Abstract
Introduction
This study investigated the effect of axial length (AL) on peripheral retinal vessel density in children and adolescents and assessed whether deep learning can detect early vascular changes in myopia.
Methods
Non-mydriatic ultra-widefield imaging was used to capture retinal images. Deep learning models based on Nested U-Net and ResNet34 segmented the vasculature, quantified vessel density in 60–30° and 100–60° fields, and classified AL from fundus images.
Results
A total of 679 eyes from 396 children and adolescents were analyzed. Participants were categorized into normal (22.96 ± 0.65 mm), medium (24.69 ± 0.50 mm), and high (27.32 ± 0.80 mm) AL groups. Across both 60–30° and 100–60° fields, the temporal retina displayed higher vessel density, while the inferior retina showed lower density. The normal AL group had significantly higher density than the medium AL group (P < 0.05), which surpassed the high AL group (P < 0.05). In the 60–30° temporal region, vessel density decreased from 7.15 ± 1.17% (normal) to 6.70 ± 1.27% (medium) and 6.16 ± 1.82% (high). Deep learning classification achieved an AUC of 0.9651, with Grad-CAM highlighting the inferotemporal vasculature.
Conclusions
As AL increases, peripheral vessel density diminishes. This pattern may suggest a potential prioritization of blood flow to the macular region, although longitudinal studies are required to confirm this hypothesis. These findings suggest that deep learning analysis of ultra-widefield images can reveal subclinical vascular changes, offering a potential tool for early detection of high myopia risk.
Keywords: Myopia, Ultra-widefield imaging, Deep learning, Retinal vessel density, Axial length
Key Summary Points
| This study quantified peripheral retinal vessel density in children and adolescents using non-mydriatic ultra-widefield (UWF) imaging and deep learning. | |
| A significant decrease in peripheral retinal vessel density was observed as axial length (AL) increased across both 60–30° and 100–60° fields. | |
| The deep learning model (ResNet34) demonstrated high diagnostic performance (AUC 0.9651) in classifying myopia severity based on fundus vasculature. | |
| The temporal retina exhibited the highest vessel density, while the inferior region showed the most significant reduction in high myopia. | |
| Automated deep learning analysis of the peripheral vasculature provides a potential screening tool for identifying subclinical vascular changes and early high myopia risk. |
Introduction
Worldwide, refractive errors are primarily caused by myopia, a prevalent ocular condition characterized by excessive elongation of the axial length (AL) [1, 2]. Refractive errors are not the only consequences of this elongation; structural changes in the eye, such as thinning of the choroid and retina, are also induced, potentially leading to blinding complications [3, 4]. Chorioretinal atrophy has been associated with increased AL exceeding 24 mm [5, 6], and choroidal vascularity and choriocapillaris perfusion are affected. The rising prevalence of myopia, especially among adolescents in China, is attributed to genetic factors and environmental influences such as increased screen time and reduced outdoor activities [1, 7]. AL growth spurts are more prominent between the ages of 6 and 9 years, slowing after age 12 [8], making adolescence a critical period for eye development. The groundwork for high myopia in adulthood is laid by this rapid expansion, underscoring the necessity for regular AL monitoring during this period. As AL increases, several ocular changes are observed, including a pronounced correlation between chorioretinal atrophy and AL growth, as well as implications on choroidal vascularity and choriocapillaris perfusion. A prior study formulated a deep learning model for predicting AL in eyes utilizing ultra-widefield (UWF) fundus images [9], but the analysis and comparison of peripheral retinal blood vessel density were not explored [10, 11].
Despite extensive research on myopia, peripheral retinal vascular density has not been a primary focus. Understanding changes in the peripheral retinal vasculature associated with axial elongation could have significant implications for the prevention or prediction of peripheral retinal degeneration and retinal breaks in patients with myopia. Investigating this aspect may provide valuable insights into the pathogenesis of myopic complications and guide early interventions.
Advancements in big data analytics and deep learning models have shown great promise in the diagnosis and management of retinal diseases [12, 13]. Deep learning algorithms, such as Convolutional Neural Networks (CNNs), have been successfully applied to predict ocular conditions and reconstruct three-dimensional structures from imaging data [14, 15]. In deep learning, particularly in the field of medical imaging, understanding how models make their decisions is paramount for validation and trust in clinical applications. Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps have emerged as a powerful tool in this context, offering visual explanations for the decisions made by CNNs [12].
In this study, a CNN-based model is utilized to analyze ultra-widefield retinal images and investigate the relationship between axial length and peripheral retinal vascular density in adolescents. By doing so, valuable insights may be offered for preventive strategies and control measures against myopia in this age group.
