Abstract
This study aimed to develop an automated computer-based algorithm to estimate axial length and subfoveal choroidal thickness (SFCT) based on color fundus photographs. In the population-based Beijing Eye Study 2011, we took fundus photographs and measured SFCT by optical coherence tomography (OCT) and axial length by optical low-coherence reflectometry. Using 6394 color fundus images taken from 3468 participants, we trained and evaluated a deep-learning-based algorithm for estimation of axial length and SFCT. The algorithm had a mean absolute error (MAE) for estimating axial length and SFCT of 0.56 mm [95% confidence interval (CI): 0.53,0.61] and 49.20 μm (95% CI: 45.83,52.54), respectively. Estimated values and measured data showed coefficients of determination of $r^2 = 0.59$ (95% CI: 0.50,0.65) for axial length and $r^2 = 0.62$ (95% CI: 0.57,0.67) for SFCT. Bland-Altman plots revealed a mean difference in axial length and SFCT of -0.16 mm (95% CI: -1.60,1.27 mm) and of -4.40 μm (95% CI, -131.8,122.9 μm), respectively. For the estimation of axial length, heat map analysis showed that signals predominantly from overall of the macular region, the foveal region, and the extrafoveal region were used in the eyes with an axial length of < 22 mm, 22-26 mm, and > 26 mm, respectively. For the estimation of SFCT, the convolutional neural network (CNN) used mostly the central part of the macular region, the fovea or perifovea, independently of the SFCT. Our study shows that deep-learning-based algorithms may be helpful in estimating axial length and SFCT based on conventional color fundus images. They may be a further step in the semiautomatic assessment of the eye.
Introduction
In ophthalmology, axial length (AL)—the front-to-back distance of the eye—and subfoveal choroidal thickness (SFCT)—the thickness of the vascular layer beneath the retina—are vital biomarkers. They are essential for managing myopia (nearsightedness) and diagnosing various maculopathies.
Currently, measuring these requires specialized, expensive equipment: optical biometers for AL and Optical Coherence Tomography (OCT) for SFCT. These devices are often unavailable in primary care or resource-limited settings.
The authors sought to determine if a deep learning algorithm could "see" these biometric properties within a standard color fundus photograph, which is much cheaper and easier to obtain.
Methods
The study utilized data from the Beijing Eye Study 2011, a large-scale, population-based study in Northern China. A total of 6,394 color fundus images from 3,468 participants (all aged 50 or older) were used. Ground truth AL was measured via optical low-coherence reflectometry (Lensstar 900), and SFCT was measured via spectral-domain OCT (Spectralis).
The team employed a convolutional neural network (CNN) based on the Inception-ResNet-v2 architecture, modified into a 164-layer model.
Key technical steps included:
- Preprocessing: Removing the dark background with circular masks and resizing images to 500 × 500 pixels.
- Quality Control: An automated module filtered out blurry or poorly illuminated images.
- Training: The model was initialized with ImageNet weights and fine-tuned using Huber loss and stochastic gradient descent.
- Interpretability: Heat map analysis (convolutional visualization layers) was used to identify which regions of the fundus the model prioritized.
Results
The model demonstrated strong performance in estimating both parameters:
- Axial Length (AL): Achieved a Mean Absolute Error (MAE) of 0.56 mm ($r^2 = 0.59$).
- Choroidal Thickness (SFCT): Achieved an MAE of 49.20 µm ($r^2 = 0.62$).
Bland–Altman plots showed a mean difference of only -0.16 mm for AL and -4.40 µm for SFCT, indicating high agreement between the AI's estimates and the "gold standard" hardware measurements. The model was particularly accurate for "normal" eyes (AL between 22–26 mm), with an MAE of 0.50 mm.
Interpretability and Dynamic Attention
The AI likely succeeds because axial elongation causes visible changes in the retina, such as "fundus tessellation" (where choroidal vessels become more visible as the overlying layers thin). The model learns these patterns as proxies for physical depth. Heat map analysis showed the model’s focus shifted based on the eye's anatomy:
- For short eyes (< 22 mm), it looked at the macular region as a whole.
- For long eyes (> 26 mm), it focused on the extrafoveal regions.
- For SFCT, it consistently prioritized the central fovea and perifovea.
Discussion and Limitations
While promising, the study has limitations. The dataset consisted entirely of Chinese participants aged 50 and older, so results may not generalize to younger populations or other ethnic groups. Additionally, the model tended to overestimate AL in very short eyes and underestimate SFCT in very thick choroids, likely due to the relative scarcity of these extreme cases in the training set.
Future iterations should include more diverse datasets and explore data augmentation to improve performance at the extremes. This could eventually lead to a "semiautomatic assessment" tool integrated into portable fundus cameras or even smartphones, enabling low-cost, accessible eye screenings.
Summary updated