Related Experiment Video
Updated: Jun 25, 2026
05:10
An Automated Method for Assessing Visual Acuity in Infants and Toddlers Using an Eye-Tracking System
Published on: March 17, 2023
DeepAdapter: a generalisable algorithm integrating self-supervised learning and unsupervised domain adaptation for
Jiaman Zhao1,2, Longhui Li1, Zhenzhe Lin1
1Sun Yat-sen University, Zhongshan Ophthalmic Center, State Key Laboratory of Ophthalmology, Guangzhou, Guangdong, China.
The British Journal of Ophthalmology
|June 12, 2026
Summary
DeepAdapter, a novel deep learning algorithm, enhances retinopathy of prematurity (ROP) screening by integrating self-supervised learning and unsupervised domain adaptation. This approach significantly improves model generalizability across diverse datasets.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Retinopathy of prematurity (ROP) screening requires accurate and generalizable AI models.
- Domain shift, caused by data distribution differences, often hinders model performance in real-world clinical settings.
Purpose of the Study:
- To develop and validate DeepAdapter, a novel deep learning algorithm for ROP screening.
- To enhance model generalizability by integrating self-supervised learning (SSL) and unsupervised domain adaptation (UDA).
Main Methods:
- SSL was applied to 500,000 unlabelled infantile fundus photographs to learn general representations.
- A UDA module was incorporated to mitigate domain shift, quantified using Correlation Alignment (CORAL) distance.
- External validation was performed on independent and multiethnic datasets.
Main Results:
- DeepAdapter significantly reduced CORAL distance, indicating effective domain shift mitigation.
- DeepAdapter achieved superior external testing accuracy (0.828) compared to the supervised method (0.739).
- The algorithm also demonstrated improved performance on a cross-ethnicity dataset (0.839 vs 0.813).
Conclusions:
- DeepAdapter effectively addresses domain shift, enhancing model generalizability for ROP screening.
- The developed algorithm provides a valuable framework for creating generalizable AI models in medical specialties.
- A web-based system was created for large-scale, multicenter clinical screening.