Methods
Patients
This was a prospective observational study. Children and adolescents in China were recruited between July 2022 and June 2024. Ethical approval for this study was obtained from the First Affiliated Hospital of Chongqing Medical University (approval number 2022-039). The study was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from the parents or legal guardians of all participants because all participants were under the age of 18. Consent for publication was not required as the images and data presented in this study are anonymized and contain no personally identifiable information. Basic visual acuity and optometry tests were conducted on the children. Measurements of the axial length and corneal curvature were conducted using the AL-Scan (NIDEK Co., Ltd., Japan). Inclusion criteria included (1) age between 6 and 18 years; (2) absence of obvious ocular diseases, including but not limited to corneal opacities, cataracts, glaucoma, uveitis, or any history of intraocular surgery. Exclusion criteria included (1) presence of retinopathy; (2) inability to cooperate, such as in cases of nystagmus; (3) images that did not meet quality standards, such as inability to identify the macula or optic disc, overexposure, or images where the center deviates from the macula or the optic disc is too far away. The enrolled eyes were categorized into the high-AL group (AL ≥ 26 mm), medium-AL group (24 mm ≤ AL < 26 mm), and normal-AL group (AL < 24 mm).
Ultra-Widefield Fundus Image Acquisition
Ultra-widefield laser scanning retinal angiography was used to image the retina. Images were acquired using the Apollo Ultra-Widefield Confocal Scanning Laser Ophthalmoscope (Suzhou Microclear Medical Instrument Co., Ltd., Suzhou, China). In an undilated state, two images were captured for each patient in confocal mode, centering on the macula and the optic nerve. These images were stitched into a single image using the manufacturer’s integrated image processing system. This system employs built-in geometric correction algorithms to minimize peripheral distortion and provides the pixel reference values for the 30°, 60°, and 100° fields of view on the image.
Retinal Image Analysis Based on Deep Learning Model
Deep learning-based analysis of retinal images was performed as shown in Fig. 1a. A trained Nested U-Net was employed to identify vascular pixels in the retinal images. On the basis of the results of this partitioning, the vessel density was calculated for each region and used for subsequent analysis.
Fig. 1.
Process of calculating vascular density and classification of axial lengths. a Calculation process via a Nested U-Net model. b Convolutional Neural Network (CNN) with a ResNet34 backbone, trained to categorize high, medium, and normal axial lengths (AL) qualitatively
The positions of the macula and optic disc were used to divide the retinal periphery into distinct regions for analysis. Specifically, as shown in Fig. 2a, subregions 1–4 are located between 100° and 60° from the optic disc, corresponding to the nasal, superior, temporal, and inferior sides, respectively [16, 17]. Subregions 5–8 lie between 60° and 30° from the optic disc, also corresponding to the nasal, superior, temporal, and inferior sides, respectively. As shown in Fig. 2b, the two main regions are centered around the macula and optic disc: region 9 lies between 30° and 60°, and region 10 lies between 60° and 100°. The CNN with a ResNet34 backbone, as demonstrated in Fig. 1b, was trained to qualitatively categorize images into high, medium, and normal axial lengths. The performance of the classification model was assessed via receiver operating characteristic (ROC) curve analysis, with the AUC serving as the metric. The Gradient-weighted Class Activation Mapping (Grad-CAM) heatmap was utilized for interpreting the areas of interest in the CNN model.
Fig. 2.
Partitioning process for retinal periphery. a Division into eight subregions within 100–60° and 60–30° of the optic disc. Subregions 1–4 correspond to the nasal, superior, temporal, and inferior sides, respectively, while subregions 5–8 mirror these alignments. b Two main regions, centered around the macula and optic disc: region 9 is within 60–30° and region 10 is within 100–60°
Statistical Methods
The basic characteristics of the dataset were described using descriptive statistics. Differences between the high-AL, medium-AL, and normal-AL groups were compared using Tamhane’s test. Differences in retinal vessel density between different regions were tested using a paired sample t test. Data conforming to a normal distribution were expressed as mean ± standard deviation. All statistical analyses were performed using SPSS (Version 26, IBM Corp., Armonk, NY, USA), and a p value < 0.05 was deemed statistically significant.
Results
Baseline Characteristics
The AL and retinal images from 679 eyes of 396 children and adolescents were collected, and the vessel density of segmented retinal vessels was calculated. The baseline data for all eyes and groups are shown in Table 1. The axial lengths of the normal-AL, medium-AL, and high-AL groups were 22.96 ± 0.65 mm, 24.69 ± 0.50 mm, and 27.32 ± 0.80 mm, respectively. The spherical equivalents were − 0.63 ± 1.17 D, − 2.81 ± 1.92 D, and − 8.47 ± 2.45 D, respectively.
Table 1.
