Reduced Retinal Vascular Density and Skeleton Length in Amblyopia

Abstract

Purpose: This study aimed to investigate the possible relationship between retinal vascular abnormalities and amblyopia by analyzing vascular structures of fundus images.

Methods: In this observational study, retinal fundus images were collected from 36 patients with unilateral amblyopia, 33 patients with bilateral amblyopia, and 36 healthy control volunteers. We developed a customized training algorithm based on U-Net to digitalize the vasculature in the fundus images to quantify vascular density (area and fractal dimension), skeleton length, and number of bifurcation points. For statistical comparisons, this study divided participants into two groups. The amblyopic eyes and the fellow eyes of patients with unilateral amblyopia formed the paired group, while bilateral amblyopic patients and healthy controls formed the independent group.

Results: In the paired group, the vascular area (P = 0.007), vascular fractal dimension (P = 0.007), and vascular skeleton length (P = 0.002) of the amblyopic eyes were significantly smaller than those of the fellow eyes. In the independent group, significant decreases in the vascular fractal dimension (P = 0.006) and skeleton length (P = 0.048) were observed in bilateral amblyopia compared to control. The vascular area was also significantly correlated with best-corrected visual acuity in amblyopic eyes.

Conclusions: This study demonstrated that retinal vascular density and skeleton length in amblyopic eyes were significantly smaller compared to control, indicating an association between the changes in retinal vascular features and the state of amblyopia.

Translational Relevance: Our algorithm presents amblyopic retinal vascular changes that are more biologically interpretable for both clinicians and researchers.

Introduction

Amblyopia, also known as lazy eye, is characterized by a reduction of vision in one or both eyes due to a failure of normal neural development in the immature visual system.^1,2 It is the most common cause of monocular visual impairment in both children and young adults, affecting 1% to 5% of the population.^3–5 This condition can cause irreversible vision loss if not detected, diagnosed, and treated in childhood.^6,7 Although early detection and intervention are crucial factors in the successful treatment of amblyopia, the most commonly used strategies for screening and clinical evaluation worldwide are orthoptic evaluations (visual acuity assessment, stereoacuity, and ocular alignment) and photoscreening.⁸ These strategies provide a comprehensive picture of visual function in amblyopic eyes. However, the structural changes of the retina and blood vessels have not been examined during such clinical evaluations. 

Recent studies using advanced imaging techniques such as optical coherence tomography angiography (OCTA) have demonstrated that amblyopia not only is a maldevelopment in the visual cortex but also causes structural changes in the blood vessels. As reported by Nishikawa et al.,⁹ in unilateral amblyopic eyes, reduced macular vessel density and thicker inner retinal layers were found compared with the contralateral eyes. Lonngi et al.¹⁰ and Liu et al.¹¹ used OCTA to study abnormal superficial and deep retinal capillary density in the macula of patients with amblyopia and found that the macular vessel density was lower in the amblyopic group than in the control group. These studies have shown that children with amblyopic eyes have attenuated macular and choriocapillaris perfusion, which can lead to reduced vision.¹² On the other hand, a meta-analysis showed that the microvascular density in the parafovea was lower in amblyopic eyes compared to healthy control eyes, but no significant difference was found in the microvascular density, foveal avascular zone area, or foveal thickness between amblyopic eyes and healthy control eyes.¹³ Although these insights into retinal structural changes in amblyopic eyes are valuable, it remains unclear how the overall retinal vasculature is altered in amblyopia. Therefore, our research aimed to provide potential clues to amblyopia through the analysis of changes in retinal vascular features. The changes in the retinal vasculature may have the potential to assist in a comprehensive diagnostic system for amblyopia. 

