Automated measurement and correlation analysis of fundus tessellation and optic disc characteristics in myopia (2025)

Over the past five decades, the prevalence of myopia has been rising rapidly worldwide, especially in East Asia^1,2. Holden et al.³ suggests that by 2050, 49.8% of the global population will be affected by myopia, with 9.8% suffering from high myopia. Previous clinical studies^4,5,6,7 have identified changes in the fundus, including alterations in fundus tessellation (FT) and the optic disc (OD), such as OD torsion and peripapillary atrophy (PPA), as primary indicators of myopic fundus changes.

Different investigations^8,9,10,11,12 have delved into the relationship between FT and OD features with the progression of myopia and its impact on visual function across diverse populations. These studies have typically employed manual grading or measurements, which were dependent on the observer’s expertise and inherently subjective and labor-intensive. As a result, these manual approaches have been constrained in their widespread application for monitoring myopia. Recently, deep learning (DL) algorithms have emerged as a promising tool for the automated identification and grading of tessellated fundus and pathological myopia from color fundus images^13,14,15,16. Several studies employed DL-based segmentation models to calculate the fundus tessellated density (FTD) and explore its associations with geographic and ocular factors in elderly communities or school-age children, but overlooked the automated measurement of OD changes^17,18.

We conducted a new hospital-based study involving myopic patients among young adults at our institution. By employing DL techniques, this study was able to automatically quantify a range of myopia-related retinal changes, including FT and OD characteristics, from fundus images of myopic eyes. Furthermore, we sought to investigate the quantitative associations between FT and OD characteristics and the severity of myopia and refractive function parameters.

This study was approved by the Institutional Review Board of Qingdao Eye Hospital of Shandong First Medical University (QEH) and was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all subjects and/or their legal guardian(s) before enrollment.

Patients and examinations

In Fig.1A, 1126 fundus images from 584 myopic patients who attended QEH between May 2021 and September 2021 were retrospectively involved in the development cohort. After image quality check, 101 poor-quality fundus images were excluded. The reserved 1,025 images were used for developing and testing the DL models for automated segmentation of myopia-associated retinal tissues.

In Fig.1B, a total of 599 young myopic patients, aged from 18 to 35, who accepted refractive surgery in QEH from November 2021 to June 2022, were prospectively enrolled. Each patient received an extensive preoperative assessment for refractive surgery, encompassing a comprehensive bilateral eye examination. This assessment included several regular examinations of best-corrected visual acuity, slit-lamp and intraocular pressure. Furthermore, a suite of ocular measurements tailored to the assessment of refractive function together with color fundus imaging were also performed. Eyes with any other co-existing ocular diseases or poor-quality fundus images were excluded. Finally, 986 eyes from 496 patients with completed fundus images and ocular examination records formed the validation cohort, which was used for automated calculation of fundus characteristics and subsequent correlation analysis.

All the fundus images were captured using non-mydriatic fundus cameras (CR-2 AF, Canon, Japan), primarily employing a 45°macula-centered imaging protocol. A total of eight ocular parameters, providing multifaceted view of each eye’s refractive condition, were recorded and utilized in our study. The spherical equivalent (SE) was taken with an auto kerato refractometer (KR-1, Topcon, Japan). Biological developmental measurements, including axial length (AL), corneal curvature (K1 and K2), central corneal thickness (CCT), anterior chamber depth (ACD), and crystal thickness (CST), were obtained using an optical biometer (OA2000, Tomey, Japan). Additionally, foveal thickness (FVT) was measured via macular OCT (HD-OCT 5000; Zeiss, Germany).

Annotation for model development

The myopia-associated tissues in the fundus images of the development cohort were labeled for model development. First, the annotation criteria and examples were provided by one senior ophthalmologists with over ten years of clinical experience. Pixel level annotations were required for fundus tessellations, peripapillary atrophy (PPA), OD, and macular fovea. Second, all the images were randomly divided into quarters and annotated by four junior ophthalmologists (each with two or three years of clinical experience) according to the established criteria. Finally, all labeled images were checked and revised by the senior ophthalmologist. The annotated images were further randomly split into three subsets for model training, validation, and testing according to the ratio of 7:1:2.

