Grading of Foveal Hypoplasia Using Deep Learning on Retinal Fundus Images.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Foveal hypoplasia is a key indicator of foveal developmental disorders.
• Accurate grading of foveal hypoplasia is crucial for predicting visual outcomes.
• Current diagnostic methods can be subjective and time-consuming.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) model for grading foveal hypoplasia.
• To compare the DL model's performance against human clinicians.
• To assess the utility of DL in diagnosing foveal developmental disorders.

Main Methods:

• Retrospective analysis of 605 retinal fundus images from 303 patients.
• Development of a DL model for binary (normal vs. abnormal) and six-grade classification.
• Comparison of model performance with senior and junior clinician assessments.

Main Results:

• The EfficientNet_b1 model achieved 84.36% accuracy in binary classification and 78.21% in six-grade classification.
• The model demonstrated superior performance compared to both junior and senior clinicians (P < 0.00001).
• Higher grades of foveal hypoplasia correlated with worse visual outcomes (P < 0.001).

Conclusions:

• The developed deep learning model is efficient and accurate for grading foveal hypoplasia.
• The AI model surpasses clinician performance in diagnosing foveal developmental diseases.
• This highlights the importance of AI in pediatric clinical evaluation using retinal fundus images.