Baseline data of all eyes in three axial length groups
| Total (679 eyes) |
Normal-AL group (499 eyes) |
Medium-AL group (152 eyes) |
High-AL group (28 eyes) |
F | P | |
|---|---|---|---|---|---|---|
| AGE | 11.44 ± 3.33 | 10.53 ± 3.03 | 13.42 ± 2.54 | 17.04 ± 1.99 | 113.18 | < 0.001 |
| SPHER | − 1.26 ± 2.06 | − 0.53 ± 1.10 | − 2.49 ± 1.81 | − 7.64 ± 2.27 | 446.49 | < 0.001 |
| CYLIN | − 0.37 ± 0.72 | − 0.22 ± 0.60 | − 0.63 ± 0.65 | − 1.65 ± 1.09 | 82.38 | < 0.001 |
| SE | − 1.44 ± 2.24 | − 0.63 ± 1.17 | − 2.81 ± 1.92 | − 8.47 ± 2.45 | 478.65 | < 0.001 |
| AL | 23.52 ± 1.23 | 22.96 ± 0.65 | 24.69 ± 0.50 | 27.32 ± 0.80 | 990.64 | < 0.001 |
| K1 | 42.67 ± 1.42 | 42.91 ± 1.37 | 42.06 ± 1.34 | 41.67 ± 1.35 | 31.09 | < 0.001 |
| K2 | 43.87 ± 1.57 | 44.07 ± 1.53 | 43.32 ± 1.55 | 43.44 ± 1.64 | 15.02 | < 0.001 |
AGE = age in years; SPHER = spherical refractive error in diopters; CYLIN = cylindrical refractive error in diopters; SE = spherical equivalent refractive error in diopters; AL = axial length in millimeters; K1 and K2 = keratometry readings (corneal curvature) in diopters for the flat and steep meridians, respectively. Values are presented as mean ± standard deviation. F and P represent the F-statistic and associated p value from one-way analysis of variance (ANOVA) comparing the three axial length groups (normal-AL group, medium-AL group, high-AL group). A p value less than 0.05 indicates a statistically significant difference among the groups
Table 2 presents the gender-based comparison of baseline characteristics across the different AL groups. In the total sample, male adolescents exhibited significantly longer AL than female adolescents (23.82 ± 1.21 mm vs. 23.25 ± 1.20 mm, F = 38.37, P < 0.001). Additionally, male adolescents had flatter corneal curvature values, as indicated by lower K1 and K2 measurements (K1, 42.18 ± 1.37 D vs. 43.13 ± 1.30 D; K2, 43.32 ± 1.53 D vs. 44.39 ± 1.43 D; both F > 85, P < 0.001). No significant gender differences were observed in age, spherical refractive error (SPHER), cylindrical refractive error, or spherical equivalent in the total group.
Table 2.
Gender-based comparison of baseline characteristics across different axial length groups
| Total | Female (n = 352) | Male (n = 327) | F | P |
|---|---|---|---|---|
| AGE | 11.46 ± 3.30 | 11.43 ± 3.37 | 0.02 | 0.882 |
| SPHER | − 1.25 ± 2.00 | − 1.27 ± 2.13 | 0.01 | 0.904 |
| CYLIN | − 0.39 ± 0.76 | − 0.34 ± 0.67 | 0.81 | 0.369 |
| SE | − 1.45 ± 2.17 | − 1.44 ± 2.32 | 0.00 | 0.973 |
| AL | 23.25 ± 1.20 | 23.82 ± 1.21 | 38.37 | < 0.001 |
| K1 | 43.13 ± 1.30 | 42.18 ± 1.37 | 85.21 | < 0.001 |
| K2 | 44.39 ± 1.43 | 43.32 ± 1.53 | 88.35 | < 0.001 |
| Normal-AL group | Female (n = 278) | Male (n = 221) | F | P |
|---|---|---|---|---|
| AGE | 10.75 ± 3.01 | 10.25 ± 3.03 | 3.31 | 0.069 |
| SPHER | − 0.57 ± 1.09 | − 0.47 ± 1.13 | 0.93 | 0.336 |
| CYLIN | − 0.27 ± 0.67 | − 0.15 ± 0.51 | 5.06 | 0.025 |
| SE | − 0.70 ± 1.16 | − 0.55 ± 1.19 | 2.20 | 0.138 |
| AL | 22.76 ± 0.67 | 23.21 ± 0.53 | 66.34 | < 0.001 |
| K1 | 43.34 ± 1.26 | 42.39 ± 1.31 | 67.29 | < 0.001 |
| K2 | 44.55 ± 1.40 | 43.46 ± 1.47 | 71.93 | < 0.001 |
| Medium-AL group | Female (n = 61) | Male (n = 91) | F | P |
|---|---|---|---|---|
| AGE | 13.59 ± 2.78 | 13.31 ± 2.36 | 0.45 | 0.502 |
| SPHER | − 3.04 ± 1.97 | − 2.12 ± 1.61 | 9.97 | 0.001 |
| CYLIN | − 0.67 ± 0.70 | − 0.61 ± 0.62 | 0.36 | 0.550 |
| SE | − 3.38 ± 2.11 | − 2.42 ± 1.69 | 9.46 | 0.002 |
| AL | 24.72 ± 0.45 | 24.67 ± 0.53 | 0.34 | 0.563 |
| K1 | 42.41 ± 1.18 | 41.82 ± 1.38 | 7.42 | 0.007 |
| K2 | 43.73 ± 1.41 | 43.04 ± 1.58 | 7.67 | 0.006 |
| High-AL group | Female (n = 13) | Male (n = 15) | F | P |
|---|---|---|---|---|
| AGE | 16.77 ± 2.35 | 17.27 ± 1.67 | 0.43 | 0.519 |
| SPHER | − 7.42 ± 1.19 | − 7.83 ± 2.94 | 0.22 | 0.642 |
| CYLIN | − 1.69 ± 1.19 | − 1.62 ± 1.04 | 0.03 | 0.858 |
| SE | − 8.27 ± 1.39 | − 8.64 ± 3.14 | 0.16 | 0.696 |
| AL | 26.84 ± 0.33 | 27.74 ± 0.86 | 12.55 | 0.001 |