Segmentation of fundus tissues via deep learning

We developed two separate DL models using the state-of-the-art Transformer-based segmentation network known as SegFormer¹⁹ to segment myopia-related tissues from fundus images (Supplementary Fig.1). First, the FT-Seg model was established to extract the FT. To enhance the contrast of FT, we proposed a color channel recalibration (CCR) algorithm to pre-process the input fundus images. The CCR algorithm applied the CLAHE²⁰ method to the red channel of the image, while the pixel values of the green and blue channels were multiplied by a factor of 0.5. Second, the OD-Seg model was trained to segment the PPA, OD, and fovea simultaneously from the original fundus images.

All fundus images inputted into the two segmentation models were cropped to remove redundant background, padded to the minimal square with zeros, and resized to 512 × 512. Common data augmentation techniques, i.e., random cropping, resizing, rotation, and horizontal flipping, were employed. Both models adopted the Dice loss²¹ and were initialized with pre-trained weights on ImageNet and updated using the Adam optimizer with an initial learning rate of 0.001. During training, the models that achieved the highest average Dices in the validation subset were picked out. Both DL networks were implemented using Pytorch (https://pytorch.org/), and the models were trained on an NVIDIA Tesla P100 GPU.

To assess the segmentation performance of our models in the testing subset, four metrics were calculated for each tissue: Dice coefficient, pixel-wise accuracy, pixel-wise sensitivity, and pixel-wise specificity. Dice measures the overlap between model prediction maps and ground truths, whereas pixel-wise accuracy represents the proportion of pixels that were correctly predicted. Pixel-wise sensitivity was defined as the proportion of positive pixels that were correctly segmented, and pixel-wise specificity estimated the proportion of negative pixels that were properly segmented.

Quantitative measurement of fundus characteristics

After applying the DL models to the fundus images of the validation cohort, we measured the myopia-associated characteristics on the segmentation predictions. The FT-Seg model segmented tessellations from whole fundus images. Three regions of interest (ROIs) were extracted using the OD-Seg model. First, the foreground of the fundus image is defined as the posterior pole zone. Second, a concentric circle of the macular fovea, with a diameter of 6.0mm, was defined as the macular zone. Third, an ellipse with a major axis of 6.0mm and a minor axis of 4.5mm was derived, centered on the OD prediction. Furthermore, we considered the tessellation segmentation results and calculated the tessellated density within the three ROIs, namely, FTD, macular tessellation density (MTD), and peripapillary tessellation density (PTD). Tessellation density was defined as the proportion of tessellated pixels within the ROI. OD characteristics were generated directly from the outputs of OD-Seg, including OD area (DA), PPA area (PA), ratio of PA over DA (RPD), and OD ovality (DV). Magnification effects are corrected for DA and PA using Littmann’s method²². The DV is defined as the ratio of the minor axis to the major axis of the fitting ellipse of the OD.

Statistical analysis

Statistical analyses were performed on myopia-associated fundus characteristics and ophthalmic examination parameters. Continuous variables are described as the mean ± standard deviation, while discrete variables are presented as counts (proportions). Moreover, we divided all subjects in the validation cohort into high myopia and mild or moderate myopia groups to explore the differences in variables between the groups. First, a Shapiro–Wilk test was performed to determine the normality of the distribution for each variable. To detect differences in variables between the two groups, a chi-square test was implemented on the gender proportion, whereas a t-test or Wilcoxon rank-sum test was used for other ocular parameters and fundus characteristics. Second, associations between fundus measurements (including FTD, MTD, PTD, DA, PA, RPD, and DV) and ocular parameters (including SE, AL, CCT, K1, K2, ACD, CST, and FVT) of each eye were assessed using Spearman correlation analysis. Furthermore, a stepwise multivariate linear regression analysis was performed to ascertain the comprehensive impact of fundus measurements on key ocular parameters. All statistical analyses were performed using the R statistical package (version 4.0.5).

