Fundus Tessellated Density Assessed by Deep Learning in Primary School Children

Abstract

Purpose: To explore associations of fundus tessellated density (FTD) and compare characteristics of different fundus tessellation (FT) distribution patterns, based on artificial intelligence technology using deep learning.

Methods: Comprehensive ocular examinations were conducted in 577 children aged 7 years old from a population-based cross-sectional study, including biometric measurement, refraction, optical coherence tomography angiography, and 45° nonmydriatic fundus photography. FTD was defined as the average exposed choroid area per unit area of the fundus, and obtained by artificial intelligence technology. The distribution of FT was classified into the macular pattern and the peripapillary pattern according to FTD.

Results: The mean FTD was 0.024 ± 0.026 in whole fundus. Multivariate regression analysis showed that greater FTD was significantly correlated with thinner subfoveal choroidal thickness, larger parapapillary atrophy, greater vessel density inside the optic disc, larger vertical diameter of optic disc, thinner retinal nerve fiber layer, and longer distance from optic disc center to macular fovea (all P < 0.05). The peripapillary distributed group had larger parapapillary atrophy (0.052 ± 0.119 vs 0.031 ± 0.072), greater FTD (0.029 ± 0.028 vs 0.015 ± 0.018), thinner subfoveal choroidal thickness (297.66 ± 60.61 vs 315.33 ± 66.46), and thinner retinal thickness (285.55 ± 10.89 vs 288.03 ± 10.31) than the macular distributed group (all P < 0.05).

Conclusions: FTD can be applied as a quantitative biomarker to estimate subfoveal choroidal thickness in children. The role of blood flow inside optic disc in FT progression needs further investigation. The distribution of FT and the peripapillary pattern correlated more with myopia-related fundus changes than the macular pattern.

Translational Relevance: Artificial intelligence can evaluate FT quantitatively in children, and has potential value for assisting in myopia prevention and control.

Introduction

Fundus tessellation (FT), defined as well-defined choroidal vessels at the posterior fundus pole outside of the parapapillary beta zone, is considered the earliest stage in the natural course of myopic maculopathy and the most common myopic fundus lesion.¹ The prevalence of FT was reported to be 48.1% among Chinese junior students and 94.3% among Chinese young high myopes.^2,3 Using manual grading methods, the higher degree of FT has been found to be associated with reduced subfoveal choroidal thickness (SFCT), longer axial length (AL), larger parapapillary atrophy (PPA) area, female sex, and larger corneal radius of curvature (CR) in children.^1–3 However, manual grading methods rely on the judgment of ophthalmologists, which is influenced easily by ophthalmologists’ experience and understanding of grading standards and are thereby limited in accuracy as well as repeatability. Recently, a new quantitative method has been proposed, which is based on artificial intelligence (AI) image processing technology to extract the exposed choroid from fundus images, and then calculate the average exposed choroid area per unit area of the fundus, named fundus tessellated density (FTD).^4,5 FTD has been identified as a reliable and objective index of FT that can be used to quantitatively assess the associations of FT with ocular and systemic parameters in the elderly.⁴ However, no studies have applied FTD to investigate factors associated with FT in children. In addition, the relationship between retinal blood flow and FT has not been evaluated comprehensively. Only one small sample study investigated changes of the peripapillary and perifoveal retinal perfusions in teenagers with FT, but did not analyze blood flow inside the optic disc.⁶ 

A previous study classified the location of FT into five primary types by manually observing the morphological characteristics of FT: in the posterior pole, macular, peripapillary, nasal, or inferior region.⁷ However, this classification is subjective and time consuming. Given that FT in the macular area and peripapillary area is clinically significant and the focus of previous researches,^1–3,8–10 the distribution of FT can be classified simply into the macular pattern (if FTD in the macular area is greater than FTD in the peripapillary area) and the peripapillary pattern (if FTD in the peripapillary area is greater than FTD in the macular area). This objective method to determine the distribution of FT based on FTD has not been used in the previous literature. 

In this population-based study of 7-year-old children, we used AI technology to obtain FTD from fundus images, explored the associations of FTD with numerous systemic and ocular factors, especially retinal blood flow, classified the distribution of FT into the macular pattern and the peripapillary pattern according to FTD, and compared the differences between these 2 patterns. 

