The Associations Between Myopia and Fundus Tessellation in School Children: A Comparative Analysis of Macular and Peripapillary Regions Using Deep Learning

Abstract

Purpose: To evaluate the refractive differences among school-aged children with macular or peripapillary fundus tessellation (FT) distribution patterns, using fundus tessellation density (FTD) quantified by deep learning (DL) technology.

Methods: The cross-sectional study included 1942 school children aged six to 15 years, undergoing ocular biometric parameters, cycloplegic refraction, and fundus photography. FTD was quantified for both the macular (6 mm) and peripapillary (4 mm) regions, using DL-based image processing applied to 45° color fundus photographs. Eyes exhibiting tessellation were classified into two groups: the macular distribution group had greater FTD in the macular area, while the peripapillary distribution group had higher FTD in the peripapillary area, allowing for a comparative analysis of axial length (AL), corneal radius, and refraction.

Results: Participants had a median age of 13 years and a median spherical equivalent (SE) of −0.75 D. The macular distribution group exhibited significantly larger AL (24.13 mm vs. 23.93 mm, P < 0.001) and more myopic refraction (−1.13 D vs. −0.75 D, P < 0.001) compared to the peripapillary group. A higher prevalence of macular-distributed FT was noted in the myopic groups (χ² = 131.675, P < 0.001). SE negatively correlated with macular (r = −0.238) and peripapillary FTD (r = −0.195), while AL positively correlated with FTD in both regions (r = 0.308; r = 0.265) (all P < 0.001).

Conclusions: The macular FT distribution pattern is significantly associated with larger AL and greater myopic refraction in school-aged children, suggesting its potential as a marker for identifying children at risk of progressing myopia.

Translational Relevance: DL analysis precisely identifies FT distribution patterns, potentially enhancing early detection of high-risk myopia in populations.

Introduction

Fundus tessellation (FT) is a prevalent myopic fundus anomaly and represents a primary category of myopic macular degeneration.^1,2 Characterized by the heightened visibility of large choroidal vessels in the fundus because of thinning of the retinal pigment epithelium and choroid, FT is usually observed by fundus photographs during myopia management for children, and the severity of FT is used clinically as an ancillary indicator of myopia progression and myopic retinopathy.^1–3 

Recent Asian studies, where trained clinicians manually assess fundus FT, indicate that the prevalence of FT stands at 42.18% in children aged seven years, escalates to 52.4% among those aged nine to 12 years with low myopia, and peaks at 94.3% among children and adolescents presenting with high myopia.^4–6 In addition, Yamashita et al.⁷ found that residents of Kumejima who were older than 40 years and who had FT in the posterior pole and macular locations showed longer axial length (AL), compared with those without FT. The authors conclude that the FT distribution pattern is valuable in both interpreting and predicting myopic changes. With the increasing global incidence of myopia and its association with FT, a profound exploration into the distribution patterns and associative parameters of FT in pediatric populations, is imperative.⁸ 

In the current study, we aimed to evaluate the refractive differences among school-aged children with macular or peripapillary FT distribution patterns. Accordingly, we quantified “fundus tessellation density” (FTD), using deep learning (DL) image processing technology that measures the average exposed choroidal area per unit area of the fundus. This metric provides a precise quantification of FT, facilitating an analysis of how FT distribution correlates with myopia in pediatric populations, thereby informing more targeted management strategies in screening and clinical settings. 

Methods

Study Population

This retrospective study included two sample groups: (1) The Nanjing group was hospital-based, which enrolled school children aged six to 15 years old from the pediatric ophthalmology clinic of the First Affiliated Hospital with Nanjing Medical University from July 2021 to September 2022, Jiangsu Province, urban east China. (2) The Mojiang group was school-based and a part of the Mojiang Myopia Progression Study.⁹ It invited all grade 7 Hani students from middle schools in Mojiang Hani Autonomous County, Yunnan Province, rural southwest China. The Hani, an ethnic minority group genetically close to the Han ethnic group, is one of the 56 officially recognized nationalities in China.¹⁰ The data from Mojiang Myopia Progression Study was collected in 2016. For both groups, we excluded children with these conditions: astigmatism ≤−3.00 D; anisometropia ≥2.50 D; hyperopia ≥3.00 D; strabismus; amblyopia; history of previous intraocular surgery; history or evidence of chorioretinal or vitreoretinal disease; history of systemic diseases. 

