Article Text

Original research
Quantitative assessment of colour fundus photography in hyperopia children based on artificial intelligence
  1. Ruiyu Luo1,
  2. Zhirong Wang1,2,
  3. Zhidong Li1,
  4. Rui Xie1,
  5. Yuan Zhang1,
  6. Guitong Ye1,
  7. Jianqi Chen1,
  8. Yue Xiao1,
  9. Jinan Zhan1,
  10. Yingting Zhu1,
  11. Yehong Zhuo1
  1. 1Ophthalmic Center State Key Laboratory of Ophthalmology, Sun Yat-Sen University Zhongshan, Guangzhou, Guangdong, China
  2. 2Foshan Women and Children's Hospital, Foshan, Guangdong, China
  1. Correspondence to Dr Yehong Zhuo; zhuoyh{at}


Objectives This study aimed to quantitatively evaluate optic nerve head and retinal vascular parameters in children with hyperopia in relation to age and spherical equivalent refraction (SER) using artificial intelligence (AI)-based analysis of colour fundus photographs (CFP).

Methods and analysis This cross-sectional study included 324 children with hyperopia aged 3–12 years. Participants were divided into low hyperopia (SER+0.5 D to+2.0 D) and moderate-to-high hyperopia (SER≥+2.0 D) groups. Fundus parameters, such as optic disc area and mean vessel diameter, were automatically and quantitatively detected using AI. Significant variables (p<0.05) in the univariate analysis were included in a stepwise multiple linear regression.

Results Overall, 324 children were included, 172 with low and 152 with moderate-to-high hyperopia. The median optic disc area and vessel diameter were 1.42 mm2 and 65.09 µm, respectively. Children with high hyperopia had larger superior neuroretinal rim (NRR) width and larger vessel diameter than those with low and moderate hyperopia. In the univariate analysis, axial length was significantly associated with smaller superior NRR width (β=−3.030, p<0.001), smaller temporal NRR width (β=−1.469, p=0.020) and smaller vessel diameter (β=−0.076, p<0.001). A mild inverse correlation was observed between the optic disc area and vertical disc diameter with age.

Conclusion AI-based CFP analysis showed that children with high hyperopia had larger mean vessel diameter but smaller vertical cup-to-disc ratio than those with low hyperopia. This suggests that AI can provide quantitative data on fundus parameters in children with hyperopia.

  • Child health (paediatrics)
  • Imaging
  • Optic Nerve
  • Optics and Refraction
  • Retina

Data availability statement

No data are available.

This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See:

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.


  • Children with hyperopia exhibit structural characteristics and differences in the posterior segment; however, colour fundus photography features in this population remain unknown. This study aimed to analyse optic nerve head and retinal vascular parameters on colour fundus photographs using artificial intelligence and to explore potential influencing factors.


  • Highly hyperopic eyes with shorter axial length have relatively smaller optic discs and smaller optic cups compared with low-hyperopic eyes. Therefore, the axial length is an effective objective predictor of optic disc and retinal vascular parameters.


  • Artificial intelligence can precisely capture fundus parameters on colour fundus photographs and may be useful for monitoring and evaluating the development of the optic nerve head and retinal vessels in children with hyperopia.


At birth, human eyes are predominantly hyperopic, and with age, eyeballs grow to become emmetropic or even myopic.1 Moderate-to-high hyperopia is common in children between 5 and 15 years of age, with a prevalence ranging from 2.1% to 19.3% in different studied populations.2 Increasing evidence suggests that children with moderate-to-high hyperopia have an increased risk for amblyopia and strabismus,3–5 lower preschool literacy, worse stereopsis and less accurate accommodation than their emmetropic peers.6–10 Moreover, the proportion of abnormal electroretinograms was found to increase with higher hyperopia.11 12

Prior studies using optical coherence tomography (OCT) have indicated that children with high hyperopia have significantly greater choroidal thickness13–15 and thicker average peripapillary retinal nerve fibre layer16 than controls.

