Related Experiment Video
Updated: May 3, 2026
Subjective Refraction Test Using a Smartphone for Vision Screening
Published on: October 18, 2024
EED-Astig: A Multimodal Dataset for Pediatric Astigmatism Severity Prediction
Haihua Liu1, Shengyang Li2,3, Yixuan Lv4
1Department of Ophthalmology Center, Peking University First Hospital, Beijing, 100034, China.
Insights
A new AI dataset, EED-Astig, uses smartphone images to predict astigmatism in young children. This approach offers a cost-effective solution for early visual impairment screening in pediatric eye health.
Area of Science:
- Ophthalmology
- Computer Science
- Pediatric Healthcare
Background:
- Astigmatism is a common refractive error in preschoolers, leading to preventable visual impairment.
- Current screening methods are often impractical for young children due to cost, equipment, and cooperation requirements.
Purpose of the Study:
- To introduce EED-Astig, a multimodal pediatric dataset for AI-driven astigmatism severity prediction.
- To develop scalable and interpretable AI tools for early detection of pediatric eye conditions, especially in underserved areas.
Main Methods:
- Collected periocular images from 640 children (aged 3-6) using smartphones under standardized conditions.
- Utilized expert-verified annotations and the Segment Anything Model (SAM) for a semi-automated annotation pipeline.
- Derived six structural parameters and incorporated behavioral/demographic metadata for supervised learning.
Main Results:
- Technical validation showed robust performance in keypoint detection and image segmentation.
- The dataset enables the derivation of clinically relevant structural parameters linked to astigmatism.
- The multimodal approach integrates imaging and metadata for comprehensive analysis.
Conclusions:
- The EED-Astig dataset supports the development of AI tools for pediatric astigmatism prediction.
- This AI-based approach offers a promising, accessible solution for early visual screening in children.
- The findings highlight the potential for improving pediatric eye health outcomes, particularly in resource-limited settings.
Abstract:
Astigmatism is a prevalent refractive error in preschool children and a leading cause of preventable early visual impairment. Conventional screening methods are often unsuitable for young children due to high costs, specialized equipment, and the need for active cooperation. To address these challenges, we present EED-Astig, a multimodal pediatric dataset for artificial intelligence based astigmatism severity prediction. The dataset comprises periocular images from 640 children aged 3-6 years, acquired with smartphones under standardized conditions, with expert-verified annotations of corneal masks and anatomical landmarks. From these, we derive six clinically relevant structural parameters, including corneal exposure ratio and eyelash orientation, that are physiologically linked to astigmatism. In addition, behavioral and demographic metadata (e.g., screen time, birth history) provide complementary predictors for supervised learning. A semi-automated annotation pipeline based on the Segment Anything Model (SAM) ensures labeling consistency and quality. Technical validation demonstrates robust performance in keypoint detection and image segmentation, supporting the development of interpretable and scalable AI tools for pediatric eye health, particularly in low-resource settings.