Methods

Study Population

The Nanjing Eye Study (NES) is a population-based cohort study in eastern China, aiming to investigate the occurrence and development of ocular diseases in children longitudinally. The details of NES have been reported previously.^11,12 In brief, all children born in Yuhuatai District, Nanjing, China, between September 2011 and August 2012, and entering kindergartens in Yuhuatai District, were invited to participate in the NES for comprehensive eye examinations annually since 2015. This study is part of the NES and the data were obtained in 2019 when these children were 7 years old. 

Institutional review board or ethics committee approval was obtained from the Ethics Committee of the First Affiliated Hospital with Nanjing Medical University. Written informed consent was obtained from the parents or legal guardians of all children, and oral assent was obtained from all participants right before examinations. For this study, we included children who completed comprehensive examinations and excluded children with strabismus, amblyopia, a history of intraocular surgery, chorioretinal or vitreoretinal abnormalities except for FT, and systemic diseases. 

Ocular and Anthropometric Examinations

Comprehensive examinations, including best-corrected visual acuity, ocular surface, anterior segment, posterior segment, refraction, ocular alignment and motility, ocular biometric parameters, intraocular pressure, and optical coherence tomography angiography (OCTA), were performed by a trained team composed of ophthalmologists and optometrists. 

Noncycloplegic refraction was performed with an autorefractor (Cannon RF10; Canon, Tokyo, Japan) and retinoscopy, and spherical equivalent was calculated as sphere plus half of cylinder. The AL and CR were measured with IOLMaster-500 (Carl Zeiss Meditec AG, Jena, Germany). A nonmydriatic fundus camera (Canon CR-2) was used to capture 45° fundus photographs centered on macula. Height and weight were measured without shoes and heavy clothing, and body mass index (BMI) was calculated as weight in kilograms divided by the square of height in meters. 

Scanning for the fundus structure and retinal blood flow was performed using optical coherence tomography angiography (Optovue RTVue XR Avanti; Optovue, Inc., Fremont, CA), as described in our previously published article.¹³ Briefly, the macula was imaged with a scan area of 6 × 6 mm and a scan depth of 2.3 mm. Then retinal thickness (from inner limiting membrane to retinal pigmented epithelium [RPE]), thickness of the RPE–Bruch's membrane (from the RPE to Bruch's membrane), superficial vessel density, and deep vessel density were automatically measured by the built-in software. The optic disc was imaged with a scan area of 4.5 × 4.5 mm and a scan depth of 2.3 mm. Then peripapillary thickness of retinal nerve fiber layer (RNFL), peripapillary vessel density, and vessel density inside optic disc were automatically measured by the built-in software. Optic disk area, optic cup area, ratio of cup/disc area, horizontal diameter of optic disc, and vertical diameter of optic disc were also calculated by the built-in software. SFCT, defined as the distance from the outer border of RPE to the inner surface of the sclera running through the center of the fovea, was measured manually by a trained ophthalmologist (H.D.) using a caliper of the device software. Intragrader reliability for the SFCT measurement was good, with intraclass correlation of 0.919 based on 50 eyes measured twice. The optic cup area, the optic disk area, the horizontal diameter of the optic disc, and the vertical diameter of the optic disc were corrected using Bennett's formula (scaling factor = 3.46 × 0.013062 × [AL – 1.82]).¹⁴ Images with overall Quality Index of six or higher from the right eye were included for analysis. 

Analysis of Fundus Photographs

AI image processing technology based on deep learning, as reported in previous studies,^4,5 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 composed of preprocessing, sample labeling, segmentation by deep learning and computer vision methods, and FTD computation. 

First, fundus images were preprocessed by four steps, including the establishment of a region of interest, denoising, normalization, and enhancement, to improve the differences between internal features of each image and reduce the differences between images.¹⁵ Second, sample labeling was performed through channel calculation, as well as manual modification and review. Third, exposed choroids and atrophic arcs were extracted by a semantic segmentation network model based on deep learning. The optic disc and macula were recognized automatically based on a visual attention mechanism to locate the optic disc center and macular fovea.¹⁶ Finally, on the basis of the results of fundus feature recognition and segmentation, FTD were calculated for the whole fundus, the macular area (a circle with a radius of 6 mm centered on the fovea), and the peripapillary area (a circle with a radius of 4 mm centered on the optic disc) (as shown in Fig.). More technical details are described in the Supplementary Materials. 

