Related Experiment Video
Updated: Jun 17, 2026
10:14
Ex Vivo OCT-Based Multimodal Imaging of Human Donor Eyes for Research into Age-Related Macular Degeneration
Published on: May 26, 2023
Performance of Deep Learning in Classifying Age-Related Macular Degeneration From Images: Systematic Review and
Yu Zhu1, Yue Niu2, Shangye Sun3
1Department of Ophthalmology, Jilin Province FAW General Hospital, Changchun, 130011, China.
Journal of Medical Internet Research
|June 15, 2026
Summary
Deep learning (DL) algorithms show high accuracy in detecting age-related macular degeneration (AMD) and differentiating its subtypes. While promising, DL tools require further validation before clinical deployment.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Age-related macular degeneration (AMD) is a primary cause of irreversible blindness globally.
- Retinal imaging combined with deep learning (DL) offers potential for scalable AMD screening.
- Accurate diagnostic performance is crucial to avoid treatment delays and unnecessary referrals.
Purpose of the Study:
- To compare the diagnostic performance of DL algorithms against ophthalmologists for AMD detection.
- To evaluate DL's ability to differentiate wet AMD (wAMD) from dry AMD (dAMD).
- To identify factors influencing DL performance in AMD diagnosis.
Main Methods:
- A systematic literature search was conducted across major databases (PubMed, Embase, Web of Science, Cochrane Library).
- Eligible studies utilized DL for AMD classification from retinal images.
- Data extraction and risk of bias assessment (PROBAST+AI) were performed by two independent reviewers.
- Bivariate random-effects models were used to pool sensitivity, specificity, accuracy, and AUC.
- Comparisons with clinicians were stratified by experience; small-study effects and evidence certainty were assessed.
Main Results:
- Twenty-eight studies (77,485 samples for AMD, 28,705 for wAMD/dAMD) were included.
- For AMD detection, DL achieved pooled sensitivity of 0.98 and specificity of 0.98.
- For wAMD vs. dAMD classification, DL achieved pooled sensitivity of 0.95 and specificity of 0.95.
- DL demonstrated higher sensitivity than senior ophthalmologists for AMD detection and higher specificity/accuracy than junior ophthalmologists for wAMD classification.
- Optical coherence tomography-based DL models showed greater consistency.
Conclusions:
- DL algorithms exhibit superior and more balanced diagnostic performance compared to ophthalmologists in current head-to-head evidence.
- DL can potentially serve as a consistent decision-support tool, mitigating human variability.
- Preliminary findings necessitate further prospective, multicenter validation due to heterogeneity and potential performance inflation.
- DL should be considered a triage adjunct requiring local calibration, not a standalone diagnostic replacement.