The study was approved by the Ethics Committee of the First Affiliated Hospital with Nanjing Medical University (2021-SR-125) and Kunming Medical University (201502), respectively. It followed the tenets of the Declaration of Helsinki. Written informed consent was obtained from the legal guardians of all children, and oral consent was also obtained before the examination. 

Eye Examinations

Eye examinations included visual acuity, ocular surface, anterior segment, posterior segment, refraction, ocular alignment and motility, ocular biometric parameters, and fundus photographs. The AL and corneal radius of curvature (CR) were obtained and calculated with the LenStar LS900 (Haag-Streit, Bern, Switzerland). The CR was computed by averaging the measurements from the flat and steep meridians, which aids in computing the axial length/corneal radius of curvature ratio (AL/CR ratio). 

In the Mojiang group, each participant was administered two drops of 1% cyclopentolate (Alcon) after a five-minute interval for cycloplegia. Thirty minutes later, a third drop was administered if the pupillary light reflex was still present or if the pupil size was less than 6.0 mm. Then optometrists measured the refractive status after cycloplegia using an autorefractor (RM-8000; Topcon, Tokyo, Japan). A fundus camera (Weiqing LUNA, Suzhou, China) was used to capture 45° color fundus images, including both macular and peripapillary areas of both eyes after cycloplegia. 

In the Nanjing group, cycloplegic refraction was obtained using retinoscopy. Cycloplegia was induced with 0.5% tropicamide eye drops (Handankangye, Handan, China). The eye drops were administered every five minutes, totaling six applications. After the last administration, a 20-minute waiting period was observed before retinoscopy was conducted to ensure full cycloplegic effect. The fundus images were also obtained after cycloplegia (Canon CR-2; Canon Inc., Tokyo, Japan). 

Analysis of Fundus Photographs

The quality of fundus images was assessed according to the following criteria: (1) Focus: the image is focused exactly on the retina and the retinal features can be clearly identified; (2) Illumination: the image is properly exposed, neither overexposed (the image is low in contrast and has white tones), nor underexposed (the image is low in brightness and difficult to identify retinal features); (3) Image field: the center of the image is located between the macula and the optic disc, and both of the optic disc and macula are at least one optic disc diameter away from the edge of the image; (4) Artifacts: there are no artifacts in the image that would interfere with the recognition of fundus features, such as dust spot, arc defect, fingerprint, camera reflex, eyelash image, and so on. Fundus images that did not meet the above criteria were identified as low quality and excluded from analysis. 

The signal strength of image data was evaluated by a fundus image quality assessment system, which consists of evaluations of exposure level and clarity of fundus images.¹¹ More technical details are described in supplementary materials. All fundus images that passed the quality control in both the processes of the image acquisition and manual FT grading had also passed the signal strength evaluation. 

Then, AI image processing technology based on deep learning, as reported in previous studies,^4,12–15 was used to extract exposed choroid from fundus images and obtain the average exposed choroid area per unit area of the fundus, named FTD. Briefly, this process is composed of preprocessing, sample labeling, feature recognition and segmentation by deep learning and computer vision methods, and FTD calculation. 

First, fundus images were preprocessed in four steps to improve the differences between the internal features of each image and reduce the differences between images 1: establishing a region of interest (ROI), denoising, normalization, and enhancement. ROI establishment involved extracting the effective region in the fundus image and eliminating invalid areas such as the background. Channel separation was performed on the color fundus image, where the background in the red channel is dark. The ROI candidate region was obtained by segmenting the red channel image using a threshold segmentation method based on the average gray value and the area ratio of the dark area. Finally, the ROI was established after evaluating the morphological characteristics and location of the ROI candidate region. Denoising was performed to decrease noise interference during photography and imaging. The image was converted from the spatial domain to the frequency domain via low-pass filtering, and then removed the low-frequency component. Normalization was conducted to adjust the color, brightness, and size of each image to a consistent range using average calibration, decreasing differences between images. Brightness normalization was accomplished by transforming the image from RGB space to LAB space, calibrating the average value of the L space, and transforming the image back to RGB space. Size normalization involved resizing each image to a uniform size. Enhancement was achieved using the Contrast-Limited Adaptive Histogram Equalization algorithm on the ROI region. It divides the image into different small blocks and applies gray-scale restriction improvement processing to various small blocks of the image, followed by gray-scale interpolation between adjacent blocks to decrease gray-scale differences at the boundaries.¹⁶ 

