Identification of ocular refraction based on deep learning algorithm as a novel retinoscopy method.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Ocular refraction evaluation is crucial in ophthalmology, typically requiring specialized equipment and procedures like cycloplegic refraction.
• Retinal fundus photographs (RFPs) offer a rich source of ocular information and present a potential for more convenient and objective refraction assessment.

Purpose of the Study:

• To develop and validate a fusion model-based deep learning system (FMDLS) for identifying ocular refraction using RFPs.
• To compare the performance of the FMDLS against standard cycloplegic refraction measurements.

Main Methods:

• A retrospective analysis of 11,973 RFPs collected between May 2020 and November 2021.
• Development of a deep learning system (FMDLS) to analyze RFPs for refractive error components (sphere, cylinder, axis).
• Evaluation of regression models using Mean Absolute Error (MAE) and correlation coefficients (r); classification models assessed using accuracy, sensitivity, specificity, AUC, and F1-score.

Main Results:

• The FMDLS achieved Mean Absolute Error (MAE) values of 0.50 D for sphere and 0.31 D for cylinder, demonstrating improved performance over single models.
• Correlation coefficients (r) for sphere and cylinder were 0.949 and 0.807, respectively.
• The classification model for cylinder axis achieved high performance with an accuracy of 0.89, specificity of 0.941, sensitivity of 0.882, AUC of 0.814, and F1-score of 0.88.

Conclusions:

• The FMDLS effectively identified ocular refraction parameters (sphere, cylinder, axis) from RFPs with good agreement to cycloplegic refraction.
• Retinal fundus photographs possess significant potential for clinical use beyond fundus imaging, including objective refractive state assessment.
• The developed FMDLS highlights the clinical value of leveraging RFPs for comprehensive eye examinations.