| K1 | 42.07 ± 1.03 | 41.33 ± 1.52 | 2.24 | 0.146 |
| K2 | 43.94 ± 1.18 | 43.02 ± 1.89 | 2.32 | 0.140 |
AGE = age in years; SPHER = spherical refractive error in diopters; CYLIN = cylindrical refractive error in diopters; SE = spherical equivalent refractive error in diopters; AL = axial length in millimeters; K1 and K2 = keratometry readings (corneal curvature) in diopters for the flat and steep meridians, respectively. Values are presented as mean ± standard deviation. F and P represent the F-statistic and associated p value from one-way analysis of variance (ANOVA) comparing the three axial length groups (normal-AL group, medium-AL group, high-AL group). A p value less than 0.05 indicates a statistically significant difference among the groups
Peripheral Retinal Vessel Density: Interregional Comparison
Table 3 displays the results of paired-sample t tests comparing mean vascular densities across regions. In the 60–30° field of view, the temporal side (TD60-30) exhibited the highest vascular density, while the inferior side (ID60-30) showed the lowest. The difference between the superior (SD60-30) and nasal (ND60-30) regions was statistically significant (P < 0.001), whereas the differences between TD60-30 and SD60-30, and between ND60-30 and ID60-30 were not significant (P = 0.291; P = 0.354). Additionally, in the 100–60° field of view, the temporal side (TD100-60) demonstrated the highest vascular density, and the inferior side (ID100-60) exhibited the lowest. Significant differences were observed among the regions in this field of view (P < 0.001). Notably, the vascular density in the 100–60° field was lower than that in the 60–30° field (P < 0.001).
Table 3.
Comparison of retinal vessel density between regions using paired sample t test
| Region 1 | Region 2 | Mean difference (95% LoA) | t | P value (t) | |
|---|---|---|---|---|---|
| ND60-30 vs SD60-30 | 7.01 ± 0.58 | 7.90 ± 0.74 | − 0.89 (− 2.06 to 0.28) | − 38.79 | 0.000 |
| ND60-30 vs TD60-30 | 7.01 ± 0.58 | 7.93 ± 0.68 | − 0.93 (− 2.35 to 0.49) | − 33.29 | 0.000 |
| ND60-30 vs ID60-30 | 7.01 ± 0.58 | 6.98 ± 0.65 | 0.02 (− 1.16 to 1.21) | 0.93 | 0.354 |
| SD60-30 vs TD60-30 | 7.90 ± 0.74 | 7.93 ± 0.68 | − 0.03 (− 1.72 to 1.65) | − 1.06 | 0.291 |
| SD60-30 vs ID60-30 | 7.90 ± 0.74 | 6.98 ± 0.65 | 0.91 (− 0.25 to 2.08) | 40.07 | 0.000 |
| TD60-30 vs ID60-30 | 7.93 ± 0.68 | 6.98 ± 0.65 | 0.95 (− 0.65 to 2.54) | 30.36 | 0.000 |
| ND100-60 vs SD100-60 | 5.59 ± 0.42 | 5.02 ± 0.41 | 0.57 (− 0.30 to 1.44) | 33.36 | 0.000 |
| ND100-60 vs TD100-60 | 5.59 ± 0.42 | 7.53 ± 0.56 | − 1.94 (− 2.96 to − 0.92) | − 97.19 | 0.000 |
| ND100-60 vs ID100-60 | 5.59 ± 0.42 | 4.45 ± 0.42 | 1.14 (0.29 to 1.99) | 68.89 | 0.000 |
| SD100-60 vs TD100-60 | 5.02 ± 0.41 | 7.53 ± 0.56 | − 2.51 (− 3.58 to − 1.43) | − 119.02 | 0.000 |
| SD100-60 vs ID100-60 | 5.02 ± 0.41 | 4.45 ± 0.42 | 0.57 (− 0.29 to 1.43) | 33.98 | 0.000 |
| TD100-60 vs ID100-60 | 7.53 ± 0.56 | 4.45 ± 0.42 | 3.08 (2.04 to 4.12) | 151.84 | 0.000 |
| MDP100-60 vs MDP60-30 | 4.98 ± 0.27 | 7.55 ± 0.47 | − 2.57 (− 3.25 to − 1.89) | − 193.29 | 0.000 |
LoA = limit of agreement; t = t value in paired t test
MDP60-30: macula-disc peripheral, 60–30°
MDP100-60: macula-disc peripheral, 100–60°
ND100-60: nasal disc, 100–60°
SD100-60: superior disc, 100–60°
TD100-60: temporal disc, 100–60°
ID100-60: inferior disc, 100–60°
ND60-30: nasal disc, 60–30°
SD60-30: superior disc, 60–30°
TD60-30: temporal disc, 60–30°
Peripheral Retinal Vessel Density: Intergroup Comparison
Table 4 presents the comparison of retinal peripheral vessel density among the three AL groups using analysis of covariance (ANCOVA), controlling for age. The results indicate that AL significantly affects vascular density in all regions except for region 3 (TD60-30), with P values less than 0.05. Across all regions, age did not have a significant independent effect on vascular density, as evidenced by P values greater than 0.05.