However, few studies have used colour fundus photography (CFP) to monitor changes in the retina among children with hyperopia; thus, data on CFP features in this population are scarce. CFP stands as an efficient and straightforward fundus examination technique,17–19 potentially representing the optimal choice for evaluating retinal features in children due to its non-invasive nature and relatively quick turnaround. Additionally, in recent years, the development of artificial intelligence (AI)-based image processing technology20 21 has enabled the effective identification of texture nuances that cannot be distinguished by human eyes.22

The assessment of the optic nerve head (ONH) is crucial in evaluating and diagnosing numerous ocular conditions, including myopia, glaucoma and optic nerve oedema, which result in structural changes in the posterior globe. In addition, studies have shown that retinal vascular calibre changes are related to different ocular conditions,23 24 and that anatomic relationships exist between the ONH and retinal vasculature.25 Fundus examination of individuals with high hyperopia often shows crowding of the optic disc due to a small disc size in adult populations.26 27 Several studies focusing on characterising ONH structure in adults with refractive errors26 28 29 have reported that the ONH of myopic eyes is significantly larger while that of hyperopic eyes is smaller compared with emmetropic eyes. While ONH parameters were not influenced by refractive errors or axial length (AL) in Singapore nor in African-American children,30 31 a previous study using Heidelberg retinal tomography (HRT) for disc measurements reported that ONH parameters were significantly smaller in the hyperopic group than in the emmetropic group.32 However, these studies only evaluated a limited number of ONH parameters, such as optic disc area, cup area and vertical cup-to-disc ratio (CDR). Moreover, reports regarding retinal vascular structure in children are scarce.25 To date, only one study has reported a decrease in retinal vein diameter with age, but this study was conducted among adults aged 40 years or older.24

Based on the above, more ONH and retinal vascular parameters should be investigated in children with hyperopia to provide insight into the process of emmetropisation. Therefore, this study aimed to explore the relationship of ONH and retinal vascular parameters with age and spherical equivalent refraction (SER) in children with hyperopia using AI-based CFP analysis.



We retrospectively reviewed the clinical data of children aged 3–12 years from the Paediatric and Ophthalmology Departments of Zhongshan Ophthalmic Center, Sun Yat-sen University. Children with an SER ≥+0.5 D who had completed at least one cycloplegic refraction, one biometric analysis and one CFP examination at the same visit were included. We used only data from the right eye because of the high correlation of ONH and retinal vascular parameters between both eyes (p<0.001). Participants who met any of the following criteria were excluded: (1) AL>25 mm; (2) ocular diseases, such as cataract, glaucoma, corneal infection, fundus diseases, ocular trauma or other ocular abnormalities; (3) history of wearing orthokeratology lenses or using tropicamide eye-drops; (4) history of ocular surgery or trauma and (5) systemic diseases or developmental disorders.

Ophthalmic examinations

All participants underwent comprehensive ocular examinations, including best-corrected visual acuity assessment (Snellen letters or Lea symbols for literate and preliterate children, respectively; Smart System, M&S Technologies,, intraocular pressure evaluation (SW-5000; Suoer, Tianjin, China), slit-lamp biomicroscopy (SL-130; Carl Zeiss Meditec,, ocular biometric parameter measurement (IOL Master 700; Carl Zeiss Meditec, Germany) and CFP (non-stereoscopic photographs of 45° of the central fundus; fundus camera type CR6-45NM; Canon, Tokyo, Japan). All evaluations were performed by a trained team composed of ophthalmologists and optometrists.

To obtain the ONH and vascular measurements in real-size units, we corrected the CFP magnification for each AL by applying Littmann’s formula.33 Standard mydriatic optometry with compound tropicamide (0.5% tropicamide plus 0.5% phenylephrine hydrochloride) three times was performed on all the participants. Two drops were applied each time at an interval of 10 min. After the third administration, the participants could undergo mydriatic optometry once they met the following criteria: no light reflexes and pupil diameter larger than 6 mm. The SER was calculated according to the format: SER=spherical degrees+(cylindrical degrees/2). According to previous studies, low hyperopia (LH) was defined as SER of +0.5 D to <+2.0 D, and moderate to high hyperopia (MHH) was defined as SER ≥+2.0 D. MHH was further categorised into moderate and high hyperopia, namely +2.0 to <+4.0 and ≥+4.0 D, respectively.34

Image analysis

In this study, we extracted the ONH and retinal vascular morphology from CFP through AI-based image processing technology, which includes computer vision and deep learning image processing technology. First, the image was preprocessed and then ONH features were segmented by a deep learning model. Based on the results of the segmentation, ONH and vascular parameters were automatically measured. Figure 1 depicts the process of image analysis.

