Quantization of Optic Disc Characteristics in Young Adults Based on Artificial Intelligence

Abstract Purpose This study aimed to automatically and quantitatively analyse the characteristics of the optic disc by applying artificial intelligence (AI) to fundus images. Methods A total of 1084 undergraduates were recruited in this cross-sectional study. The optic disc area, cup-to-disc ratio (C/D), optic disc tilt, and the area, width, and height of peripapillary atrophy (PPA) were automatically and quantitatively detected using AI. Based on axial length (AL), participants were divided into five groups: Group 1 (AL ≤ 23 mm); Group 2 (23 mm < AL≤ 24 mm); Group 3 (24 mm < AL≤ 25 mm); Group 4 (25 mm < AL< 26 mm) and Group 5 (AL ≥ 26 mm). Relationships between ocular parameters and optic disc characteristics were analysed. Result A total of 999 undergraduates were included in the analysis. The prevalence of optic disc tilting and PPA were 47.1% and 92.5%, respectively, and increased with the severity of myopia. The mean optic disc area, PPA area, C/D, and optic disc tilt ratio were 1.97 ± 0.46 mm2, 0.84 ± 0.59 mm2, 0.18 ± 0.07, and 0.81 ± 0.08, respectively. In Group 5, the average optic disc area (1.84 ± 0.41 mm2) and optic disc tilt ratio (0.79 ± 0.08) were significantly smaller and the PPA area (1.12 ± 0.61 mm2) was significantly larger than those in the other groups. AL was negatively correlated with optic disc area and optic disc tilt ratio (r=–0.271, –0.219; both p < 0.001) and positively correlated with PPA area, width, and height (r = 0.421, 0.426, 0.345; all p < 0.01). A greater AL (β = 0.284, p < 0.01) and a smaller optic disc tilt ratio (β=–0.516, p < 0.01) were related to a larger PPA area. Conclusion The characteristics of the optic disc can be feasibly and efficiently extracted using AI. The quantization of the optic disc might provide new indicators for clinicians to evaluate the degree of myopia.


Introduction
Myopia has become a public health problem worldwide, with a strikingly high prevalence. 1 In some parts of East and Southeast Asia, the prevalence of myopia in young adults is approximately 80-90%. 2,3Myopic axial elongation results in a series of changes to the optic disc, including the development of parapapillary atrophy (PPA), 4 optic disc tilt, 5 and torsion, 6 and may cause conversion to a pathological myopia, such as glaucoma 7 and posterior staphyloma. 8ventually, pathological myopia can lead to irreversible vision loss and blindness. 9The high prevalence of myopia, especially in Asian populations, highlights the importance of analysing changes in the fundus, which may cause the deformations seen in pathological myopia.Clinical studies 10,11 have reported that with increasing myopia severity and axial length (AL), the area of PPA and the degree of optic disc tilting increase.As optic disc characteristics are easily visualized noninvasively on fundus images, a previous study demonstrated the relationships between the degree of myopia and optic disc tilt and the PPA area via manual identification of optic disc characteristics. 12Additionally, evidence has indicated that optic disc deformation and PPA formation are risk factors for myopia progression and the development of pathologic myopia. 12,13n ophthalmologist can analyse tiny changes in the optic disc characteristics with AL elongation using repeated images and then suggest the best treatment for preventing the development of pathological myopia.However, in current clinical practice, the identification and quantification of optic disc characteristics relies on manual observation by ophthalmologists, a time-consuming and laborious task for doctors to perform on an individual patient basis that requires a very high medical skillset.Additionally, as the shape of the optic disc is not a regular circle, subjective deviations in manually measuring the disc are inevitable. 14onsequently, an automatic segmentation system without the need for human intervention or manual delineation is highly desired.
In recent years, artificial intelligence (AI) diagnosis systems have emerged as powerful tools that can overcome the shortage of fundus screening programs. 15Previous studies have developed AI-based systems for the automated detection of optic disc abnormalities 16,17 and PPA. 18The application of AI in ophthalmology could provide more objective results than human graders, reduce operator dependence, and increase scalability and reproducibility, thereby potentially providing more accurate results. 157][18] However, to our knowledge, there is limited research investigating the quantitative assessment of the optic disc morphology and PPA area via AI, even though it is important for monitoring myopia progress.Thus, this study aimed to quantitatively assess the optic disc characteristics in fundus images using AI and explore the associations with myopia in young individuals.