Fundus imaging offers a holistic view of the retina and provides detailed structural information that complements OCTA data. Although OCTA excels in visualizing retinal blood flow, depth-resolved imaging, angiographic analysis, and quantitative measurements, fundus imaging provides a widefield view of the retina, allowing for a comprehensive evaluation of the entire retina.¹⁴ It captures detailed images of the optic nerve head, macula, and peripheral retina, making it useful for screening, monitoring, and diagnosing various retinal conditions. In addition, fundus imaging is a non-invasive, relatively cost-efficient and quick procedure that can be performed in a routine clinical setting.¹⁵ 

Classification methods of retinal vessel segmentation have proven to be useful in providing retinal information for relevant disease detection, but efficient methods are necessary to extract and analyze vascular information from retinal fundus images with cloudlike textures specifically for amblyopic eyes.¹⁶ In this study, to find discrepancies between the normal and amblyopic retinal vascular structure, fundus photographs of amblyopic patients were analyzed using an algorithm developed based on the U-Net training algorithm. Our algorithm automatically extracted the complete vasculature from fundus images and produced retinal vascular data that could be analyzed to provide biologically interpretable results. 

Although early detection and intervention are critical, we acknowledge that amblyopia is often not diagnosed until later in childhood. Therefore, the focus of the current study was to investigate potential retinal vascular indications of amblyopia. Our study does not offer information suitable for widespread screening or early detection. Nonetheless, delving into analytical methods such as retinal vasculature analysis can reveal the link between retinal abnormalities and amblyopia, which may be useful in developing a diagnostic framework for the condition in future studies. 

Methods

Retinal Fundus Image Dataset and Preprocessing

The study was approved by the Ethics Committee of Shenzhen Eye Hospital (k20211109-07) and followed the tenets of the Declaration of Helsinki. Consent was obtained from all participants or their guardians. The retinal photographs were collected from 33 young patients with bilateral anisometropic amblyopia (BA), 36 patients with unilateral anisometropic amblyopia (UA), and 36 young healthy control (HC) volunteers. Among the patients with UA, 33 of them had amblyopic left eyes and three of them had amblyopic right eyes. The criteria for diagnosing anisometropic amblyopia was anisometropia with an interocular spherical equivalent difference of at least 0.5 diopters (D) or cylinder difference of at least 1.0 D and best-corrected visual acuity (BCVA) of the amblyopic eye equivalent to 0.05 to 0.8 logarithm of the minimum angle of resolution (logMAR).^17–19 All amblyopic patients received prior spectacle or occlusion treatments. The visual acuity was performed using a standard logarithmic visual acuity chart. Compound tropincaramide with mydriatic optometry was used to measure the refractive state. The stereo acuity of the amblyopic patients was not measured. The HC volunteers had normal corrected visual acuity for both eyes, without other ocular disease, surgery, or neurological disorders. Detailed clinical data are summarized in Table 1. Raw images were acquired using a ZEISS VISUCAM 200 (Carl Zeiss Meditec, Jena, Germany) with 45° viewing angle for all participants. All raw images were clipped to square and scaled to the same size. The images were then enhanced with the contrast-limited adaptive histogram equalization (CLAHE) algorithm. 

Table 1.

 

Clinical Characteristics

 

Many image processing methods and deep learning algorithms have been developed that are dedicated to improving the accuracy and efficiency of retinal vessel segmentation.^16,20,21 However, the confounding influence of cloudlike texture exists in the retinal fundus images of amblyopic teenagers, so few studies have provided valid results for retinal vasculature analyses. Therefore, this study combined three different methods that produced promising results to obtain an accurate binary representation of retinal vasculature by eliminating the influence of cloudlike texture observed in teenage retinal fundus images. 

Retinal Vessel Segmentation

Vessel segmentation recognizes vasculature in the original retinal fundus images and produces binary images of blood vessels without other unnecessary retinal information. Because the calculation of vascular features is based on the binary images of blood vessels, vessel segmentation is a key step in retinal feature extraction. The three commonly used image processing methods for vessel segmentation are based on vascular tracking, morphological processing, and deep learning. Three specific vessel segmentation methods were used in this work: weighted line detector,²² COSFIRE filters,²³ and a U-Net–based convolutional neural network.²⁴ Weighted line detector and COSFIRE filters were two of the best morphological processing-based algorithms at the time, and U-Net was the most commonly used neural network algorithm in vessel segmentation. 