Figure 1

The image processing and analysis flow diagram. ROI, region of interest.

The image preprocessing can be generally divided into four steps22 35: region of interest establishment (figure 2B), denoising (figure 2C), normalisation (figure 2D) and enhancement (figure 2E). Next, the ONH was identified by the deep learning target detection network and its boundary was determined based on the mechanism of visual attention. The vascular structure of the retina was marked automatically by the deep learning target detection network, which precisely identifies the vascular borderline through brightness threshold segmentation and colour discrimination. The trained segmentation model ResUnet we employed is an improved semantic segmentation algorithm based on ResNet and U-Net. It uses ResNet to obtain high-level feature representations and uses U-Net’s downsampling and upsampling operations for pixel-level segmentation. This enables the identification of the characteristics of the ONH and retinal vasculature.

Figure 2

The image preprocessing steps: (A) original image; (B) region of interest establishment; (C) denoising; (D) normalisation; (E) enhancement.

Statistical analysis

Statistical analyses were performed by using IBM SPSS Statistics V.22.0 (IBM). The normality of data distribution was verified using the Shapiro-Wilk normality test. As all data showed skewed distribution, non-parametric tests were used. The Mann-Whitney U test was employed to compare parameters according to sex and the Kruskal-Wallis test was used for evaluating between-group differences. Statistical significance was set at p<0.05 (two sided). Spearman’s correlation analysis was applied to examine the correlation between fundus parameters and age. Univariate linear regression analysis was conducted to explore the association between fundus parameters and AL or SER. Significant variables (p<0.05) in the univariate analysis were included in a stepwise multiple linear regression model.


Participants’ characteristics

A total of 324 children were enrolled in this study, including 174 boys (53.70%) and 150 girls (46.30%), with an average age of 6.25±2.07 years, mean SER of 2.33±1.62 D and mean AL of 22.21±0.93 mm. As shown in online supplemental table 1, compared with girls, boys had larger AL, smaller vertical disc diameter and smaller superior and inferior neuroretinal rim (NRR) width (all p<0.05). A partial correlation test was performed to test the association between age, ONH and retinal vascular parameters (online supplemental table 2). When controlling for sex and SER, optic disc area, vertical disc diameter, and inferior NRR width were statistically correlated with age (partial r=−0.113, –0.112, −0.116, respectively, all p<0.05) while none of the retinal vascular parameters were statistically correlated with age.

Comparison between different SER groups

Among the 324 children included, 172 (53.09%) were classified in the LH and 152 (46.91%) in the moderate-to-high hyperopia groups. Table 1 shows the median values and interquartile ranges of all parameters in the different groups. Compared with those with LH, children with high hyperopia had larger superior NRR width (p<0.001), larger mean vessel (p=0.005) and vein diameter (p=0.003), and smaller vertical CDR (p=0.003). Compared with those with moderate hyperopia, children with high hyperopia had smaller optic cup area (p=0.004), smaller horizontal and vertical cup diameters (p=0.002 and p=0.004), smaller cup-to-disc area ratios (p=0.004) and smaller horizontal and vertical CDR (p=0.008 and p<0.001). Furthermore, optic disc area (p=0.011), horizontal disc diameter (p=0.008) and mean artery diameter (p=0.014) were larger in children with moderate hyperopia than in those with LH.