Subjects and ocular examination
This population-based cross-sectional study was conducted in accordance with the tenets of the Declaration of Helsinki and approved by the Ethics Committee of the Eye Institute of the Shandong University of Traditional Chinese Medicine (HEC-KS-2021006KY).All participants were recruited and informed about the purposes and methods of the study, and each volunteer signed a consent form.
The study enrolled a total of 1084 young individuals using a stratified random sampling method from the Shandong University of Traditional Chinese Medicine in November 2021.Using stratified random cluster sampling, 6 classes (clusters) from the 4 colleges of the university (stratification) were randomly selected, with approximately 50 students in each class.The body mass index (BMI) was calculated using the following formula: weight (kg)/[height (m)] 2 .A detailed medical history was recorded for each participant.All participants underwent comprehensive ophthalmic examinations, including measurement of best-corrected visual acuity (BCVA), cycloplegic refraction, and intraocular pressure (IOP, NIDEK NT-510; NIDEK; JA-PAN), and slit-lamp microscopy of the anterior segment of the eye and axial length (AL) measurement using an IOLMaster (V5.0;Carl Zeiss Meditec AG; Jena; Germany).Refraction data were converted to spherical equivalent refraction (SER), which was defined as the spherical refractive power plus half of the cylindrical refractive power.The BCVA was converted into the logarithm of minimal angle resolution (logMAR).
For cycloplegia, the participants first received one drop of 0.4% oxybuprocaine (Santen Corp., Shiga, Japan) for topical anaesthesia.Two minutes later, the first of three drops of 1% cyclopentolate (Alcon, Fort Worth, Tex., USA) were instilled at 5-min intervals.If the pupil diameter was at least 6 mm 30 min later, refractometry was performed.If the pupil diameter was less than 6 mm, a fourth drop of cyclopentolate was applied.Autorefraction measurements were taken by autorefractor NIDEK (ARK-700A; NIDEK; JA-PAN).The right eye was selected for the analyses.
The inclusion criteria included the following:

Fundus photography
For all participants, fundus images were obtained for each eye with fundus photography (TRC-NW 400, Topcon, Japan).To obtain the optic disc morphology and PPA zone measurements in real size units, the magnification of the fundus image was corrected for each AL by applying Littmann's formula. 19The optic disc tilt was defined as the optic disc tilt ratio, the ratio of the minimum-to-maximum disc diameter; a tilted optic disc was defined as a tilt ratio of 0.80 or less. 20tic disc characteristics extracted and quantified using AI The optic disc morphology and the area of PPA were extracted and quantized from the fundus with an AI system that combines computer vision and deep learning image processing technology and integrates the colour, brightness, texture, shape and other characteristics of the optic disc and PPA presented in the fundus to perform quantitative analyses.First, the image was preprocessed, and then the optic disc characteristics were segmented by a deep learning segmentation model.Based on the results of the segmentation, the morphological parameters (maximum diameter of optic disc, minimum diameter of optic disc, PPA width, PPA height, etc.) were automatically measured.
The image preprocessing involves four steps: region of interest (ROI) establishment (Figure 1(B)), denoising (Figure 1(C)), normalization (Figure 1(D)), and enhancement (Figure 1(E)), 21,22 all of which aim to remove the interference of the background region on feature extraction and enhance the target features.Then, the optic disc is identified through the deep learning target detection network; thus, the position of the optic disc and PPA can be located, and the boundary of the optic disc is determined based on the mechanism of visual attention, 23 accurately segmenting the optic disc (Figure 1(G)) and extracting its maximum and minimum diameters.Finally, a deep learning model is used to segment the PPA area (Figure 1(H)) and optic cup (Figure 1(F)), thus identifying the characteristic zone of the PPA and optic cup.
For the segmentation of the optic disc, the single shot detection (SSD) object detection model is first used to locate the position of the optic disc in the fundus image and draw the bounding box.Then, the edge of the optic disc is detected in the polar coordinate-transformed image and inverted to the rectangular coordinate system to obtain the segmentation region of the optic disc. 14The purpose of segmenting the optic cup and PPA is to allow the preprocessed fundus images to be input into the corresponding trained semantic segmentation model ResUnet, which then outputs a binary image of the optic cup and PPA.ResUNet combines residual connections and a U-Net-shaped model, the former to simplify training and promote full propagation of feature maps in both advanced and low-level stages, while the latter is incorporated to leverage its powerful feature extraction capability.ResUNet includes a coding layer, bridge layer and decoding layer; the coding layer consists of a convolutional group (including a convolutional layer, batch normalization, and rectified linear unit, ReLU) composed of 3 sets of residual connections.Symmetrically with the coding layer, the decoding layer also consists of 3 convolutional groups, where the input size for each convolutional group is the output size of the corresponding coding layer.The coding and decoding layers are connected by a skip connection.
The accuracy, sensitivity, and specificity are calculated by comparing the pixel-based annotation with manual annotation. 24The three metrics were used to test the performance of the segmentation model.Both the recognition and segmentation accuracy of the optic disc and PPA were 0.998, the sensitivity of the optic disc segmentation and PPA segmentation were 0.969 and 0.874, respectively, and the specificity was 0.999.

