Estimation of best corrected visual acuity based on deep neural network.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate estimation of best-corrected visual acuity (BCVA) is crucial for diagnosing and managing eye conditions.
• Current methods for BCVA assessment can be subjective and require specialized equipment and trained personnel.
• There is a need for objective, automated methods for BCVA estimation, particularly in resource-limited settings.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN)-based framework for estimating BCVA from fundus images.
• To assess the accuracy and reliability of the proposed CNN schemes.
• To investigate the regions of interest within fundus images utilized by the CNN for BCVA prediction.

Main Methods:

• Collected a dataset of 53,318 fundus photographs with 11-level BCVA categorization.
• Designed four BCVA estimation schemes using transfer learning with pre-trained ResNet-18 and EfficientNet-B0 models.
• Employed both regression and classification approaches for prediction, and Guided Grad-CAM for interpretability.

Main Results:

• Achieved 94.37% prediction accuracy with a 3-level tolerance.
• Reported a mean squared error of 0.028 and an [Formula: see text] score of 0.654.
• Demonstrated accurate prediction for extreme BCVA values (0.0 or 1.0) and identified macula and surrounding blood vessels as key predictive areas.

Conclusions:

• CNN-based schemes provide accurate and objective BCVA estimation from fundus images.
• The developed framework has potential for widespread application in medical screening and remote patient monitoring.
• The interpretability analysis validates the clinical relevance of the CNN's predictive features.