Table 4.
Differences in retinal vascular density across three groups assessed with ANCOVA
| Total (679 eyes) |
Normal-AL group (499 eyes) |
Medium-AL group (152 eyes) |
High-AL group (28 eyes) |
AL | Age | |||
|---|---|---|---|---|---|---|---|---|
| F | P | F | P | |||||
| Region 1: ND60-30 | 7.01 ± 0.58 | 7.14 ± 0.49 | 6.72 ± 0.58 | 6.09 ± 0.67 | 174.55 | < 0.001 | 0.53 | 0.466 |
| Region 2: SD60-30 | 7.90 ± 0.74 | 8.11 ± 0.61 | 7.47 ± 0.66 | 6.51 ± 0.70 | 259.51 | < 0.001 | 0.98 | 0.323 |
| Region 3: TD60-30 | 7.93 ± 0.68 | 7.94 ± 0.68 | 7.93 ± 0.62 | 7.83 ± 0.89 | 0.08 | 0.784 | 0.36 | 0.550 |
| Region 4: ID60-30 | 6.98 ± 0.65 | 7.13 ± 0.58 | 6.66 ± 0.59 | 6.09 ± 0.81 | 174.65 | < 0.001 | 0.28 | 0.595 |
| Region 5: ND100-60 | 5.59 ± 0.42 | 5.63 ± 0.41 | 5.56 ± 0.41 | 5.18 ± 0.60 | 14.54 | < 0.001 | 0.01 | 0.936 |
| Region 6: SD100-60 | 5.02 ± 0.41 | 5.04 ± 0.40 | 4.98 ± 0.41 | 4.90 ± 0.63 | 7.87 | 0.005 | 0.03 | 0.858 |
| Region 7: TD100-60 | 7.53 ± 0.56 | 7.68 ± 0.48 | 7.23 ± 0.45 | 6.48 ± 0.63 | 226.72 | < 0.001 | 0.08 | 0.772 |
| Region 8: ID100-60 | 4.45 ± 0.42 | 4.50 ± 0.39 | 4.35 ± 0.41 | 4.19 ± 0.67 | 34.23 | < 0.001 | 1.07 | 0.299 |
| Region 9: MDP60-30 | 7.55 ± 0.47 | 7.69 ± 0.38 | 7.26 ± 0.42 | 6.70 ± 0.54 | 313.94 | < 0.001 | 0.19 | 0.663 |
| Region 10: MDP100-60 | 4.98 ± 0.27 | 5.01 ± 0.26 | 4.93 ± 0.26 | 4.78 ± 0.45 | 28.04 | < 0.001 | 0.57 | 0.451 |
MDP60-30: macula-disc peripheral, 60–30°
MDP100-60: macula-disc peripheral, 100–60°
ND100-60: nasal disc, 100–60°
SD100-60: superior disc, 100–60°
TD100-60: temporal disc, 100–60°
ID100-60: inferior disc, 100–60°
ND60-30: nasal disc, 60–30°
SD60-30: superior disc, 60–30°
TD60-30: temporal disc, 60–30°
Performance Evaluation of the Deep Learning Model
Figure 3 displays the ROC curve of the classification model, which was trained over 91 epochs. The model achieved a training loss of 0.0942 with an accuracy of 97.67%, and a validation loss of 0.2101 with an accuracy of 92.25%. Additionally, the model exhibited an average AUC of 0.9651 on the test dataset.
Fig. 3.
ROC curve of deep learning model for classification of axial length
Grad-CAM Heatmap Analysis
Figure 4 shows heatmaps for eyes with normal axial length (a, b), medium axial length (c, d), and high axial length (e, f). As depicted, the temporal and inferior vessels of the optic disc were the focus of the deep learning model when determining the axial length group of a retinal image.