Statistical analysis
Statistical analyses were performed using the SPSS software package (SPSS for Windows, V. 25.0; SPSS, Inc., Chicago, Illinois, USA).Continuous variables are shown as the mean ± SD.The normality of the variable distribution was verified using the Shapiro-Wilk normality test.One-way analysis of variance (ANOVA), the nonparametric K-W test and the X 2 test were used to examine differences among the five groups.Pearson's and Spearman's correlation analyses were performed to determine the simple linear correlation between variables, where "r" is the correlation coefficient.Multivariate linear regression analyses were used to identify explanatory variables with a statistically significant contribution to the PPA area.Statistical significance was considered at p < 0.05.

Subject characteristics
Among the 1084 subjects enrolled in the study, 62 were excluded because of a history of excimer laser surgery, 11 for having poor-quality fundus images, 10 for having an IOP higher than 21 mm Hg, and another 2 for having an SER less than 10 dioptres (D).Finally, 999 (92.16%) subjects were included in the analysis, including 674 (67.47%) females and 325 (32.53%) males.
Based on the AL, the 999 subjects were divided into five groups: Group 1 (n ¼ 25, 2.5%), Group 2 (n ¼ 131, 13.1%), Group 3 (n ¼ 253, 25.3%), Group 4 (n ¼ 319, 31.9%), and Group 5 (n ¼ 271, 27.1%).The demographic and ocular characteristics of the included subjects are presented in Table 1.There were no significant differences between the groups in terms of age, BMI or BCVA (all p > 0.05).However, compared with Group 1 and Group 2, the other groups had a lower frequency of females, lower SER, greater AL, and higher IOP (all p < 0.05).

