Construction of an automatic optic disc and cup segmentation and cup-to-disc ratio calculation system for ocular fundus image and its application in glaucoma screening
10.3760/cma.j.cn115989-20250609-00191
- VernacularTitle:眼底图像视杯和视盘自动分割与杯盘比计算系统的构建及其在青光眼筛查中的应用
- Author:
Xiaoxuan LYU
1
;
Yang YANG
;
Jiani ZHAO
;
Qiuli YU
;
Cheng WAN
Author Information
1. 南京航空航天大学，南京 211106
- Publication Type:Journal Article
- Keywords:
Glaucoma;
Computer-aided diagnosis;
Image segmentation;
Deep learning;
Cup-to-disc ratio
- From:
Chinese Journal of Experimental Ophthalmology
2025;43(11):1007-1016
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a deep learning-based automated analysis system for precise segmentation of the optic cup and disc in fundus images and automatic measurement of the vertical cup-to-disc ratio (CDR) for early risk assessment and screening of chronic glaucoma.Methods:The proposed automated system comprised three modules: a dual coding-attention U-net (DCoAtUNet) segmentation network for optic cup and disc segmentation, a conditional random field (CRF) post-processing module, and a CDR measurement and glaucoma screening module based on the segmentation results.The system was designed to enhance boundary detection accuracy and measurement stability and its performance was evaluated on the publicly available Drishti-GS dataset.The dataset was divided into a training set and a test set in a 1∶1 ratio.Dice coefficient and intersection over union (IoU) were used to quantify segmentation accuracy and regional consistency, while accaracy, precision, recall, and F1-score were employed to assess glaucoma screening performance.Results:The DCoAtUNet combined with CRF post-processing achieved Dice coefficients of 0.976 0 for the optic disc and 0.908 1 for the optic cup, with corresponding IoU values of 0.953 4 and 0.837 9, demonstrating high segmentation precision and stability in boundary identification and region overlap.In glaucoma screening, the system achieved an accuracy of 0.843 1, precision of 0.840 9, recall of 0.973 7, and F1-score of 0.902 4, indicating good diagnostic sensitivity and accuracy.Conclusions:By integrating high-precision segmentation and automated measurement strategies, the DCoAtUNet+ CRF model significantly improves the accuracy and stability of CDR evaluation.It effectively assists in identifying high-risk individuals during early glaucoma screening and shows strong potential for clinical application in computer-aided diagnosis workflows.