Fig. 4.
Heatmaps of retinal images for normal-AL, medium-AL, and high-AL eyes, highlighting the areas of interest. a, b Heatmaps of two eyes with normal AL, corresponding to right and left eyes with axial lengths of 21.53 mm and 22.48 mm, respectively. c, d Heatmaps of two eyes with medium AL, corresponding to right and left eyes with axial lengths of 24.25 mm and 24.14 mm, respectively. e, f Heatmaps of two eyes with high AL, corresponding to right and left eyes with axial lengths of 26.84 mm and 26.58 mm, respectively. As shown, the CNN model paid more attention to the inferior temporal region of the retina
Discussion
In this study, deep learning models were utilized to analyze peripheral retinal vessel density in children and adolescents using non-mydriatic ultra-widefield retinal imaging. Characteristic changes in the peripheral retinal vasculature during myopia progression were identified, suggesting that these vascular alterations may serve as novel early predictive markers for myopia-related retinal pathologies. The use of deep learning in assessing peripheral retinal vessels provides a promising tool for early detection and intervention, potentially improving preventive strategies against myopic retinal diseases.
This study revealed that, in the early stages of axial elongation, peripheral retinal vessel density significantly decreases, particularly in the inferior and temporal regions. In the mid-peripheral retina, vessel density decreases significantly in all directions except for the temporal region, where the reduction is less pronounced. Notably, these areas of reduced vascular density correspond to regions where high myopia-related peripheral retinal lesions, such as lattice degeneration and retinal breaks, are commonly observed [16–18]. This coincidence suggests that the decreased peripheral retinal vessel density associated with axial elongation may lead to sustained nutritional deficits in the peripheral retina. Such deficits could contribute to the development of peripheral retinal changes seen in high myopia. Therefore, changes in peripheral retinal vessel density might serve as sensitive indicators of axial differences. Increased attention should be paid to the density and morphological changes of peripheral retinal vessels during the early progression of myopia, using them as indicators to predict myopia progression. These results suggest that clinicians should prioritize monitoring the peripheral and mid-peripheral retinal regions with significantly reduced vascular density during follow-up of patients with myopia, even when no overt fundus abnormalities are observed.
It was found that the distribution of retinal blood vessels is uneven; vessel density decreases towards the periphery. Specifically, compared with other directions, the temporal side of the optic disc has a higher blood vessel density, while the inferior side has a lower blood vessel density. This may be attributed to the fact that the macula, which is responsible for crucial visual functions, is encompassed within the TD60-30 region. The macula is a region of high metabolic demand and requires abundant oxygen and nutrient supply, hence necessitating blood perfusion [19, 20]. Although axial elongation leads to reduced vascular density in peripheral regions, the macular blood supply remains relatively stable [21, 22]. We hypothesize that this may represent a potential compensatory mechanism during axial elongation that prioritizes blood flow to the macula. However, given the cross-sectional nature of this study, causality cannot be established. Longitudinal studies are needed to confirm whether a true redistribution of blood flow occurs during the progression of myopia.
In this study, the Grad-CAM technique was employed to explain the decision-making process of the deep learning models [23]. It was shown that the models particularly focused on the inferotemporal vascular region when classifying axial length, further validating the critical role this area plays in axial elongation. It is noteworthy that while our ANCOVA results showed no significant difference in vascular density in the temporal region (region 3, P = 0.784), the deep learning model still identified the inferotemporal vasculature as a key discriminative feature. This discrepancy highlights the distinct capabilities of the two analytical methods. Statistical analysis quantified the absolute reduction in capillary perfusion, whereas the deep learning model (ResNet34) likely extracted high-dimensional morphological features beyond simple density, such as the straightening of the retinal vascular arcades and changes in vessel tortuosity, which are characteristic mechanical responses to axial elongation. Furthermore, the “inferotemporal” region highlighted by Grad-CAM encompasses the vascular arcades extending into the inferior retina (region 4), where our statistical analysis confirmed a highly significant reduction in vessel density (P < 0.001). Thus, the model likely integrates both the morphological stretching in the temporal sector and the density reduction in the inferior sector to classify axial length. This finding suggests that the inferotemporal region may contain key morphological information that reflects axial length changes. Grad-CAM not only enhances model interpretability but also offers clinicians a visual understanding of the model’s focus, increasing confidence in the application of deep learning models in clinical practice.
Grad-CAM provides more than just model interpretability; targeted diagnostic cues are also offered to clinicians. For example, the regions highlighted by the model may guide clinicians in their examination of specific areas of the retina, potentially improving the sensitivity and accuracy of clinical assessments.