Second, sample labeling was performed through channel calculation, manual modification, and review. This process consisted of automatic labeling and manual correction. Automatic sample labeling was achieved by channel separation and channel subtraction of the color fundus image to obtain initial labels of the exposed choroid regions. The initial labeled samples were then manually modified and reviewed by two experts to obtain the final labeled samples. One expert performed the initial modification, and another expert conducted the review, reducing subjectivity and enhancing the accuracy and consistency of the annotations. 

Third, a semantic segmentation network model based on deep learning was used to extract exposed choroid. Specifically, we used the ResNet-FCN (Fully Convolutional Network) model as the training model with the labeled samples as training data. The ResNet component introduces residual connections to effectively solve the problems of gradient vanishing and gradient explosion in deep neural networks, allowing the network to learn image features more deeply. The FCN component achieves pixel-level classification of input images through a fully convolutional design. The combination of ResNet and FCN's high-level feature extraction capabilities enables more accurate capture of image details and contextual information, which is particularly effective for segmenting the irregular and randomly distributed patterns of fundus tessellation. We first extracted high-level features through ResNet18 and then performed deconvolution to obtain the segmentation area, outputting the tessellation (leopard spot) confidence map and obtaining the confidence probability of each pixel belonging to exposed choroid. After the model was trained to the optimal effect, it was used as an inference model for subsequent exposed choroid extraction. 

The optic disc and macula were automatically recognized using a visual attention mechanism to locate the optic disc center and macular fovea. This mechanism simulates the human visual feature filtering process by selecting features in the regions of interest to remove irrelevant areas and highlight areas of concern. Brightness was used as the primary feature for localization. The image was processed through filtering and Laplace transformation to obtain brightness feature distribution results, where the optic disc and macula appear as extremely bright and extremely dark regions, respectively. By identifying the extreme brightness points, the optic disc was localized based on high-brightness extrema, and the macula was localized based on low-brightness extrema.¹¹ 

Finally, based on the results of fundus feature recognition and segmentation, FTD was calculated for the whole fundus, the macular area (a circle with a diameter of 6 mm centered on the fovea), and the peripapillary area (a circle with a diameter of 4 mm centered on the optic disc).¹² FTD is defined as the average exposed choroid area per unit area of the fundus and calculated using the formula: ρ = S1/S, where ρ is the FTD, S1 is the area of the exposed choroid, and S is the area of the preprocessed ROI region. The use of fixed region sizes provides statistical uniformity and stability in FTD measurements. Although individual anatomical variations exist, fixed region definitions allow for consistent comparisons across participants. Masking of retinal vessels was not performed, as it could introduce uncertainties and affect the stability of FTD calculations. 

Classification of the Distribution of Fundus Tessellation

Fundus photographs were stratified into three groups based on the distribution characteristics of FT (Fig. 1). Photographs exhibiting the macular distribution pattern, which is defined as a greater FTD within the macular area (6mm) compared to the peripapillary area (4mm), were classified as the macular distribution group. Conversely, photographs demonstrating a higher FTD in the peripapillary area were assigned to the peripapillary distribution group. Additionally, photographs where both macular and peripapillary FTD equated to zero were categorized into the no tessellation group, indicating an absence of tessellation. 

Figure 1.

 

Flowchart for classifying FTD patterns. (A) An eye in the no Tessellation group; (B) An eye in the macular distributed group; (C) An eye in the peripapillary distributed group. (A1, B1, C1) Original image. (A2, B2, C2) Fundus tessellation in macular area (6 mm, yellow circle). (A3, B3, C3) Fundus tessellation in peripapillary area (4 mm, purple circle).

