Source-linked AI summary
Deep learning for predicting refractive error from retinal fundus images
Avinash V. Varadarajan, Ryan Poplin, Katy Blumer, Christof Angermueller, Joe Ledsam, Reena Chopra, Pearse A. Keane, Greg S. Corrado, Lily Peng, Dale R. Webster
TL;DR
Uncorrected refractive error is a common cause of visual impairment, and prediction from retinal fundus images was not previously thought possible. The study trained deep learning models on fundus images, used attention techniques to visualize predictive features, and found high-accuracy prediction with attention extending beyond the foveal region.
Problem
Uncorrected refractive error is one of the most common causes of visual impairment.
Method
The study trained models on fundus images from two datasets and used attention techniques to visualize features associated with refractive error.
Results
Deep learning models predicted refractive error from retinal fundus images with high accuracy, although attention maps indicated that features outside the foveal region also contributed.
Takeaways & Limitations
The findings demonstrate that deep learning can make novel predictions from medical images and may be relevant where autorefractors are not readily available.
Takeaways & Limitations
The model estimated the spherical component of refractive error fairly well but performed less well for the cylindrical component.
Abstract
from arXiv · showhide
Refractive error, one of the leading cause of visual impairment, can be corrected by simple interventions like prescribing eyeglasses. We trained a deep learning algorithm to predict refractive error from the fundus photographs from participants in the UK Biobank cohort, which were 45 degree field of view images and the AREDS clinical trial, which contained 30 degree field of view images. Our model use the "attention" method to identify features that are correlated with refractive error. Mean absolute error (MAE) of the algorithm's prediction compared to the refractive error obtained in the AREDS and UK Biobank. The resulting algorithm had a MAE of 0.56 diopters (95% CI: 0.55-0.56) for estimating spherical equivalent on the UK Biobank dataset and 0.91 diopters (95% CI: 0.89-0.92) for the AREDS dataset. The baseline expected MAE (obtained by simply predicting the mean of this population) was 1.81 diopters (95% CI: 1.79-1.84) for UK Biobank and 1.63 (95% CI: 1.60-1.67) for AREDS. Attention maps suggested that the foveal region was one of the most important areas used by the algorithm to make this prediction, though other regions also contribute to the prediction. The ability to estimate refractive error with high accuracy from retinal fundus photos has not been previously known and demonstrates that deep learning can be applied to make novel predictions from medical images. Given that several groups have recently shown that it is feasible to obtain retinal fundus photos using mobile phones and inexpensive attachments, this work may be particularly relevant in regions of the world where autorefractors may not be readily available.
Abbreviations
The study addresses refractive error, a major cause of visual impairment, by training deep learning models on retinal fundus images from UK Biobank and AREDS. The models predicted spherical equivalent with substantially lower MAE than baseline, while attention maps highlighted the macula and additional retinal features.
- Background: Refractive error is a common cause of visual impairment, although spectacles or contact lenses can treat it; limited eye-care access affects many patients.Portable fundus-imaging tools may support screening where expert graders or autorefractors are limited.
- Method: The study trained a deep learning model using UK Biobank and AREDS fundus images, then applied attention techniques to identify predictive image features.The architecture combined ResNet and soft-attention components and predicted spherical equivalent.
- Results: 0.56D MAE was achieved on UK Biobank clinical validation data versus 0.91D on the AREDS clinical validation dataset.These results summarize joint prediction of spherical equivalent from the two datasets.
- Results: 86% of UK Biobank predictions and 65% of AREDS predictions were within 1D of actual values, compared with 50% and 45% for baseline, respectively.The model-baseline differences were significant at all margins of error.
- Interpretation: Attention maps consistently highlighted the macula, while retinal vessels, retinal-pigment cracks, and regions outside the fovea also contributed.The highlighted observations generalized across many images.
- Limitations and analyses: The model predicted the spherical component of refractive error fairly well but performed less well for the cylindrical component.Excluding patients with cataracts and/or AMD produced a small but significant accuracy improvement.
DISCUSSION
The study showed that deep learning can predict refractive error from retinal fundus images, a task not previously thought possible. Attention maps highlighted the fovea and other retinal regions, while dataset differences and limited validation constrain interpretation and clinical generalization.
- Deep learning models predicted refractive error from retinal fundus images with high accuracy, despite this task not previously being considered feasible.
- Attention maps consistently highlighted the foveal region, while the optic disc, temporal vessel arcades, and other features contributed to predictions to a lesser extent.
- Attention maps looked similar for hypermetropic and myopic images, suggesting that the network used the same regions across a spectrum of refractive errors.
- Attention maps identify anatomic correlates but do not establish causation; further validation across more diverse populations and independent settings is needed.
- The lower performance in AREDS may reflect its smaller field of view, fewer images, preexisting macular pathology, and different refraction measurement method.
- The model predicted spherical power better than cylindrical power, consistent with retinal images containing less information about corneal or lenticular toricity.
FIGURES & TABLES
The figures and tables describe the study cohorts, model architecture, validation performance, refractive-error components, subgroup analysis, and attention patterns across UK Biobank and AREDS.
- Population characteristics: The UK Biobank and AREDS cohorts are presented side by side with demographic and spherical-equivalent characteristics.Table 1 reports age, ethnicity, spherical equivalent, and refractive-error categories for both datasets.
- Model performance: Table 2 compares algorithm and mean-prediction baseline performance for spherical-equivalent estimation using MAE and R2.Confidence intervals are reported, with values expressed in diopters.
- Margin accuracy: Table 3 evaluates the model and baseline by whether spherical-equivalent predictions fall within specified error margins.The baseline uses the most populated sliding histogram window for each margin.
- Refractive-error components: Table 4 evaluates cylindrical and spherical refractive-error component predictions against mean-value baselines in UK Biobank.The table reports MAE and R2 with 95% confidence intervals in diopters.
- Model architecture: Figure 1 shows fundus images entering a residual network with attention, which outputs spherical equivalent, cylindrical component, and spherical component.The model learns parameters from input-output examples.