Automated quantitative analysis of optic disc morphology
The optic disc morphologies are shown in Table 1.There were 471 (47.1% sample prevalence) cases of tilted optic discs.Figure 2 intuitively shows that after stratifying by myopia severity, there were statistically significant differences in the prevalence of optic disc tilting.The average optic disc area, cup-to-disc (C/D) ratio, and optic disc tilt ratio were 1.97 ± 0.46 mm 2 , 0.18 ± 0.07, and 0.81 ± 0.08, respectively.Split violin plots (Figure 3) show the distribution of the optic disc area in the five groups.The median optic disc area was smaller in Group 5 than in the other groups.Table 2 demonstrates the associations between sex, BMI, IOP, SER, AL and optic disc morphology characteristics.There were no significant associations between BMI and disc area, C/D ratio, or optic disc tilt ratio (all p > 0.05).Male sex was positively associated with the disc area, C/D ratio, and optic disc tilt ratio (r ¼ 0.143, p < 0.001; r ¼ 0.150, p < 0.001; r ¼ 0.152, p < 0.001, respectively).There were no significant associations between IOP and disc area, C/D ratio, or optic disc tilt (all p > 0.05).SER was positively correlated with the disc area, C/D, and optic disc tilt ratio (r ¼ 0.296, p < 0.001; r ¼ 0.076, p ¼ 0.016; r ¼ 0.227, p < 0.001, respectively).AL was negatively correlated with the disc area and optic disc tilt ratio (r¼-0.271,p < 0.001; r¼-0.219,p < 0.001, respectively).However, AL was not significantly associated with the C/D ratio (r¼-0.014,p ¼ 0.663).
After stratifying by myopia severity, Figure 4 shows the relationship between the AL and the optic disc area and optic disc tilt ratio among the 5 groups.Although a negative correlation between optic disc area and AL was only present in Group 3, the optic disc area presented a decreasing trend in the first four groups, and in Group 5, a slightly increasing trend was shown.In eyes with an AL less than 24 mm, there was no significant negative correlation between optic disc tilt and AL, and in Group 3, there was an obvious negative relation between the optic disc tilt ratio and AL.However, in eyes with an AL more than 25 mm, the degree of optic disc tilt tended to be stable.

Automated quantitative analysis of PPA
The characteristics of PPA are shown in Table 1.There were 924 (92.5% sample prevalence) subjects of PPA.The PPA area, width, and height were 0.84 ± 0.59 mm 2 , 0.44 ± 0.24 mm, and 1.61 ± 0.56 mm, respectively.In addition, Figure 2 intuitively shows that after stratifying by myopia severity, there were statistically significant differences in the prevalence of PPA among the groups.The split violin plots (Figure 3) show the distribution of PPA area among the five groups.The median PPA area increased with greater AL.Table 2 presents the correlation between sex, BMI, IOP, SER, AL and PPA characteristics.There were no significant relations between BMI and IOP and PPA area, width, or height (all p > 0.05).Male sex was negatively associated only with the PPA width (r=-0.080,p ¼ 0.011).SER was revealed to be negatively correlated with PPA area, width, and height (r¼-0.287,p < 0.001, r¼-0.420,p < 0.001, r¼-0.316,p < 0.001, respectively).AL was positively correlated with PPA area, width, and height (r ¼ 0.292, p < 0.001, r ¼ 0.426, p < 0.001, r ¼ 0.345, p < 0.001, respectively).
After stratifying by myopia severity, Figure 4 shows the correlation between AL and PPA area among the 5 groups.Overall, the area of PPA increased with greater AL.Although a significant correlation between the AL and PPA was presented only when the AL was longer than 24 mm, there was a significant increasing trend in the PPA area with increasing AL.

Correlations between optic disc morphology and PPA characteristics
Table 3 shows the correlation between disc morphology and PPA characteristics.The PPA area was negatively correlated with the disc area, C/D ratio, and disc tilt ratio (r¼-0.267,p < 0.001; r¼-0.051,p ¼ 0.017; r¼-0.425,p < 0.001, respectively).PPA width and height were negatively correlated with the disc area and disc tilt ratio (all p < 0.001) and positively correlated with the C/D ratio (all p < 0.001).Furthermore, we constructed a multivariate linear regression model to explore the independent factors associated with PPA area.The multivariate model showed that AL and optic disc tilt ratio were independently associated with PPA area (both p < 0.001).According to the model, every 1 mm increase in AL was associated with a 0.14 mm 2 increase in PPA area, and every 0.1 increase in the disc tilt ratio was associated with a 4.045 mm 2 decrease in PPA area.The combination of these factors yielded an adjusted R 2 of 0.424 (Table 4).After stratifying by myopia severity, the multiple regression model showed that the optic disc tilt correlated negatively with the PPA area in all groups.The overall R 2 values derived from the multiple regression model for the PPA area were 0.335, 0.368, 0.401, 0.292 and 0.331 (Table 4).