Evaluation of fundus segmentation algorithms

We assessed the segmentation performance of the DL models using 205 independent testing images selected from the development cohort. For FT, PPA and OD, the Dice coefficients were 72.23%, 81.23% and 92.94%, respectively, while the pixel-wise accuracies were 95.28%, 99.69% and 99.80%, respectively; the pixel-wise sensitivities were 73.00%, 81.93% and 93.43%, and the pixel-wise specificity was 96.87%, 99.85% and 99.90%, respectively. When the FT-Seg model was trained without CCR pre-processing, the Dice, accuracy, sensitivity, and specificity of FT dropped to 70.61%, 94.09%, 70.81% and 96.11%, respectively. Figure2 illustrates two examples showcasing manual annotations and model predictions, along with their corresponding calculated measurement values.

Validation cohort characteristics

A total of 986 eyes of 496 patients from the validation cohort were included in the statistical analysis (Fig.1). The mean age of the 496 patients was 24.44 ± 5.59years (range, 18–35), and 58.32% were female. In this study, myopia was defined as an SE ≤ -0.5D, with mild, moderate, and high myopia were defined as SE of ≤ − 0.5DD, ≤ − 3.0D, and ≤ − 6.0D, respectively^23,24. Subsequently, 325 (32.96%) eyes exhibited high myopia and the remaining 661 eyes (67.04%) were treated as mild or moderate myopia. In terms of the ocular parameters, the mean SE was − 6.06 ± 2.44 D (range − 0.50 D to − 21.75 D); the mean AL was 26.01 ± 1.17mm (range 22.64–30.96mm); the mean CCT was 531.64 ± 31.05μm (range 433–646μm); the mean K1 and K2 were 42.77 ± 1.39 D (range 38.44–48.35 D) and 44.15 ± 1.60 D (range 39.47–49.71 D); the mean ACD, CST and FVT were 3.82 ± 1.09mm (range 2.98–4.94mm), 3.64 ± 0.27mm (range 2.90–6.28mm) and 251.00 ± 18.69μm (range 156–321μm) (Table 1). As for myopic fundus measurements, the mean FTD, MTD, and PTD were 0.067 ± 0.081 (range 0–0.470), 0.091 ± 0.122 (range 0–0.614) and 0.049 ± 0.063 (range 0–0.354); the mean PA, DA, RPD, DV were1.139 ± 1.163 mm² (range 0–14.43 mm²), 2.397 ± 0.516 mm² (range 0.863–5.235 mm²), 0.498 ± 0.488 (range 0–3.900) and 0.788 ± 0.095 (range 0.496–0.989) (Table 2).

Association between fundus measurements and myopia severity

In the group effect analysis, the tessellated density in each of the three ROIs showed similar increasing trends, with myopia severity approaching from mild or moderate to high (Table 2; Fig.3). The mean FTD increased from 0.041 ± 0.056 in the mild or moderate myopia group to 0.120 ± 0.097 (p < 0.001) in the high myopia group, the MTD increased from 0.052 ± 0.080 to 0.171 ± 0.151 (p < 0.001), and the PTD increased from 0.031 ± 0.045 to 0.088 ± 0.076 (p < 0.001). The PA was significantly larger in the high myopia group than in the other groups (p < 0.001), and a similar trend was observed for RPD (p < 0.001). A larger DA was observed in the high myopia group than in the other group (p < 0.001). The mean DV was 0.770 ± 0.105 in the high myopia group and 0.796 ± 0.089 in the non-high myopia group (p < 0.001) (Table 2; Fig.4).

Association between fundus measurements and ocular parameters

In the correlation analysis, all the fundus measurements were significantly associated with SE and AL, with p-values less than 0.001 for both parameters, indicating a highly statistically significant relationship. In addition, FTD, MTD, PA and DA showed a significant correlation with Keratometric values (K1 and K2), while DV was highly significantly (p < 0.001) associated with K2 and FVT (Table 3).

