
• Performance was evaluated using Dice similarity coefficient (DC), Jaccard coefficient (JC), and other metrics on in-house and public datasets (DRISHIT-GS, RIM-ONE v3).

Main Results:

• The R-DCNN achieved high segmentation accuracy, with Dice coefficients of 98.51% for OD and 97.63% for OC on the in-house dataset.
• Performance on public datasets (DRISHIT-GS, RIM-ONE v3) demonstrated robust generalization capabilities.
• The model's performance was comparable to that of experienced ophthalmologists.

Conclusions:

• The proposed R-DCNN demonstrates high accuracy and robustness for joint OD and OC segmentation.
• This automated approach shows significant potential for computer-assisted glaucoma screening.
• The tool can facilitate more efficient and objective CDR measurement in clinical practice.