For normal fundus images, all three methods exhibited good performance; however, some images from amblyopic eyes showed cloudlike textures that were incorrectly identified as vessels (Fig. 1A). To improve the quality of blood vessel extraction, some typical images were selected as training samples for the U-Net neural network algorithm. Overall, 20 images from the DRIVE open dataset and 10 images with cloudlike textures from our data were used to train the U-Net model so that the cloudlike textures would not be recognized as blood vessels. 

Figure 1.

 

Example images of retinal vessel segmentation. (A) Left: A sample image with cloudlike textures. Middle: Vessel segmentation by U-Net training on the DRIVE dataset, in which blurry white spots can be observed around the cloudlike texture. Right: Vessel segmentation by U-Net training on samples picked from images set to be processed, in which the white spots disappeared. (B) Better results from merged outputs of the three vessel segmentation methods.

The ground truth vessel images were the expected output of the training samples and were prepared through the following steps: (1) open a sample picture in the image processing tool, (2) add a transparent masked layer, (3) mark the vessels manually on the layer, and (4) export the masked layer as the ground truth for training. After training, the network automatically learned to differentiate cloudlike textures from actual blood vessels. Based on these new training data, additional enhanced training was performed on all of the samples. 

As traditional image processing methods take time to adjust arguments or to add extra processing steps with no guaranteed improvement, the model for deep learning algorithms can automatically learn from datasets with particular features. In Figure 1B, the U-Net model lacks fine recognition of vascular ends, the weighted line detector model does not recognize enough details, and the COSFIRE model has a poor recognition effect on the optic disc and macular area. As a result, comparison of the outputs of these vessel segmentation methods showed that each one contained some imperfect details. The annotated results are binary images, but the predicted results of the model are probabilities of each pixel being part of a blood vessel. By trying multiple groups of threshold values (0.31, 0.32, 0.33, …, 0.70) and comparing them with the annotated results in the training set, a binary vessel image was generated based on a proper threshold value. For all three methods, the output value range was 0.0 to ∼1.0, corresponding to the confidence of each pixel within vessels. The error between the predicted and the annotated binary images was evaluated using the area under the curve (AUC) metric. The predictions from the three methods were merged by assigning different weighted sums to each of the methods to find the best result. A set of weights and a threshold that yielded the best AUC result were selected as the default values. Additionally, a semi-automated tool was developed for final observations and manual adjustments to obtain a better outcome. Most of the binary images were generated using default values, but a small number of images were manually adjusted. The final image was obtained after denoise processing (removing small blobs; see Fig. 1B). 

Calculation of Vascular Density-Related Features

Two features were employed to quantify vascular density: vascular area and fractal dimension. To evaluate fundus images of different sizes, all of the images were scaled to the same size with a 624 × 624 resolution. The vascular area was calculated based on the number of pixels within vessels. The vascular fractal dimension measures the distribution of vessel density by reflecting the distribution patterns of vascular networks.²⁵ The wider a vascular network spreads, the larger the fractal dimension for this network. The binary vessel image was divided into n × n small squares, and the number of blocks with vessel points was counted. As n took values from 10 to 100, a group of x (number of total blocks) and y (number of vessel blocks) was collected. Performing a linear regression on log(x) and log(y), the slope of the regression line was defined as the fractal dimension (Fig. 2). Detailed calculation methods can be found in our previous article.²⁶ 

Figure 2.

 

Calculation of vascular fractal dimension. Left: Vessel image is divided into 25 × 25 blocks; Right: Scatter points corresponding to n = 10 ∼ 100. The x-axis is the logarithm of number of total blocks; the y-axis is the logarithm of number of vessel blocks.

Calculation of Vascular Skeleton Length and Its Bifurcation Points

Overall vascular skeleton length and the number of bifurcation points for the skeletonized vessel tree were also employed as vascular features. To calculate the vascular skeleton length, the vessel caliber width was reduced to 1 by applying a binary blob thinning operation and turning the vessel image into a skeletonized form, forming a skeleton tree (Fig. 3, middle). The number of white pixels corresponds to the overall vascular skeleton length. The bifurcation points on the vascular skeleton tree split the tree into small branches (Fig. 3, right). The number of bifurcation points could be an indicator of the morphological richness of the vessel tree. 

