Related Experiment Video
Updated: Jan 13, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Enhancing ocular sign detection: AI-based strategic segmentation for improved accuracy and privacy protection
Chaoyu Lei1,2, Chen Zhao1, Jiayu Chen3
1State Key Laboratory of Eye Health, Department of Ophthalmology, Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
NPJ Digital Medicine
|January 6, 2026
Summary
This study introduces a novel AI model for detecting eye disease signs using limited facial images, significantly improving diagnostic accuracy while protecting patient privacy. The developed system demonstrates high performance and generalizability across diverse populations.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Current AI for eye disease diagnosis uses external images, often including non-essential data, which hinders performance and compromises patient privacy.
- Accurate detection of ocular signs is crucial for the early diagnosis of various eye conditions.
Purpose of the Study:
- To develop a privacy-preserving AI system for accurate ocular sign detection in thyroid eye disease.
- To minimize data exposure by focusing on essential ocular regions and enhance diagnostic performance.
Main Methods:
- A multinational retrospective study involving 2360 eyes from 1180 half-face images across diverse racial groups and hospitals.
- Development of a Dense Squeeze-and-Excitation Network (DSE-Net) for segmenting key ocular structures (eyelid, conjunctiva, lacrimal caruncle, eyeball).
- Implementation of SegmenView, utilizing convolutional neural networks (LeNet, AlexNet, ResNet50, VGGNet16) for detecting specific ocular signs like edema and exophthalmos.
Main Results:
- DSE-Net achieved high segmentation performance with Dice coefficients up to 95.1%, outperforming seven other models.
- SegmenView demonstrated strong diagnostic capabilities with internal Area Under the Curve (AUC) scores up to 90.07% and external AUCs ranging from 55.58% to 84.29%.
- The developed system outperformed existing half-face and periocular models and balanced diagnostic accuracy with privacy preservation effectively.
Conclusions:
- The proposed AI approach, SegmenView, offers accurate, generalizable, and privacy-preserving detection of ocular signs.
- The system shows potential for lightweight deployment in clinical settings for early eye disease diagnosis.
- Gradient-weighted Class Activation Mapping (Grad-CAM) visualization enhanced model interpretability.