Definitions

The spherical equivalent (SE) was calculated as the spherical power plus half of the cylinder power. Mild hyperopia was defined as 1.00 D<SE<3.00 D. Emmetropia was defined as −0.50 D<E ≤ 1.00 D; mild myopia was defined as −3.00 D<SE ≤ −0.50 D; moderate myopia was defined as −6.00 D<SE ≤ −3.00 D. High myopia was defined as SE ≤ −6.00 D, according to the guideline from International Myopia Institute.¹⁷ 

Statistical Analysis

Data were analyzed using IBM's Statistical Package for the Social Sciences (SPSS) software, version 13.0 (Chicago, IL, USA). P values were two-sided, and values < 0.05 were considered indicative of statistical significance, with 95% confidence intervals (CIs) provided. Nonnormal continuous variables were represented as medians and quartiles. Categorical measures were described using frequency counts and percentages. Analysis exclusively incorporated data from the right eye of each participant. All analyze were performed using data from Nanjing and Mojiang groups separately, as well as in combination. 

The Mann-Whitney U test was used to evaluate differences between the Nanjing and Mojiang groups. The Kruskal-Wallis H test was used to discern ocular feature variances across three distinct FT distribution patterns. The chi-square tests were conducted to compare the percentages of sex distribution and FT patterns, whereas Fisher's exact test was applied in instances of expected values less than 5. Then, Spearman correlation analysis was used to examine relationships between continuous measures with SE, AL, and FTD. 

Results

General Characteristics

As shown in Table 1, a total of 1942 people were enrolled in this study, including 993 boys (51.13%) and 949 girls (48.87%), with a median age of 13 years old, median SE of −0.75 D, median AL of 23.86 mm and median AL/CR ratio of 3.05. The hospital-based Nanjing Group (N = 1048) was younger (9.00 vs. 14.00 years old), more myopic (−1.25 vs. 0.25 D), with longer AL (24.22 vs. 23.51 mm), larger CR (7.81 vs. 7.79 mm), larger AL/CR ratio (3.09 vs. 3.01), larger macular FTD (0.004 vs. 0.003) and larger peripapillary FTD (0.003 vs. 0.000) than the school-based Mojiang group (all P < 0.05). There was no statistical difference in sex between the two groups (P = 0.826). 

Table 1.

 

General Characteristics and Comparison Between Different Distribution Groups of Fundus Tessellation

 

Comparison of Ocular Biometric and Refraction Characteristics Between the Macular and Peripapillary FT Distribution Patterns

Of the 1942 children, 39.55% (768/1942) were classified as macular pattern, 34.19% (664/1942) were classified as peripapillary pattern, and the remaining 26.26% (510/1942) had no FT (Table 2). Table 2 shows the comparison of parameters between macular and peripapillary distribution patterns. The median values of eyes with FT macular distribution pattern show lower SE, longer AL, and larger ALCR ratio compared to eyes with FT peripapillary distribution pattern across all groups: total children (−1.13 vs. −0.75 D; 24.13 vs. 23.93 mm; 3.08 vs. 3.06), the hospital-based Nanjing group (−1.75 vs. −1.13 D; 24.47 vs. 24.19 mm; 3.11 vs. 3.09), and the school-based Mojiang group (0.13 vs. 0.38 D; 23.79 vs. 23.51 mm; 3.03 vs. 3.00) (Fig. 2, Table 2, all P < 0.05). 

Table 2.

 

General Characteristics and Comparison Between Different Distribution Patterns of Fundus Tessellation

 

Figure 2.

 

(A) Comparison of axial length by fundus tessellation distribution pattern. (B) Comparison of spherical equivalent by fundus tessellation distribution pattern.

The distribution patterns of FT vary across refractive groups, with the myopic groups exhibiting a higher prevalence of macular-distributed FT (χ² = 131.675, P < 0.001, Fig. 3) compared to the hyperopia/emmetropia group. For total children, the proportion of the macular-distributed FT was not statistically different from that of the peripapillary pattern in the hyperopic or emmetropic children (N = 840, 31.9% vs. 29.9%), mild myopic children (N = 859, 41.9% vs. 39.1%), or high myopic children (N = 22, 63.6% vs. 36.4%) (all P > 0.05). However, in moderate myopic children, there were more children with macular FT pattern than peripapillary FT pattern (N = 221, 57.0% vs. 31.2%, P < 0.05). 