Figure.

 

Quantification of FTD by AI technology. (A) An eye in the macular distributed group. (B) An eye in the peripapillary distributed group. (A1, B1) Original image. (A2, B2) FT in the macular area (6mm, green circle). (A3, B3) FT in the peripapillary area (4 mm, red circle).

The optic disc–fovea distance (DFD), defined as the distance from the optic disc center to the macular fovea, and PPA area were also calculated based on the AI image processing as described elsewhere in this article. In this study, the PPA area only includes the area of parapapillary beta and gamma zone, because the parapapillary alpha zone is difficult to be identified by AI image processing technology. 

Classification of FT Distribution

The distribution of FT was classified into two patterns according to the characteristics of FTD obtained in the macula area and peripapillary area: for eyes with FTD in the macular area greater than FTD in the peripapillary area, they were classified as the macular distributed group (macular pattern); for eyes with FTD in the peripapillary area greater than FTD in the macular area, they were classified as the peripapillary distributed group (peripapillary pattern). 

Statistical Analysis

We performed data analysis using the IBM Statistical Package for the Social Sciences program statistical package V13.0 (SPSS, Chicago, IL). Data obtained from right eye were included for analysis. To evaluate relative factors of FTD, univariate linear regression models were used, with FTD as a dependent variable, and ocular and systemic factors as independent variables. Factors with a P value of <0.05 in univariate linear regression models were included in multivariate linear regression model using the stepwise method. Independent samples t-tests were used to explore differences in ocular and systemic factors between different distribution patterns. All P values were two-sided with significance set at a P value of <0.05. The 95% confidence intervals (CI) were also given. Continuous variables were described as mean ± standard deviation and categorical variables as frequency count and percentage. 

Results

General Characteristics

A total of 577 children were enrolled in this study, including 295 boys (51.13%) and 282 girls (48.87%), with an average age of 7.44 ± 0.28 years, mean spherical equivalent of –0.34 ± 0.53 diopters, mean AL of 22.99 ± 0.75 mm, and mean AL/CR ratio of 2.95 ± 0.08. The mean FTD in the whole fundus, the macular area, and the peripapillary area were 0.024 ± 0.026, 0.023 ± 0.027, and 0.037 ± 0.046, respectively. As shown in Table 1, boys had lager BMI, AL, CR, and AL/CR (all P < 0.05) than girls, but had lower deep vessel density, vessel density inside optic disc, and peripapillary vessel density (all P < 0.05). 

Table 1.

 

General Characteristics of Enrolled Children

 

Associated Factors of FTD

In univariate linear regression analysis, a longer AL, larger CR, larger PPA area, longer DFD, thinner retinal thickness, thinner SFCT, higher vessel density inside optic disc, lower peripapillary vessel density, thinner RNFL, larger optic disk area, larger horizontal diameter of optic disc, and larger vertical diameter of optic disc (all P < 0.01) were significantly correlated with greater FTD of the whole fundus (Table 2). 

Table 2.

 

Univariate Analysis for Associations of FTD With Ocular and Systemic Parameters

 

In multivariate linear regression analysis, larger PPA area (B, 0.000; 95% CI, 0.000–0.000; P < 0.001; beta, 0.235), longer DFD (B, 0.000; 95% CI, 0.000–0.000; P = 0.04; beta, 0.074), thinner SFCT (B, 0.000; 95% CI, 0.000–0.000; P < 0.001; beta, –0.354), higher vessel density inside optic disc (B, 0.001; 95% CI, 0.000–0.001; P = 0.01; beta, 0.089), thinner RNFL (B, 0.000; 95% CI, 0.000–0.000; P = 0.02; beta, –0.088), and larger vertical diameter of optic disc (B, 0.013; 95% CI, 0.003–0.023; P = 0.01; beta, 0.092) remained significantly associated with greater FTD of the whole fundus (Table 3). 

Table 3.