Figure 3.

 

Illustration of calculation of skeleton length and bifurcation points. Left: Vessel segmentation image. Middle: Vascular skeleton tree. Right: Marked bifurcation points on the skeleton tree (green points).

Statistical Analysis

The vascular features on two groups of fundus data were analyzed using SPSS Statistics 19 (IBM, Chicago, IL): (1) the paired group for within-subjects comparison and (2) the independent group for between-subjects comparison. In the paired group, 36 patients with UA were recruited for the pairing tests. The Wilcoxon signed-rank test was used to compare vascular features between fellow eyes and amblyopic eyes of the same patients with UA. In this study, the term “fellow eyes” refers only to the eyes with normal corrected vision in unilateral amblyopic patients. In the independent group, 36 HCs and 33 patients with BA were also recruited for the collection of fundus images from both eyes. Hence, 72 HC images and 66 amblyopic images from patients with BA were used in the independent group. To analyze the difference in independent samples, the Mann–Whitney U test was used to compare the vascular area between HCs and patients with BA. We performed t-tests on the vascular features of fellow eyes of patients with UA and HCs. In addition, the Kendall correlation coefficients were calculated to analyze the potential correlation between vascular features and the best-corrected visual acuity of the amblyopic eyes from patients with UA and patients with BA. 

Results

Vascular Density Analysis

Vascular Area

A non-parametric sample pairing test was performed on fundus data from patients with UA between their fellow eyes and amblyopic eyes, in which mean rank (MR) values of each data point were compared. The vascular area decreased significantly in amblyopic eyes compared to that in fellow eyes (MR_A<H = 18.70, MR_A>H = 17.89; n_UA = 36; Z = –2.702; P = 0.007), as shown in Figure 4A. In the independent group, a trend, but no statistical significance, was observed for a decrease in the vascular area in patients with BA (MR_BA = 65.33) compared to HCs (MR_HC = 73.33; n_BA = 66, n_HC = 72; U = 2100.5; P = 0.240), as shown in Figure 4B. 

Figure 4.

 

Comparison of data distribution of the vascular area. (A) Comparison of 36 patients with UA showed a significant decrease in amblyopic eyes compared to fellow eyes in vascular area. (B) Comparison between 72 fundus images of 36 healthy participants and 66 images of 33 patients with BA showed no statistically significant decrease in BA. **P < 0.01. Data are presented as medians (center lines) and quartiles (boxes) in violin plots in which the shaded areas represent the data distribution.

Vascular Fractal Dimension

The fractal dimension demonstrates the space-filling capacity of the blood vessel system on the fundus. The results on paired eyes showed that the vascular fractal dimension decreased significantly from fellow eyes to amblyopic eyes (MR_A<H = 18.74, MR_A>H = 17.78; n_UA = 36; Z = −2.718; P = 0.007) (Fig. 5A). The absolute mean rank value of the fractal dimension of the HCs was 78.50, and that of patients with BA was 59.68. The vascular fractal dimension was significantly lower in the fundi of patients with BA compared to HC (n_BA = 66, n_HC = 72; U = 1728; P = 0.006) in the independent group (Fig. 5B). 

Figure 5.

 

Comparison of data distribution of the vascular fractal dimension. (A) Comparison of 36 patients with UA showed a significant decrease in amblyopic eyes compared to fellow eyes in vascular fractal dimension. (B) Comparison between 72 HC images and 66 BA images showed a significant falling tendency in BA compared to HC in vascular fractal dimension. **P < 0.01. Data are presented as medians (center lines) and quartiles (boxes) in violin plots in which the shaded areas represent data distribution.

