Refractive error detection in smartphone images via convolutional neural network

Background and Objective: Refractive error, a common vision impairment, can cause serious problems such as amblyopia. Current vision screening relies on expensive equipment and trained optometrists, limiting accessibility, especially in less developed regions. Recent studies suggest that smartphone images can be analyzed for refractive errors, which can potentially democratize vision screening. This study investigates using CNN-based models to accurately estimate refractive error and to screen visually significant myopic refractive error. Methods: Data were collected from 93 participants aged 7 to 23 years (mean age 10.3, standard deviation 2.61). Our proposed method sarts with CNN models pre-trained on common images from the ImageNet dataset, which are then fine-tuned with data augmentation to address the challenge of data insufficiency. We explore different ways of applying the learned CNN features to improve the robustness and efficiency of the model in two applications, namely refractive error estimation and binary classification. Specifically, this study explored the use of MobileNetV2, EfficientNetB0, and ResNet18. Results: The best model, achieved by MobileNetV2, demonstrated promising performance in refractive error estimation, achieving a mean absolute error of approximately 0.616, and around 85.3% accuracy for binary refractive error detection. Conclusions: This study is the first to use CNN-based models to estimate refractive error and to screen for visually significant myopic refractive error. The proposed method shows potential as an accessible and efficient solution for vision screening.