 

Multivariate Analysis for Associations of FTD With Ocular and Systemic Parameters

 

Comparison Between Different FT Distribution Groups

Among the 577 children included, 207 (35.7%) were classified as the macular distributed group and 371 (64.3%) as the peripapillary distributed group. Table 4 showed the comparison of systemic and ocular parameters between these two groups. The macular distributed group had more boys (56.80% vs 47.98%; P = 0.04) and larger AL/CR (2.96 ± 0.07 vs 2.94 ± 0.08; P = 0.04) than the peripapillary distributed group. However, the peripapillary distributed group had larger PPA area (0.052 ± 0.119 vs 0.031 ± 0.072; P = 0.008), greater FTD (0.029 ± 0.028 vs 0.015 ± 0.018 in the whole fundus [P < 0.001]; 0.026 ± 0.029 vs 0.017 ± 0.021 in the macular area [P < 0.001]; 0.052 ± 0.050 vs 0.010 ± 0.018 in the peripapillary area [P < 0.001]), thinner SFCT (297.66 ± 60.61 vs 315.33 ± 66.46; P = 0.002), and thinner retinal thickness (285.55 ± 10.89 vs 288.03 ±1 0.31; P = 0.008) than the macular distributed group. There were no statistical differences in height, weight, BMI, best-corrected visual acuity, intraocular pressure, AL, CR, spherical equivalent, DFD, superficial vessel density, deep vessel density, thickness of RPE–Bruch's membrane, peripapillary vessel density, vessel density inside optic disc, thickness of RNFL, or size of optic disc between these two groups (all P > 0.05). 

Table 4.

 

Comparison Between Different Distribution Patterns

 

Discussion

In this study, we creatively applied AI technology to obtain FTD in children, explored the associations between FTD and numerous factors, especially retinal blood flow, and compared the characteristics of different FT distribution patterns. The FTD was negatively correlated with SFCT and RNFL thickness, and positively correlated with PPA area, vessel density inside the optic disc, vertical diameter of the optic disc, and DFD. Eyes with an FTD greater in the peripapillary area had larger PPA, greater FTD, thinner SFCT, thinner retinal thickness, and smaller AL/CR than eyes with FTD greater in the macular area. 

The association between FT and SFCT has been reported in many studies.^10,17–19 Progression of FT has been assumed to accelerate with the thinning and hypoperfusion of choroid.²⁰ Previous studies using manual grading methods showed that the SFCT decreased significantly as the degree of FT increased.^2,3 In the Beijing Eye Study, which used the same AI technology as our study, the FTD was found to be associated closely with a thinner SFCT in the elderly population, which is consistent with the findings in our study.⁴ Therefore, we hypothesize that FT might be caused partially by structural changes of the choroid, and FTD can be used as a quick quantitative biomarker to estimate SFCT, not only in the elderly people, but also in primary school children, as in our population. In addition, it has been accepted widely that the thinning of the SFCT is highly related to myopia and myopic maculopathy.^17,19,21 The distribution pattern of the choroidal thickness in tessellated eyes was different from that in eyes without maculopathy, but similar to that in eyes with significant myopic maculopathy, suggesting that FT might be the first sign for myopic eyes to become pathological.²² Thus, there might be a reciprocal relationship between FT, choroid, and myopic macular degeneration that needs further investigation. 

PPA has been found to be a predictive indicator for the development of advanced myopic chorioretinal atrophy and myopic macular degeneration in later life.^4,23 Our study showed that a larger PPA area was also an important factor associated with FTD, similar to the findings of the Beijing Eye Study in the elderly population.⁴ This finding indicates the same mechanism shared by the development of FT and enlargement of PPA, both of which might play a role in the development of myopic macular degeneration. However, a school-based study using manual methods to identify FT and PPA found no association between FT and PPA among Chinese junior students.² These contradictory results may be due to the difference in the range of PPA. In our study and the Beijing Eye Study, PPA included the parapapillary beta and gamma zones, whereas in the school-based study of Chinese junior students, only the parapapillary gamma zone was analyzed. In addition, we included retinal blood flow in the analysis to control for confounding factors. 