Figure 3.

 

Fraction of fundus tessellation distribution pattern by refractive status: (A) Total. (B) The hospital-based Nanjing Group. (C) The school-based Mojiang group.

Correlation Analysis of AL, SE, and the percentage of FTD

Based on the analysis of all children included, SE was negatively correlated with AL (r = −0.665), macular FTD (r = −0.238), peripapillary FTD (r = −0.195), the percentage of FTD of macular area / whole image (r = −0.323) and the percentage of FTD of peripapillary area/whole image (r = −0.210). (Table 3, all P < 0.001). The macular FT group has higher correlation coefficients in all analyses: FTD with SE and AL, and the percentage of FTD with SE and AL, compared to the peripapillary FT group. 

Table 3.

 

Spearman's Correlation Analysis of Spherical Equivalent, Axial Length, and the Percentage of Fundus Tessellated Density Among 1942 School Children

 

Discussion

In this study of school-aged children, our findings indicate that those exhibiting a macular FT pattern not only presented with longer AL and higher AL/CR ratios but also more pronounced myopic refraction, compared to their peers with a peripapillary FT pattern. Additionally, our analysis revealed distinct variations in FT distribution across different degrees of myopia. These findings suggest that the macular region may play a pivotal role in the manifestation of myopia, highlighting the potential of macular FT as a marker for identifying children at risk of myopia progression. 

Our study included two distinct cohorts: the hospital-based Nanjing group and the school-based Mojiang group, which differed significantly in age, refractive status, and ethnicity (Table 1). The Nanjing group was younger (median age: 9 years) and had more myopic refractive errors (median SE: −1.25), while the Mojiang group was older (median age: 14 years) with less myopic refraction (median SE: 0.25 D). Despite these differences, the associations between FT distribution patterns and ocular biometric parameters were almost consistent across both groups (Tables 2, 3). 

In both cohorts, children with the macular FT pattern exhibited longer AL and more myopic SE compared to those with the peripapillary FT pattern. However, the strength of these associations varied slightly between the groups. For instance, the correlation between macular FTD and AL was slightly stronger in the Mojiang group (r = 0.311, P < 0.001) than in the Nanjing group (r = 0.299, P < 0.001) (Table 3). In addition, the significant association between FTD ratio of the respective region versus whole image in the population combining the two cohorts is only observed for the macular distribution group instead of the peripapillary distribution group. These variations may be attributed to differences in refractive status, age, environmental factors, or genetic backgrounds between the two populations. 

The correlation between the spatial distribution of FT and AL has been explored in a cross-sectional, population-based study focusing on middle-aged and senior individuals.⁷ This study undertook a manual classification of FT, revealing that individuals within the “no tessellation” category exhibited a significantly shorter AL compared to those identified in the posterior pole and macular groups, suggesting that FT distribution patterns are valuable in interpreting myopic changes.⁷ Similarly, our preceding study used AI based FTD quantification found that children with FT macular pattern tend to have more pronounced myopic refraction, although the statistical significance was marginal due to limitations in non-cycloplegic refraction measurements.¹² By using cycloplegic refraction in the current study, we have reinforced the evidence supporting the significant refractive differences associated with FT patterns. 

Previous research has explored the morphological changes of the eyeball during the progression of myopia, proposing various models of ocular expansion, including posterior pole, equatorial, global, and axial expansions.^18–20 It's worth noting that the influence of visual cues on scleral expansion is remarkably localized.²¹ Research in animal models has uncovered that retinal growth exhibits a localized characteristic, with distinct regions of the retina independently expanding according to their specific retinal blur signals.^22–24 In alignment with these findings, we observed that different degrees of myopia appear to be intricately linked with changes in the spatial distribution of FT. Emmetropic and mildly myopic eyes predominantly exhibited fundus tessellation (FT) in the peripapillary area. However, as myopia progressed to the moderate stage, a distinct shift in FT distribution towards the macular area became apparent (refer to Fig. 3). Similarly, Hsiang et al.²⁵ found that the macular type was the most prominent type (52.7%) of staphylomas in high myopes. This shift may reflect localized scleral remodeling and differential ocular growth patterns in the macular region associated with increasing myopia severity. Supporting this notion, Varadarajan et al.²⁶ also found that fundus characteristics at the foveal region in the fundus images can be used to predict refractive error in middle-aged and older adults in the United Kingdom and the United States, using a deep learning algorithm. 