Results of Vascular Skeleton Length and Its Bifurcation Points

Vascular Skeleton Length

A significant decrease of vascular skeleton length in the amblyopic eyes was observed compared to the fellow eyes in the paired group (MR_A<H = 21.04, MR_A>H = 12.73; n_UA = 36; Z = −3.032; P = 0.002) (Fig. 6A). Meanwhile, the independent test scoring 1913 (P = 0.048) demonstrated that healthy participants (n_HC = 72; MR = 75.93) had significantly higher vascular skeleton length compared with patients with BA (n_AD = 66; MR = 62.48) (Fig. 6B). 

Figure 6.

 

Comparison of data distribution of the vascular skeleton length. (A) Comparison of 36 patients with UA showed a significant reduction in amblyopic eyes compared to fellow eyes in vascular skeleton length. (B) Comparison between 72 HC images and 66 BA images confirmed the consistent tendency of a decrease of vascular skeleton length in BA compared to HC. *P < 0.05; **P < 0.01. Data are presented as medians (center lines) and quartiles (boxes) in violin plots in which the shaded areas represent data distribution.

Number of Bifurcation Points

The results of the paired group did not find any significant difference between amblyopic eyes and fellow eyes of patients with UA (MR_A<H = 18.95, MR_A>H = 16.38; n_UA = 36; Z = −1.671; P = 0.095), as shown in Figure 7A. Meanwhile, the results of the independent group (n_BA = 66, MR_BA = 69.42; n_HC = 72, MR_HC = 69.58; U = 2370.5; P = 0.981) also did not show changes with statistical significance, as shown in Figure 7B. We also analyzed the differences in vascular features between the healthy (right) eyes of patients with UA and the right eyes of HCs and did not find any significant difference. The results of the statistical tests can be found in Table 2. 

Figure 7.

 

Comparison of data distribution of the number of bifurcation points. No significant results were found in Wilcoxon signed-rank tests for UA (A) or Mann–Whitney U test between HC and BA (B). Data are presented as medians (center lines) and quartiles (boxes) in violin plots in which the shaded areas represent data distribution.

Table 2.

 

Comparisons Between Fellow Eyes of Patients With UA and HCs

 

Correlation Analysis

The correlation analysis aimed to investigate the association between retinal vascular features and the decline in BCVA among individuals with amblyopia. Because visual acuity is measured in logMAR units, visual acuity decreases when the number increases. Specifically, the vascular area (r = −0.121, P = 0.039) showed a significant correlation with the decline in visual acuity among all of the amblyopic eyes (Fig. 8A). However, the vascular fractal dimension (r = −0.093, P = 0.244), skeleton length (r = −0.053, P = 0.398), and number of bifurcation points (r = −0.018, P = 0.373) did not exhibit significant correlation with visual acuity. 

Figure 8.

 

Correlation between retinal vascular features and visual acuity. (A) The vascular area was significantly correlated with visual acuity. (B–D) The vascular fractal dimension (B), vascular skeleton length (C), and number of bifurcation points (D) were not significantly correlated with visual acuity.

Discussion

Our study analyzed multiple retinal vascular features in amblyopia based on retinal fundus images, including retinal vascular density (vascular area and fractal dimension), vascular skeleton length, and its bifurcation points. We demonstrated that the retinal vascular density and vascular skeleton length of amblyopic eyes were significantly smaller under both paired and unpaired conditions compared to healthy controls. Although the difference in the number of bifurcation points was not as significant as the other three features, our results indicated an overall reduction in the values of vascular features of amblyopic eyes compared to the HCs. Thus, not only is the macular microvascular structure altered in amblyopic eyes,^9,10 but also the overall retinal vascular features in the posterior segment are reduced. The involvement of both the macular microvascular structure and the overall retinal vasculature in the posterior segment indicates a widespread vascular remodeling throughout the retina under amblyopic conditions. This information provides an indication of the retinal vessel difference which could potentially be used for amblyopia screening in future research. 