Despite the significant findings, this study has some limitations. First, retinal thickness was not considered in the analysis, which may have influenced vascular density distribution. In future studies, retinal thickness should be integrated as a factor to provide a more comprehensive and precise assessment of retinal health. Second, although the overall dataset was substantial, there was an imbalance in sample size across groups. Specifically, the high-AL group included only 28 eyes compared to 499 eyes in the normal-AL group. This relatively small sample size for the high myopia cohort may limit the statistical power and the generalizability of the findings for this specific subgroup. Larger, multicenter studies are needed to validate the generalizability of these results. Future research directions include further optimizing the performance of deep learning models, exploring more biomarkers in relation to vascular density, and developing automated screening tools for clinical use. Such tools would assist clinicians in more efficiently managing patients with high myopia.
Conclusions
This study demonstrates the potential application of deep learning models in managing myopia in adolescents, particularly in detecting retinal vascular changes associated with early axial elongation. The deep learning models not only provided high accuracy in axial length classification but also revealed subtle changes in retinal vessel distribution, which are often undetectable in traditional clinical examinations. By identifying these changes early, more proactive interventions can be taken to slow the progression of axial elongation and reduce the risk of high myopia-related complications, ultimately improving long-term visual outcomes for patients.
Acknowledgements
We thank the participants of the study.
Author Contributions
Yong Wang, Tong Zhang, and Wenjuan Wan were involved in the design and conduct of the research and preparation of the manuscript. Yanlin Yang, Yongguo Xiang, Kexin Sun, Xin Yang, Jiuyi Xia, Shenglan Yi, and Shijie Zheng participated in the collection of the data. Hangjia Zuo analyzed the data. Ke Hu and Wenjuan Wan participated in the management of this program. All authors reviewed and approved the final manuscript.
Funding
This work was supported in part by grants from the National Natural Science Foundation of China (Grant numbers 81870650; 82000883; 81970832), Natural Science Foundation Project of Chongqing (Grant number cstc2021jcyj-msxm3178), the Project Foundation of Chongqing Science and Technology Commission of China (Grant numbers CSTC2021jscx-gksb-N0017; cstc2020jcyj-msxmX0829; cstc2021jcyj-msxmX0967, CSTB2022NSCQ-MSX1561), the First batch of key Disciplines On Public Health in Chongqing, and Chongqing Medical University (Grant number YJSZHYX202118), the National Key Research and Development Program of China (Grant nos. 2020YFC2008200, 2020YFC2008204). The journal’s publication fee was funded by the National Natural Science Foundation of China (Grant number 81870650). The sponsors and funding organizations had no role in the design or conduct of this research.
Data Availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Conflict of Interest
Yong Wang, Tong Zhang, Hangjia Zuo, Yanlin Yang, Yongguo Xiang, Kexin Sun, Xin Yang, Jiuyi Xia, Fanfan Huang, Shenglan Yi, Shijie Zheng, Ke Hu, and Wenjuan Wan have nothing to disclose.
Ethical Approval
This study was performed in accordance with the Helsinki Declaration of 1964 and its later amendments. Ethical approval was obtained from the Ethics Committee of The First Affiliated Hospital of Chongqing Medical University (Approval number 2022-039). Written informed consent to participate in the study was obtained from the parents or legal guardians of all participants because all participants were under the age of 18. Consent for publication was not required as the images and data presented in this study are anonymized and contain no personally identifiable information.
Footnotes
Yong Wang and Tong Zhang contributed equally as co-first authors.
Ke Hu and Wenjuan Wan contributed equally as co-corresponding authors.
Contributor Information
Ke Hu, Email: cqhuke@hospital.cqmu.edu.cn.
Wenjuan Wan, Email: wanwenjuancqums@163.com.