Discussion
To our knowledge, this is the first study to quantitatively extract optic disc characteristics from fundus images using AI and explore their influencing factors.In our basic cross-   sectional study, there were high incidence of PPA (92.5%) and optic disc tilting (47.1%) in young individuals, and the prevalence of both significantly increased with higher degrees of myopia.Participants with a greater AL were found to have a significantly smaller optic disc area and optic disc tilting ratio and a larger PPA area.The multilinear regression model showed that a greater AL (p < 0.01) and a larger degree of optic disc tilt (p < 0.01) were related    to a larger PPA area.This is a clinically significant observation given that myopia-related optic disc characteristics may lead to pathologic myopia and, eventually, irreversible visual impairment, suggesting that it is crucial to focus on the optic disc characteristics of young myopic patients at an early stage.
When leveraged as an automatic screening tool for myopia, fundus-based AI systems could have significant benefits when embedded within new clinical workflows.In recent years, to improve the efficiency and accuracy of optic disc morphology and PPA zone identification and detection tasks, many automatic identification methods based on AI have been proposed.Deep learning has become an efficient and effective methodology for fundus image segmentation tasks. 257][28][29] This study proposed a more efficient method for segmenting optic disc characteristics from fundus images using the a ResUNet model, which combines the ResNet and U-Net models.][31][32] However, although the U-Net architecture is the current state-of-the-art owing to its training simplicity and data efficiency, its multiscale skip connection tends to use unnecessary information, and low-level encoder features are insufficient, leading to a relatively poor performance. 24dditionally, ResNet is a simplified version of residual blocks that uses artificial neural networks.In residual blocks, the skip connections principle simplifies and accelerates the deep learning process in complex networks. 33In contrast, ResUNet al.lows comprehensive standby of convolutional blocks, which is substantially more effective in terms of training time, memory usage, and accuracy than baseline methods. 34ur study demonstrated the relationships between optic disc characteristics and AL.The overall negative correlation between the optic disc area and AL is consistent with the findings of previous studies. 12,35The reason for this result may be due to the fact that part of the nasal optic disc is covered by an overhanging Bruch's membrane when the optic disc is distorted by mechanical stretching of the eyeball. 36Additionally, after stratifying by myopia severity, we explored the relationship between optic disc area and AL for different groups of AL.We further found that the optic disc area was slightly increased in Group 5.This result is similar to that of previous studies. 37,38The optic disc enlarges with greater axial length or more myopic refractive error, starting at a cut-off value of approximately À8 dioptres or an axial length of approximately 26.5 mm.Consistent with the results of previous studies, 39,40 we also found that a greater AL was correlated with a larger PPA area and a higher degree of optic disc tilt.During asymmetrical elongation of the posterior ocular segment, the skewed insertion of the optic nerve into the globe causes optic disc tilt, 41 and the progressive thinning of the choroid, disappearance of choroidal vessels, and loss of RPE and photoreceptors cause PPA area enlargement. 42Additionally, our study demonstrated that there was no significant correlation between the degree of optic disc tilt and AL in Group 5. The asymmetric elongation of the eyeball may cause the optic disc to experience an inconsistent direction of traction, so the degree of optic disc tilt is compensated in the progress of eye elongation.Therefore, research on the traction direction of the optic disc during myopia needs to be further investigated.
The potential relationship between optic disc tilt and the PPA area was explored in a previous study. 43In this crosssectional observational study, we quantitatively assessed the optic disc morphology and PPA.We demonstrated that a higher degree of optic disc tilt was significantly correlated with a larger PPA area.Recently, several studies have reported that both tilted discs and PPA were risk factors for myopia developing into pathologic myopia. 44A larger PPA area was correlated with the progression of myopia maculopathy, 44,45 myopic retinopathy, 46 and worsening glaucoma. 47A higher degree of optic disc tilt was also correlated with worse staphyloma. 48Although most young participants in the present study did not have pathologic myopia, a larger PPA area and a higher degree of optic disc tilt with greater AL were observed.During the elongation of the eye, the asymmetrically stretched fundus, tilted disc and PPA formation exacerbate the degree of choroid thinning. 12,43ence, close monitoring of myopia-related optic disc characteristic changes and the relationship between optic disc morphology characteristics and the PPA area are essential for better managing myopia patients.