The correlation analysis further supports the significance of macular FT in relation to myopia severity. We found stronger correlations between macular FTD and both AL and SE compared to peripapillary FTD (Table 3). This suggests that the macular region may be more susceptible to the biomechanical forces associated with axial elongation. 

Our findings have potential screening and clinical implications. The ability to quantify and analyze FT distribution patterns using DL technology offers a noninvasive and accessible method for early identification of children at risk for myopia progression. Incorporating macular FT assessment into routine screening could enhance the precision of myopia management strategies, allowing for timely interventions to slow down myopia progression and prevent high myopia and its associated complications at the population level. 

Moreover, the integration of AI-based FTD analysis into screening and clinical practice represents an advancement in ophthalmic diagnostics. Automated and objective measurements reduce inter-observer variability and increase the efficiency of screenings, which is valuable in large-scale public health initiatives aimed at addressing the rising prevalence of myopia globally. 

This study’s merits are underscored by its extensive sample size of a school-aged population, the application of cycloplegic refraction, and the use of FTD, an objective metric powered by a ResNet-FCN deep learning model, which has been previously validated in different populations and achieved high segmentation accuracies.^13,27 

Despite these strengths, areas for enhancement remain. The study's retrospective cross-sectional design constrains the longitudinal analysis possible for AL, SE, and FT changes. This limitation precludes a detailed analysis of how refractive differences evolve over time among varied FT patterns. Moreover, there are discrepancies in age, refractive status, cycloplegic drops and fundus camera models between the Nanjing group and Mojiang group. To address this, data were presented both collectively and separately for the two groups, yielding consistent results that bolster the overall reliability of the findings. Another limitation is the lack of extensive external validation of our deep learning model on entirely independent datasets. Although our model has shown consistent high accuracy in previous studies, we acknowledge that deep learning models might exhibit reduced performance on external test sets because of variations in data distribution and imaging conditions.²⁸ Future studies should include comprehensive external validation using diverse datasets from different regions and imaging devices to further assess and enhance the model's generalizability. Finally, we did not account for other potential confounding factors such as genetic predisposition and environmental influences. Including these variables in future studies could provide a more comprehensive understanding of the associations. 

Notably, the conclusion should be drawn carefully by considering disease history and differential diagnosis for certain retinopathy and vitreoretinopathy, such as retinopathy of prematurity and familial exudative vitreoretinopathy, which involves pathological fundus changes that may affect FT patterns. The use of AI-powered FTD for objective FT classification enhances the reliability of our analysis, offering an accessible promising tool for future myopia evaluations. Future research should therefore not only focus on refining these AI models to include more detailed fundus characteristics for greater accuracy and broader applicability but also on exploring the integration of such technologies into routine screening and clinical practice. 

Conclusions

Assessing FT patterns, particularly in the macular region, may aid clinicians in identifying children at risk of myopia. The integration of AI-based FTD into clinical practice represents a promising tool, offering a nuanced approach to understanding and managing myopia. Future research should focus on refining these AI models and exploring their integration into routine screening and clinical applications. 

Acknowledgments

Authors thank all the members of the Maternal and Child Healthcare Hospital of Yuhuatai District, Nanjing, China, for their helpful advice and support. 

Supported by the National Natural Science Foundation of China (Grant No.82273159; No. 82003475), Jiangsu Province's Science and Technology Project (Grant No. BE2020722) and Scientific Research Project of Jiangsu Province Association of Maternal and Child Health (Grant No. FYX202340); the Specialized Diseases Clinical Research Fund of Jiangsu Province Hospital (Grant No. DL202405). The funding organizations had no role in the design or conduct of this research. 

Disclosure: D. Huang, None; X. Lin, None; H. Zhu, None; S. Ling, None; Z. Dong, None; X. Ke, None; T. Long, None; Y. Qian, None; Q. Yan, None; R. Li, None; H. Zhong, None; H. Liu, None