The decreased vascular density suggests a diminished coverage of blood vessels within the retina, potentially compromising the delivery of oxygen and nutrients to retinal tissues. Subasi et al.²⁷ found that amblyopic children suffered from a deficiency of phosphorus, selenium, molybdenum, iodine, chromium, boron, beryllium, folate, and vitamin B12. On top of such nutritional deficiency, decreased contrast sensitivity and impaired visual acuity may exacerbate the visual deficits observed in amblyopia.²⁸ Similarly, a decrease in the vascular skeleton length suggests a less elaborate and organized vasculature that may disrupt the intricate network of retinal blood vessels that also play a crucial role in delivering oxygen and nutrients to retinal cells. Our findings indicate that amblyopia is not solely a disorder of visual acuity but also involves alterations in the retinal vasculature. This broader involvement of the retinal vasculature suggests that vascular remodeling is not limited to a specific region but extends throughout the retina. 

A decrease in retinal vessel density was found in amblyopia compared to controls in our current study, which is consistent with previous studies.²⁹ One important note is that the data in our study were from post-treatment patients. Interestingly, Gunzenhauser et al.³⁰ and Huang et al.³¹ both showed an increase in retinal vessel density in post-treatment amblyopic children compared to pre-treatment. Taken together, these results indicate a decrease in fundus vascular density in the amblyopia state itself. In addition, the correlation analysis suggests a potential link between retinal vascular area and the deterioration of visual acuity in individuals with amblyopia. These findings indicate that changes in retinal vasculature are closely related to changes and visual deficits in the amblyopic state, which improves our understanding of the pathogenesis of amblyopia. Future studies may examine the correlation between the progression or treatment of amblyopia and retinal structural changes during different stages. Although it cannot yet claim clinical application nor replace traditional diagnostic strategies, our study aimed to focus on vascular density and other related vascular features to offer a more objective perspective for primary eyecare providers. 

One limitation of our work is that the generalizability of our U-Net model has not been proven with multiple datasets from different medical centers. The model was effective in recognizing retinal blood vessels, but our data regarding retinal fundus images were collected from Shenzhen Eye Hospital only. To test its practical generalizability, we hope that it can implemented by other research entities in future studies involving the analysis of teenagers with amblyopia. Another limitation is that the differences between eyes or groups do not provide information that can be directly used for large-scale screening or diagnosis. Our study is based on a small population of amblyopia and does not conclusively determine whether a child is amblyopic based on these measurements. Also, we analyzed only a limited number of vascular features, and the potential of other features as indicators of amblyopia remained unexplored. 

Overall, the decreases in retinal vessel density and vascular skeleton length observed in amblyopia indicate underlying structural and functional changes within the retinal vasculature. This study shed light on the association between amblyopia and retinal vascular changes, specifically the decreases in retinal vessel density and vascular skeleton length, which are also correlated with the decrease in visual acuity. These alterations in the retinal vasculature can have important implications for visual deficits experienced by individuals with amblyopia and may contribute to our understanding of the pathophysiology of amblyopia.³² Further work to follow up our current study includes elucidating the underlying mechanisms involved and investigating the practical application of retinal vascular features in the diagnostic system. 

Acknowledgments

Supported by grants from the Ministry of Science and Technology (2022ZD0208604, 2022ZD0208605 to JZ; 2022ZD0210000 to BY; 81970790 to XY), National Natural Science Foundation of China (31771195, 81790640, 81773513, 82071200 to JZ; 32100803 to BY), Shanghai Municipal Science and Technology Major Project (2018SHZDZX01 to JZ), Key Research and Development Program of Ningxia (2022BEG02046 to JZ), Sanming Project of Medicine in Shenzhen (SZSM202011015 to JZ and XY), Key Scientific Technological Innovation Research Project of the Ministry of Education (JZ), and Shenzhen Science and Technology Program (SGDX20211123120001001, JCYJ2020010914500181 to XY). 

Disclosure: W. Su, None; L. Ma, None; K. Li, None; Y. Hu, None; Y. Mao, None; W. Xie, None; X. Hu, None; T. Huang, None; J. Lv, None; M. Wang, None; B. Yan, None; X. Yao, None; X. Yan, None; J. Zhang, None