Axial length prediction Model based on screening fundus photography data in school-age children

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Pediatric Health

Background:

• Axial length (AL) is a critical parameter in pediatric eye development and refractive error progression.
• Accurate AL prediction is essential for monitoring and managing childhood myopia.
• Current methods for AL measurement can be invasive or require specialized equipment.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for predicting axial length (AL) in schoolchildren.
• To assess the impact of integrating clinical data (age, Diopter Sphere, sex) with color fundus photographs (CFPs) for AL prediction.
• To interpret DL model predictions using heatmaps to understand feature importance.

Main Methods:

• Utilized 2,779 minimally abnormal color fundus photographs (CFPs) from 6-10-year-old children.
• Employed a ResNet101 architecture for DL model development, integrating clinical parameters into the fully connected layer.
• Partitioned data into training (70%), validation (20%), and test (10%) sets for robust model evaluation.
• Used Grad-CAM heatmaps for model interpretability.

Main Results:

• A CFP-only DL model achieved high predictive accuracy (R² = 0.70).
• Integrating age and Diopter Sphere (DS) with CFPs further improved AL prediction accuracy (R² = 0.75).
• Incorporating sex information alongside CFPs, age, and DS significantly reduced predictive efficacy (R² = 0.41).
• Heatmaps indicated that retinal vasculature and perivascular tissues were key features for AL prediction.

Conclusions:

• Deep learning models can effectively predict pediatric axial length using color fundus photographs.
• Age and refractive error (DS) enhance DL-based AL prediction, while categorical variables like sex may degrade performance.
• DL models leverage subtle fundus vascular changes for AL prediction, offering insights into refractive development.