The prevalence of tilted optic discs in the current study sample was 47.1%, similar to that reported in relatively recent population-based studies by Tong (44.9%) 49 and Samarawickrama (37%). 50The incidence of PPA has been studied in several population-based studies: 72% in African eyes, 64% in European eyes, 51 and 72.6% in Japanese eyes. 52n a cross-sectional study in China, Li et al. 5 reported that the proportions of optic disc tilting and b-zone peripapillary atrophy were 81.2% and 92.8%, respectively.Chen et al. 43 found that the prevalence of tilted optic discs and PPA in a university-based study in China were 56.6% and 79.9%, respectively.A recent study by Zhang et al. 53 reported that the prevalence of optic disc tilt was 37.5% in young Chinese adults.The difference between these studies in terms of the prevalence of tilted optic disc and PPA may be partly explained by differences in the populations, the different definitions of optic disc tilting, the different subzones for identifying PPA and differences in the research methods used to extract optic disc characteristics from fundus images.First, in our study, we sampled a cohort of young college students with a relatively high prevalence of myopia and with a higher severity of myopic refractive errors.Second, the definition of optic disc tilting was the ratio between the shortest and longest diameters of the optic disc, which was recently widely adopted.Third, we applied an AI model to identify and extract optic disc characteristics, which can overcome the effects of the subjective deviation in manual measurement of optic disc characteristics by ophthalmologists.Finally, the present study did not identify the subzone for measuring PPA.In addition, our results showed that the prevalence of PPA was more than 50% in eyes with ALs less than 24 mm, and its area was significantly smaller than that in eyes with ALs longer than 24 mm.This may suggest that pathological PPA can be accurately defined through accurate identification and quantitative extraction of the PPA area based on AI.
The present study had limitations.First, it was a university-based study; our participants were enrolled from the Shandong University of Traditional Chinese Medicine, which has a higher proportion of female than male.Additionally, there were higher rates of myopia prevalence in females compared with males.As shown in supplementary table 1, epidemiological studies of the general population presented the difference in the prevalence of myopia between female and male.Thus, the statistical power of the relationships between optic disc characteristics and sex was limited.Second, the sample size of Group 1 was small; hence, the results might not be truly representative of the whole population.Finally, this was a cross-sectional study, and the cause-and-effect relationship between the evolution of optic disc characteristics and the increase in AL could not be established.Longitudinal studies are needed to explore this relationship further.
In conclusion, deep learning has become an efficient and effective methodology for a wide variety of ophthalmological image segmentation tasks.As changes in optic disc morphology characteristics and increases in PPA area could reflect the exacerbation of myopia severity, timely follow-up of these optic disc characteristics can help monitor the progress of myopia.Applying AI to automatically and quantitatively detect optic disc characteristics can play a critical role in regular fundus examination of the entire population at-risk of myopia due to accessibility.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Figure 1 .
Figure 1.Automated recognition and segmentation of optic disc characteristics from fundus images by AI. (A) Original image; (B) Image showing the automatic recognition of optic disc area (amaranth), maximum optic disc diameter (line a), minimum optic disc diameter (line b), optic cup (yellow), PPA area (violet), PPA height (line h) and PPA height (line h); (C-E) Automated segmentation of optic disc, optic cup and PPA.

Figure 2 .
Figure 2. The prevalence of optic disc tilting and PPA in different groups.(A) The prevalence of optic disc tilting, (B) the prevalence of optic disc tilting in the PPA area.

Figure 4 .
Figure 4. Correlation between optic disc morphology, PPA area and AL among the 5 groups.(A) The relationship between optic disc area and AL among the 5 groups, (B) the relationship between optic disc tilt ratio and AL among the 5 groups, and (C) the relationship between PPA area and AL among the 5 groups.

Table 1 .
Demographics and clinical characteristics.

Table 2 .
Correlation analysis between optic disc characteristics, PPA and ocular parameters.

Table 4 .
Multivariate regression analysis of associations with PPA.