SpecF2M network estimates pediatric myopia from fundus photos with 0.53mm AL accuracy

Measuring pediatric myopia typically requires dedicated biometry devices and cycloplegic refraction—expensive and impractical for large-scale screening. Fundus photography offers a cheaper alternative, but the visual cues for myopia are low-contrast, diffuse, and multi-scale. Now, researchers have developed SpecF2M, a spectral-aware multi-task network that directly estimates axial length (AL), sphere (SPH), and cylinder (CYL) from standard 45-degree fundus images.

SpecF2M combines three novel components: a deterministic anatomy-guided enhancement module that highlights myopia-related posterior-pole changes, a hybrid spatial-spectral backbone built from MixCNN and Hybrid Spectral Learning (HSL) blocks to capture both local fine features and global spectral patterns, and an expert-routing head that assigns specialized sub-networks to each output component (AL, SPH, CYL). This design acknowledges that these measurements share partially overlapping but non-identical anatomical correlates.

In validation on a pediatric cohort of 4,359 eligible visits (6,966 fundus images), SpecF2M outperformed conventional CNN and ViT baselines, achieving mean absolute errors of 0.5347 mm for axial length and 0.7062 D for spherical equivalent refraction. Notably, component-level analysis revealed asymmetric task coupling: cylinder (CYL) showed weaker association with fundus-derived myopic patterns than AL or SPH, suggesting that CYL prediction is fundamentally harder from these images.

The authors emphasize that fundus-based screening is feasible for pediatric myopia indicators, but external validation on diverse populations is required before clinical deployment. This work represents a step toward accessible, low-cost myopia screening for children, potentially replacing or triaging expensive biometry in resource-limited settings.