A multivariate regression analysis was conducted on the significant outcomes of the correlation analysis (Table 4). With SE as the dependent variable, FTD and RPD were dropped because of their high collinearity with other variables, and gender was used to adjust the model. The DA was dropped because of a lack of significant association. Lower MTD (p < 0.001), lower PTD (p < 0.01), smaller PA (p < 0.001), and smaller DV (p < 0.001) were ultimately included, where the adjusted r-squared was 0.3537. With AL as the dependent variable, all fundus measurements were included as the independent variables. Furthermore, the FTD and RPD were removed to eliminate collinearity, and gender was used to adjust the model. In the optimal model, longer AL was associated with higher MTD (p < 0.001), higher PTD (p < 0.01), larger PA (p < 0.001), and larger DA (p < 0.001), and the adjusted r-squared was 0.4011.

FT is an important feature in myopic fundus and has been analyzed in many studies. Yoshihara et al.⁸ developed an objective and quantitative method to determine the degree of tessellation called tessellated fundus indices (TFIs). The TFIs were calculated using the image processing program ImageJ; however, the classifications of tessellation and the selection of the calculation area were performed by retina specialists. Gao et al.⁹ conducted a school-based study to determine the prevalence of tessellated fundus and its association with ocular and general parameters among junior students. This study revealed that the prevalence of FT was relatively high in Chinese teenagers and that the degree of FT is a surrogate for choroidal thickness in teenagers. OD and peripapillary changes in high myopia cannot be neglected and may help in a better understanding of the pathophysiology or mechanism of myopia progression¹⁰. PPA is a clinical discovery related to choroid-retinal thinning and interruption of the RPE in the peripapillary area, which can occur in both benign and pathologic states of high myopia^11,12.

Shao et al.¹⁷ first reported quantitative measurements of FTD from color photographs using deep convolutional neural network (DCNN) based image processing technology. Their color photographs were taken from the Beijing Eye Study, a population-based epidemiological study in which an age of no less than 50years was the only inclusion criterion. Huang et al.¹⁸ conducted comprehensive ocular examinations in 577 children aged 7years old from a population-based cross-sectional study. In this research, FT measurements were also determined automatically using a DCNN model, serving as a rapid quantitative biomarker for estimating subfoveal choroidal thickness in primary school-aged children. While both studies employed DL models for their analyses, they did not thoroughly investigate the relationship between fundus changes and the progression of myopia within the myopic population.

In this study, we developed DL models to automatically quantify myopia-related fundus changes including FT and OD signs. We employed a state-of-art segmentation model, SegFormer, to generate tessellation prediction maps from the fundus images. Uneven image quality and diverse color styles inevitably influence the computer-aided diagnosis of retinal diseases. An image preprocessing method CCR was proposed to highlight tessellations in fundus images and reduce the influence of uneven illumination, low contrast, and irrelevant tissues (Fig.2). This could contribute a gain of 1.62% in Dice and 2.19% in sensitivity for segmenting FT, respectively. Additionally, a SegFormer-based model was specifically trained to delineate the macular and peripapillary zones, enabling the calculation of the FTD, MTD, and PTD. Concurrently, this model was also utilized to segment the OD and PPA. Following the segmentation, key characteristics around the OD were determined, including DA, PA, RPD, and DV, providing a detailed analysis of the OD region. Our study revealed that the automated calculations of fundus characteristics, including both FT features (FTD, MTD, and PTD) and OD attributes (DA, PA, RPD, and DV), demonstrated highly significant (p < 0.001) differences when comparing the high myopia group with the moderate or mild myopia group within our validation cohort. This finding indicates the potential utility of these quantitative fundus parameters in the identification of high myopia.

