Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Non-Invasive Estimation of Cycloplegic Refraction in Children from Color Fundus Photographs Using a Multi-Task Deep Learning Framework

Authors
Zhenhan Wang, Danli SHI

Presenting
Zhenhan Wang

PURPOSE:
Cycloplegic refraction is the clinical gold standard for pediatric refractive assessment but is impractical for population-level screening. This study developed and evaluated a multi-task deep learning model to non-invasively estimate cycloplegic spherical equivalent refraction (SER) and axial length (AL) from color fundus photographs in children, with mechanistic interpretability analysis.

METHODS:
A total of 37,523 fundus images from 12,940 pediatric patients were partitioned by a strict temporal split (training: validation: test=8:1:1). A ConvNeXt-Base backbone with parallel task-specific heads was trained to simultaneously predict SER, AL, and five-class refractive status (hyperopia, emmetropia, pre-myopia, myopia, high myopia) using Huber and cross-entropy losses. Gradient-weighted Class Activation Mapping (Grad-CAM) and Testing with Concept Activation Vectors (TCAV) provided mechanistic explainability. All metrics were bootstrapped for 95% confidence intervals.

METHODS:

A total of 37,523 fundus images from 12,940 pediatric patients were partitioned by a strict temporal split (training: validation: test=8:1:1). A ConvNeXt-Base backbone with parallel task-specific heads was trained to simultaneously predict SER, AL, and five-class refractive status (hyperopia, emmetropia, pre-myopia, myopia, high myopia) using Huber and cross-entropy losses. Gradient-weighted Class Activation Mapping (Grad-CAM) and Testing with Concept Activation Vectors (TCAV) provided mechanistic explainability. All metrics were bootstrapped for 95% confidence intervals.

RESULTS:
The model achieved near-clinical-grade discrimination for high myopia (AUC = 0.983; 95% CI: 0.969–0.994) and myopia (AUC = 0.933; 95% CI: 0.925–0.941), comparable to leading dedicated myopia cohort systems. Continuous predictions yielded strong correlations for SER (r = 0.831, R² = 0.660) and AL (r = 0.832, R² = 0.687). Heteroscedasticity analysis confirmed that residual variance in high myopia reflects biological heterogeneity of posterior segment deformations rather than algorithmic failure. Grad-CAM and TCAV identified peripapillary atrophy and disc tilt as dominant predictive features, consistent with established biomarkers of axial elongation.

RESULTS:

The model achieved near-clinical-grade discrimination for high myopia (AUC = 0.983; 95% CI: 0.969–0.994) and myopia (AUC = 0.933; 95% CI: 0.925–0.941), comparable to leading dedicated myopia cohort systems. Continuous predictions yielded strong correlations for SER (r = 0.831, R² = 0.660) and AL (r = 0.832, R² = 0.687). Heteroscedasticity analysis confirmed that residual variance in high myopia reflects biological heterogeneity of posterior segment deformations rather than algorithmic failure. Grad-CAM and TCAV identified peripapillary atrophy and disc tilt as dominant predictive features, consistent with established biomarkers of axial elongation.

CONCLUSIONS:
This model enables non-invasive estimation of cycloplegic refraction from fundus photographs with clinically meaningful accuracy and near-clinical-grade high myopia discrimination, supporting its potential as a scalable adjunct to pediatric refractive screening.