Because the retinal microvascular network plays an important role in the development and maintenance of the fundus structure, we explored the relationship between retinal blood flow and FTD, and found that greater vessel density inside optic disc was related to greater FTD. The increase in blood flow inside optic disc was also observed in high myopic adults.²⁴ However, peripapillary retinal perfusion was not associated with FTD in our study, but was found to be decreased among teenagers with FT and adults with high myopia in previous studies.^6,24 We hypothesize that the increased blood flow inside optic disc might be a compensation for the peripapillary retinal perfusion to ensure the normal function of retinal tissue in the early stage of FT and myopia. With the increase of age and the progression of FT and myopia, the increased blood flow inside optic disc might not be able to compensate for the decreased peripapillary retinal perfusion. Our 7-year-old children might have a stronger capacity for compensation than teenagers and adults to maintain the normal peripapillary retinal perfusion and have increased blood flow inside the optic disc as FTD progresses. However, more studies are needed to clarify the mechanism. 

The associations of FTD with AL, CR, and refraction were not significant in the multivariate model. Axial elongation was hypothesized to be associated with changes in the vascular and nonvascular tissues of the choroid. The significance of AL in the univariate analysis disappeared in the multivariate analysis, which may be explained by the emmetropic and mild myopic refractive status of the participants. However, the significance of DFD and RNFL thickness in the multivariate analysis confirms the impact of the extension of fundus on FTD in another aspect. A recent study using manual grading methods in students aged 9 to 12 years reported that higher degree of FT was associated with larger CR, which was not found in our study. This might be due to younger age of the participants and different methods for analyzing FT in our study. 

Age was found to be correlated positively with the severity of FT in a study among the elderly.⁹ In contrast, neither the present study nor another study of adolescents showed an association between age and FT.³ This finding implies that the effect of age on FT might be slight in young people, but might become obvious as age advances. In addition, the association between gender and FT remains controversial. Yan et al.⁹ found that a higher degree of FT was associated with male sex in the elderly. However, an association between female sex and FT was reported by Cheng et al.³ in highly myopic children and adolescents. In our study, no association was found between gender and FTD. These discrepancies might be due to differences in age and refractive status of the study participants. Height, weight, and BMI were not related to FTD in our study, which was consistent to the findings in previous studies. 

Traditionally, FT has been manually classified into five categories: posterior pole, macular, peripapillary, nasal, and inferior. A study among individuals more than 40 years old found that eyes with FT in the posterior pole and macular region had a longer AL and concluded that the location distribution of FT was important for the correct interpretation and prediction of myopic changes.⁷ In the present study, we divided eyes with FT into two patterns based on FTD obtained from the AI technology, a method that is more objective and simple. In comparison, we found that eyes with FT mainly distributing in the macular area had larger AL/CR, suggesting that this FT distribution pattern might be more related to myopic status. In contrast, eyes with FT mainly distributing around the optic disc had more myopia-related fundus changes, including a larger PPA, greater FTD, thinner SFCT, and thinner retinal thickness. Because myopic macular degeneration is significantly associated with the thinning of SFCT and the enlargement of PPA,^10,25,26 the peripapillary FT distribution pattern might be a warning of myopic macular degeneration in the future. 

The first strength of this study is the use of AI technology to obtain the quantitative index of FT, an objective method with confirmed accuracy.⁴ Second, the relationship between FT and retinal blood flow was comprehensively evaluated by optical coherence tomography angiography in a large sample size. To the best of our knowledge, this study is the first to indicate a possible compensatory role of blood flow inside optic disc as FT progresses. This study also has some shortcomings. First, this is a cross-sectional study, and cohort studies are still needed to investigate causal relationships and longitudinal changes of FT. Second, we failed to provide a reasonable explanation for the association between FTD and the vertical diameter of the optic disc found in this study. Finally, cycloplegic refraction was not available, making the analysis of the association between refractive status and FTD less reliable, although we included both noncycloplegic refraction and AL/CR ratio in the analysis. 

Conclusions

A larger PPA area, longer DFD, thinner SFCT, greater vessel density inside the optic disc, thinner RNFL, and larger vertical diameter of optic disc were significantly associated with greater FTD. FTD can be applied as a quick quantitative biomarker to estimate SFCT in primary school children. The role of blood flow inside optic disc in FT progression needs further investigation. The peripapillary FT distribution pattern correlated more with myopic fundus changes than the macular pattern, including larger PPA, greater FTD, thinner SFCT, and thinner retinal thickness. Longitudinal studies are needed to verify the findings in this cross-sectional study. 

Acknowledgments

The 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) and Jiangsu Province's Science and Technology Project (Grant No. BE2020722). The sponsor or funding organization had no role in designing or conducting this research. 

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