Table 1

Comparison between different refractive error groups

Relationship of fundus parameters with AL and SER

The relationship of ONH and retinal vascular parameters with AL and SER, adjusted for age and sex, is summarised in table 2. In the superior and temporal quadrants of the ONH, NRR width was significantly correlated with AL (β = −3.030, p<0.001 and β=−1.469, p=0.020, respectively). Decreased vertical CDR was significantly associated with increasing hyperopia (p=0.047). As for the retinal vascular structure, the mean retinal vessel diameter significantly decreased with increasing AL (β=−0.076, p<0.001), and a similar relationship was observed between the mean artery and vein diameters and AL (β=−0.028 and −0.061, respectively, both p<0.05). Moreover, the retinal vascular structure showed a significant positive correlation with the SER, suggesting that eyes with higher myopia have smaller retinal vessel size.

Table 2

Univariate linear regression analyses*

In the stepwise linear regression models adjusted for age and sex, smaller superior NRR width and smaller mean vessel diameter were statistically significant predictors of increased AL, as shown in table 3. Mean vessel diameter alone explained 33.3% (F=54.74; p<0.001) of the variance in participation. When superior NRR width was included in the model, the explained variance increased to 35.0% (F=44.54; p<0.001). In the regression model for SER, all significantly correlated variables were included, that is, superior NRR width, vertical CDR, mean vessel diameter, mean artery diameter and mean vein diameter. Superior NRR width and mean vessel diameter were kept in the model (R2=7.6%; F=7.60; p<0.001), as shown in table 3.

Table 3

Stepwise multivariate linear regression analyses*


In this study, we conducted a thorough evaluation of ONH and retinal vascular parameters in Chinese children with hyperopia by analysing CFP using AI. To our knowledge, CFP has not been systematically employed to investigate ONH and retinal vascular parameters in children with hyperopia. Our results indicate that children with high hyperopia have larger superior NRR width, larger mean vessel and mean vein diameters, but a smaller CDR compared with those with low and moderate hyperopia. Among the total population, with an SER ranging from +0.50 to +10.38 D, significant relationships were found between AL and some parameters, including superior and temporal NRR width, mean vessel diameter, and mean artery and vein diameters. However, on conducting multiple linear regression, no significant correlation was observed for SER, indicating that AL serves as a superior objective predictor of ONH and retinal vascular measurements when CFP is used.

Prior studies have analysed normative fundus image data by OCT and HRT in children. Jnawali et al used spectral-domain OCT to analyse ONH parameters in 5–16 years, healthy American children.36 Their study included children with a mean SER of +0.84 ± 1.12 D in the non-myopia group. However, the horizontal and vertical CDRs reported in their study were smaller than those found in our population, likely due to the difference in mean SER. Notably, the vertical CDR decreased significantly with increasing AL and myopic SER in their study, contrasting our findings that eyes with high hyperopia exhibited the smallest vertical CDR. Previous HRT studies have found that ONH parameters were not influenced by age,30 31 contradicting our observations. In our study, we found negative correlations between age and optic disc area, vertical disc diameter and inferior NRR width, although the correlation coefficients were relatively small. Pang et al investigated ONH parameters in African-American children with refractive errors up to +6.5 D.31 Their study reported significantly larger average disc and cup areas in hyperopic eyes compared with our findings. This discrepancy is likely attributed to racial differences in the ONH between African-American and Chinese children. OCT and HRT, as confocal scanning ophthalmoscopes, offer objective and precise measurements of ONH parameters and retinal nerve fibre layer thickness. These techniques are invaluable in the paediatric population due to their non-invasive nature and rapid completion. However, their high cost restricts their widespread use, particularly in remote areas where families may not be able to afford them. In contrast, CFP offers several advantages over OCT or HRT. It is rapid, direct, simple, economical and suitable for large-scale screening. For instance, CFP can provide fast and effective diagnosis during school screenings, making it a practical and efficient tool for assessing ONH and retinal vascular parameters in children.

