Deep learning for predicting refractive error from multiple photorefraction images.

Area of Science:

• Ophthalmology
• Computer Vision
• Artificial Intelligence

Background:

• Refractive error detection is crucial for preventing myopia development.
• Current methods require improvement in efficiency and accuracy.
• Deep learning offers potential for enhanced detection systems.

Purpose of the Study:

• To develop an advanced system for predicting refractive errors using multiple eccentric photorefraction images.
• To improve the accuracy and efficiency of refractive error detection.

Main Methods:

• A novel Refractive Error Detection Network (REDNet) was developed, combining Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) capabilities.
• Pupil area images were extracted from multiple photorefraction images.
• REDNet's convolutional layers extracted image features, and recurrent layers fused these features for prediction.

Main Results:

• The system achieved a mean absolute error (MAE) of 0.1740 D for spherical power.
• The MAE for cylindrical power was 0.0702 D.
• The MAE for spherical equivalent was 0.1835 D.

Conclusions:

• The developed REDNet method significantly outperforms current state-of-the-art deep learning techniques.
• The system demonstrates high accuracy, effectiveness, and practicality for refractive error detection.