Deep Transfer Learning for Ethnically Distinct Populations: Prediction of Refractive Error Using Optical Coherence...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning models struggle with performance disparities across ethnic groups due to training and testing data distribution mismatches.
• Posterior segment optical coherence tomography (OCT) images offer potential for refractive error prediction.

Purpose of the Study:

• To develop and validate a deep transfer learning model for predicting uncorrected refractive errors (spherical equivalent and keratometry) using OCT images.
• To reduce performance discrepancies in deep learning models across ethnically diverse populations.

Main Methods:

• A deep learning model was pre-trained on OCT images from a South Korean dataset.
• Transfer learning with adaptation training was applied using OCT images from an Indian dataset for domain adaptation.
• Models were trained to predict spherical equivalent (SE) and mean keratometry (K) values.

Main Results:

• Adaptation training significantly improved SE and K prediction models compared to those without adaptation (P < 0.001).
• The best performance for myopia/hyperopia classification and SE estimation was achieved using optic disc OCT images (74.7% accuracy, 1.58 D MAE).
• The optic nerve horizontal model demonstrated the lowest MAE (1.85 D) for predicting K values.

Conclusions:

• Adaptation training via transfer learning is effective for estimating refractive errors and K values from OCT images in ethnically diverse populations.
• The study highlights the potential of using macular and optic nerve OCT images for refractive error assessment.
• Further research with larger, diverse datasets is recommended to validate the algorithm's feasibility.