Related Experiment Video
Updated: Dec 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Optic disc and optic cup segmentation based on anatomy guided cascade network
Xuesheng Bian1, Xiongbiao Luo1, Cheng Wang1
1Fujian Key Laboratory of Sensing and Computing for Smart Cities, Department of Computer Science, School of Informatics, Xiamen University, Xiamen 361005, China.
Computer Methods and Programs in Biomedicine
|September 21, 2020
Summary
A new deep neural network accurately segments optic discs and cups in fundus images for glaucoma screening. This method achieves state-of-the-art performance, improving early diagnosis of irreversible vision damage.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Glaucoma is a leading cause of irreversible blindness worldwide.
- Early diagnosis is crucial for preventing vision loss.
- Retinal fundus photography is a cost-effective screening method, with the cup-to-disk ratio being a key diagnostic indicator.
Purpose of the Study:
- To develop a precise fundus image segmentation method for calculating the cup-to-disk ratio.
- To improve glaucoma screening accuracy through automated image analysis.
Main Methods:
- A deep neural network incorporating anatomical knowledge for fundus image segmentation.
- An attention-based cascade network designed for accurate segmentation of small targets like the optic disc and cup.
- Focus on accelerating training convergence and preserving fine details in segmentation.
Main Results:
- Achieved a 93.31% dice score for optic disc segmentation and 88.04% for optic cup segmentation.
- Demonstrated high performance in the MICCAI REFUGE challenge.
- Obtained a high cup-to-disk ratio (CDR) evaluation score, vital for glaucoma screening.
Conclusions:
- The proposed method effectively integrates anatomical knowledge into the segmentation task.
- Achieved state-of-the-art results in fundus image segmentation.
- The method supports both automatic and semi-automatic segmentation for clinical application.
