Ocular image-based deep learning for predicting refractive error: A systematic review.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Uncorrected refractive error is a significant global cause of vision impairment.
• Deep learning (DL) offers advanced solutions for ophthalmological diagnostics.
• There is a need for a systematic review on DL models for refractive error detection.

Purpose of the Study:

• To systematically review and evaluate ocular image-based deep learning models for refractive error prediction.
• To summarize the performance of DL models across different imaging modalities.

Main Methods:

• Systematic search of PubMed, Scopus, and Web of Science databases up to June 2023.
• Inclusion of studies reporting refractive error outcomes, regardless of publication year.
• Extraction and evaluation of continuous and categorical outcomes, imaging modalities, DL models, and performance metrics following PRISMA guidelines.

Main Results:

• Nine studies were categorized into retinal photo-based (5), OCT-based (1), and external ocular photo-based (3).
• Retinal photo-based models showed high performance for myopia prediction (AUC 0.91-0.98) and continuous prediction (MAE 0.31-2.19D).
• External ocular photo-based models demonstrated strong performance (AUC 0.91-0.99, MAE 0.07-0.18D).

Conclusions:

• Deep learning models integrated with ocular imaging show promising results for refractive error detection.
• Retinal and external eye photo-based DL models exhibit particularly strong performance.
• Further evaluation of real-world clinical utility and implementation strategies is necessary.