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

This study developed deep learning (DL) models to predict axial length (AL) in 6-10-year-old schoolchildren using minimally abnormal color fundus photographs (CFPs), while evaluating the impact of integrating age, Diopter Sphere (DS), and sex. Following quality assessment of 5460 initial CFPs from 3840 children, 3840 images from 2779 children were utilized and partitioned into training (70 %), validation (20 %), and test (10 %) sets. ResNet101 served as the core architecture, with supplemental clinical parameters integrated into the fully connected layer for continuous AL prediction. Model interpretation employed Grad-CAM-generated heatmaps. Comparative analysis demonstrated that DS and age achieved moderate predictive accuracy (R2 = 0.37), a CFP-only model showed significantly stronger performance (R2 = 0.70), and combining CFPs with DS and age further improved accuracy (R2 = 0.75). However, incorporating sex alongside CFPs, DS, and age substantially reduced efficacy (R2 = 0.41). Heatmaps revealed that regions critical for AL predictions anatomically corresponded to retinal vasculature and immediate perivascular tissues. These findings collectively indicate that DL may leverage near-normal CFPs for pediatric AL prediction, with selective enhancement by age and DS, but degradation when categorical variables (such as sex) are included. Subtle changes in the fundus vasculature may help DL to identify the cause of CFP changes with AL.