As AL increases, posterior scleral changes occur, particularly at the posterior pole near the ONH, resulting in morphological changes in the ONH. Myopic eyes exhibit distinctive features in the ONH, including marked parapapillary atrophy, shallow disc cupping, a macrodisc with an abnormal elongation, optic disc tilt and torsion.37–40 These fundus changes have typically been attributed to the globe’s axial elongation, resulting in mechanical tissue strain and vascular changes due to the stretching process.41 Our study revealed that the mean vessel diameter, along with the mean artery and vein diameters, increase significantly with decreasing AL and increasing hyperopic SER, corroborating the above theory. Previous studies also discovered a curvilinear relationship in adults, exhibiting a sharp increase in optic disc area in highly myopic eyes starting at a refractive error of −8 D, followed by a decline in optic disc area in highly hyperopic eyes with a refractive error of +4 D.42 Similarly, our current study observed that children with moderate hyperopia had the largest optic disc and cup areas across different SER groups, indicating a curvilinear relationship and a unique development pattern of the ONH in hyperopic children.

Regarding the correlation between retinal vascular and ONH parameters, our findings revealed that eyes with a smaller vertical optic disc diameter exhibited a narrower retinal vein diameter, regardless of age and sex. Prior studies indicated an anatomic relationship between the optic disc and retinal vasculature, suggesting that the association between smaller optic discs and narrower retinal vessels may be due to the crowding at the lamina cribrosa in eyes with smaller optic discs. Our results are consistent with observations in children and older adults,25 43 but differ in that vertical CDR was negatively associated with the mean retinal artery diameter.

Using AI for quantitative analysis of CFP indices allows for the swift and straightforward acquisition of ONH and retinal vascular parameters. This AI-based approach offers a significant advantage over manual measurement by ophthalmologists, as it overcomes the limitations of subjective deviation. Nevertheless, it is crucial to acknowledge that AI screening based on images may occasionally yield false-positive and false-negative results. In clinical settings, false-positives can lead to unnecessary referrals and a waste of medical and public health resources while false-negatives can result in misdiagnosis or missed diagnosis, potentially leading to severe visual impairment. Therefore, further research is imperative to investigate how AI can be effectively integrated into clinical practice, aiming to achieve safety, efficacy, convenience and high economic benefits.

One limitation of this study lies in its cross-sectional design, which precludes the analysis of causality and restricts us to identifying mere correlations. Longitudinal studies in the future could potentially monitor the relationship between ONH and vascular parameters with AL and SER during emmetropisation, thereby shedding light on the causal impact of refractive development on the ONH. Additionally, the study’s focus on a Chinese paediatric population limits the generalisability of our findings to other ethnic groups. To address this, future multiethnic, multicentre research is imperative to provide a more comprehensive understanding of these relationships.


In summary, compared with those with low and moderate hyperopia, children with high hyperopia have relatively smaller ONH areas and larger retinal vessel diameter. Thus, optic discs in eyes with high hyperopia are more crowded as they are smaller with shorter AL.

Data availability statement

No data are available.

Ethics statements

Patient consent for publication

Ethics approval

This study involves human participants and was approved by the local research ethics committee of Zhongshan Ophthalmic Center, Sun Yat-sen University (number: 2021KYPJ85). Participants gave informed consent to participate in the study before taking part.


Supplementary material


  • RL, ZW and ZL are joint first authors.

  • Contributors RL, ZW, ZL, YZhu and YZhuo contributed to the study concept and design. RL, RX, YZhang, YZhu and YZhuo contributed to the acquisition of data; RL, GY, JC and YX contributed to the assessment of data. RL and JZ contributed to the assessment of images; RL, ZW and ZL contributed to the data analysis. RL and YZhu contributed to interpretation of the results; RL contributed to the drafting of the manuscript. YZhu and YZhuo contributed to the study supervision. All authors contributed to the critical revision of the manuscript for important intellectual content, article and approved the submitted version. RL is responsible for the overall content as guarantor. In this study, we extracted the ONH and retinal vascular morphology from CFP through AI-based image processing technology, which includes computer vision and deep learning image processing technology.

  • Funding This work was supported by the National Key R&D Project of China (2020YFA0112701); the National Natural Science Foundation of China (82171057); Science and Technology Program of Guangzhou, China (202206080005); Major Science and Technology Project of Zhongshan City (2022A1007).

  • Disclaimer The sponsor or funding organisation had no role in the design or conduct of this research.

  • Competing interests None declared.

  • Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.