References
- 1.Morgan IG, Ohno-Matsui K, Saw SM. Myopia. Lancet. 2012;379(9827):1739–48. 10.1016/S0140-6736(12)60272-4. [DOI] [PubMed] [Google Scholar]
- 2.Gonzalez Blanco F, Sanz Fernandez JC, Munoz Sanz MA. Axial length, corneal radius, and age of myopia onset. Optom Vis Sci. 2008;85(2):89–96. 10.1097/OPX.0b013e3181622602. [DOI] [PubMed] [Google Scholar]
- 3.Yang J, Ouyang X, Fu H, et al. Advances in biomedical study of the myopia-related signaling pathways and mechanisms. Biomed Pharmacother. 2022;145:112472. 10.1016/j.biopha.2021.112472. [DOI] [PubMed] [Google Scholar]
- 4.Tian F, Zheng D, Zhang J, et al. Choroidal and retinal thickness and axial eye elongation in Chinese junior students. Invest Ophthalmol Vis Sci. 2021;62(9):26. 10.1167/iovs.62.9.26. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Naik GT, Achar P, Dsouza NDA, Beary MSM. Fundus changes in high myopia in relation to axial length of the globe. J Pharm Bioallied Sci. 2022;14(Suppl 1):S649–53. 10.4103/jpbs.jpbs_828_21. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Wu H, Xie Z, Wang P, et al. Differences in retinal and choroidal vasculature and perfusion related to axial length in pediatric anisomyopes. Invest Ophthalmol Vis Sci. 2021;62(9):40. 10.1167/iovs.62.9.40. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Wang W, Xiang Y, Zhu L, et al. Myopia progression and associated factors of refractive status in children and adolescents in Tibet and Chongqing during the COVID-19 pandemic. Front Public Health. 2022;10:993728. 10.3389/fpubh.2022.993728. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Sanz Diez P, Yang LH, Lu MX, Wahl S, Ohlendorf A. Growth curves of myopia-related parameters to clinically monitor the refractive development in Chinese schoolchildren. Graefes Arch Clin Exp Ophthalmol. 2019;257(5):1045–53. 10.1007/s00417-019-04290-6. [DOI] [PubMed] [Google Scholar]
- 9.Oh K, Kang HM, Leem D, Lee H, Seo KY, Yoon S. Early detection of diabetic retinopathy based on deep learning and ultra-wide-field fundus images. Sci Rep. 2021;11(1):1897. 10.1038/s41598-021-81539-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Oh R, Lee EK, Bae K, Park UC, Yu HG, Yoon CK. Deep learning-based prediction of axial length using ultra-widefield fundus photography. Korean J Ophthalmol. 2023;37(2):95–104. 10.3341/kjo.2022.0059. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Dong L, Hu XY, Yan YN, et al. Deep learning-based estimation of axial length and subfoveal choroidal thickness from color fundus photographs. Front Cell Dev Biol. 2021;9:653692. 10.3389/fcell.2021.653692. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Chen TC, Lim WS, Wang VY, et al. Artificial intelligence-assisted early detection of retinitis pigmentosa - the most common inherited retinal degeneration. J Digit Imaging. 2021;34(4):948–58. 10.1007/s10278-021-00479-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Dong Z, Xiao Y, Wei P, Lin L. Decoder-only LLMs are better controllers for diffusion models. In: Proceedings of the 32nd ACM International Conference on Multimedia. New York, NY, USA: Association for Computing Machinery; 2024:10957–10965. 10.1145/3664647.3681639
- 14.Patefield A, Meng Y, Airaldi M, et al. Deep learning using preoperative AS-OCT predicts graft detachment in DMEK. Transl Vis Sci Technol. 2023;12(5):14. 10.1167/tvst.12.5.14. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Zhang T, Huang F, Gao N, et al. Three-dimensional quantitative description of the implantable collamer lens in the ocular anterior segment of patients with myopia. Am J Ophthalmol. 2023;252:59–68. 10.1016/j.ajo.2023.03.005. [DOI] [PubMed] [Google Scholar]
- 16.Maltsev DS, Kulikov AN, Burnasheva MA. Lattice degeneration imaging with optical coherence tomography angiography. J Curr Ophthalmol. 2022;34(3):379–83. 10.4103/joco.joco_94_22. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Nguyen HTT, Hoang TT, Pham CM, Nguyen TM, Dang TM, Fricke TR. Prevalence and related factors of myopic retinopathy - a hospital-based cross-section study in Vietnam. Clin Exp Optom. 2023;106(4):427–30. 10.1080/08164622.2022.2039555. [DOI] [PubMed] [Google Scholar]
- 18.Khatwani N, Makhija S, Ahuja A. Clinical profile and distribution of peripheral retinal changes in myopic population in a hospital-based study in North India. Indian J Ophthalmol. 2022;70(4):1280–5. 10.4103/ijo.IJO_946_21. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Snodderly DM, Weinhaus RS, Choi JC. Neural-vascular relationships in central retina of macaque monkeys (Macaca fascicularis). J Neurosci. 1992;12(4):1169–93. 10.1523/JNEUROSCI.12-04-01169.1992. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Wen C, Pei C, Xu X, Lei J. Influence of axial length on parafoveal and peripapillary metrics from swept source optical coherence tomography angiography. Curr Eye Res. 2019;44(9):980–6. 10.1080/02713683.2019.1607393. [DOI] [PubMed] [Google Scholar]
- 21.Durbin MK, An L, Shemonski ND, et al. Quantification of retinal microvascular density in optical coherence tomographic angiography images in diabetic retinopathy. JAMA Ophthalmol. 2017;135(4):370–6. 10.1001/jamaophthalmol.2017.0080. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Lin T, Yang Y, Lin J, et al. A comparative study of macular and choroidal thickness and blood-flow parameters in patients with intermediate and simple juvenile moderate myopia. Int J Gen Med. 2021;14:1343–8. 10.2147/IJGM.S301702. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Itoh H, Misawa M, Mori Y, Kudo SE, Oda M, Mori K. Positive-gradient-weighted object activation mapping: visual explanation of object detector towards precise colorectal-polyp localisation. Int J Comput Assist Radiol Surg. 2022;17(11):2051–63. 10.1007/s11548-022-02696-y. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.