Furthermore, we assessed the correlations between the abovementioned FT and OD measurements and a spectrum of myopia-related ocular parameters. Initially, we observed a highly significant correlation (p < 0.001) between all calculated FT and OD characteristics with both SE and AL. Multivariate regression analysis utilizing SE as the dependent variable indicated that MTD, in conjunction with PTD, PA, DV, and gender, collectively accounted for a substantial portion of the dioptric power. When AL was considered as the dependent variable, MTD, PTD, PA, DA, DV and gender emerged as significant predictors. These results highlight the potential of fundus imaging as a valuable tool for estimating SE and AL, offering important insights for refractive function evaluation.

This study not only enhances our understanding of the morphological fundus changes associated with myopia but also helps identify specific patterns that could be used for early detection and monitoring of myopia. Looking forward, we plan to conduct a longitudinal cohort study to investigate whether incorporating FT and OD measurements can enhance the prediction of myopia severity and refractive function. By tracking these parameters over time, we aim to determine if they provide additional predictive value beyond traditional methods such as refraction tests and axial length measurements. This future research will help us assess the potential of these morphological features as supplementary indicators in the comprehensive management of myopia, potentially leading to more personalized and effective treatment strategies.

There are several potential limitations to our study. First, the number of images used was small. However, as only myopic patients undergoing corneal refractive surgery were included, the interference of other fundus diseases could be minimized. Second, this work is a single-center study, external datasets from other manufacturers have not been collected, and the generalization of our method has yet to be further verified.

In summary, this research marks a significant advance in utilizing DL technology to automate the measurement of key fundus features, such as FT and OD characteristics, in young patients with myopia. Notably, both FT and OD characteristics demonstrate a strong correlation with the progression of myopia, especially as it escalates from mild or moderate to severe stages. Furthermore, a substantial correlation is observed between these fundus measurements and a range of ocular parameters associated with refractive function. The findings underscore the potential of these characteristics to serve as biomarkers for assessing the severity of myopia and its impact on refractive function.

AL:

Axial length

DL:

Deep learning

FT:

Fundus tessellations

OD:

Optic disc

FTD:

Fundus tessellated density

PPA:

Peripapillary atrophy

CCR:

Color channel recalibration

ROI:

Region of interest

MTD:

Macular tessellated density

PTD:

Peripapillary tessellated density

AOD:

Area of optic disc

APPA:

Area of peripapillary atrophy

ARAD:

Area ratio of peripapillary atrophy over optic disc

OID:

Ovality index of optic disc

CNT:

Corneal thickness

CST:

Crystal thickness

FVT:

Foveal thickness

SE:

Spherical equivalent

ACD:

Anterior chamber depth

Authors and Affiliations

1. Eye Institute of Shandong First Medical University, Qingdao Eye Hospital of Shandong First Medical University, 5 Yanerdao Road, Qingdao, 266071, China

   Zhen Guo,Yan Gao,Qianqian Liang,Shuyue Xue,Qing Du,Zhichun Zhang&Jun Li

2. State Key Laboratory Cultivation Base, Shandong Provincial Key Laboratory of Ophthalmology, Qingdao, China

   Zhen Guo,Yan Gao,Qianqian Liang,Shuyue Xue,Qing Du,Zhichun Zhang&Jun Li

3. School of Ophthalmology, Shandong First Medical University, Qingdao, China

   Zhen Guo,Yan Gao,Qianqian Liang,Shuyue Xue,Qing Du,Zhichun Zhang&Jun Li

4. Ping An Healthcare Technology, Beijing, China

   Lingzhi Chen,LiLong Wang,Bin Lv,Guanzheng Wang&Guotong Xie

5. Ping An Healthcare and Technology Company Limited, Shanghai, China

   Guotong Xie

Contributions

All authors contributed to the study’s conception and design. J.L. designed the experiments. Z.G., L.C., and L.W. analyzed the data and wrote the first draft of the manuscript, and all authors commented on previous versions of the manuscript. All the authors read and approved the final manuscript.

Corresponding author

Correspondence to Jun Li.

Competing interests

The authors declare no competing interests.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions