Deep Imbalanced Regression Model for Predicting Refractive Error from Retinal Photos.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) models show promise for predicting refractive error from ocular images.
• Existing studies often neglect dataset imbalances and external validation, limiting real-world applicability.

Purpose of the Study:

• To integrate deep imbalanced regression (DIR) into ResNet and Vision Transformer models for refractive error prediction.
• To address biases from imbalanced datasets and improve prediction accuracy using retinal photographs.

Main Methods:

• Developed and compared ResNet34 and SwinV2 (Vision Transformer) models with and without DIR integration.
• Incorporated Label Distribution Smoothing and Feature Distribution Smoothing techniques.
• Validated models on internal datasets (Singapore Epidemiology of Eye Diseases Study, UK Biobank) and external datasets (Singapore Prospective Study, Beijing Eye Study).

Main Results:

• DIR-integrated models (ResNet34-DIR, SwinV2-DIR) significantly outperformed baseline models in predicting spherical and spherical equivalent (SE) power.
• Lower Mean Absolute Errors (MAE) were achieved by DIR models: 0.84D (ResNet34-DIR) and 0.77D (SwinV2-DIR) for spherical power; 0.78D (ResNet34-DIR) and 0.75D (SwinV2-DIR) for SE power.
• Consistent performance improvements were observed across both internal and external test datasets.

Conclusions:

• Deep imbalanced regression effectively mitigates data imbalances in refractive error prediction models.
• DIR-integrated DL models demonstrate significant potential for accurate refractive error prediction using retinal images.
• This approach facilitates opportunistic screening for refractive errors, especially in settings with existing retinal imaging technology.