[Joint optic disc and cup segmentation based on residual multi-scale fully convolutional neural network]

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Glaucoma is a leading cause of irreversible blindness, often presenting with subtle early symptoms.
• Early detection through screening is vital for preventing vision loss.
• The optic cup-to-disc ratio is a key metric in clinical glaucoma assessment, requiring precise optic cup and disc segmentation.

Purpose of the Study:

• To develop and evaluate an accurate deep learning model for simultaneous optic cup and disc segmentation in fundus images.
• To improve the accuracy of calculating the optic cup-to-disc ratio for enhanced glaucoma screening.

Main Methods:

• A W-Net based fully convolutional neural network architecture was employed.
• A novel residual multi-scale convolution module was integrated into the W-Net backbone.
• Multi-scale input using an image pyramid and a side output layer for early classification were utilized.
• A new multi-tag loss function was introduced to guide the segmentation process.

Main Results:

• The proposed method achieved a mean intersection over union (IoU) of 0.904 for optic cup segmentation and 0.955 for optic disc segmentation on the REFUGE dataset.
• The overlapping error was reported as 0.178 for the optic cup and 0.066 for the optic disc.
• The model demonstrated effective joint segmentation of both the optic cup and disc.

Conclusions:

• The developed deep learning approach significantly enhances the accuracy of optic cup and disc segmentation.
• This improved segmentation accuracy holds promise for facilitating large-scale, early glaucoma screening initiatives.
• The method provides a robust tool for precise optic cup-to-disc ratio calculation, aiding ophthalmologists in clinical decision-making.