A Novel Artificial Intelligence-Based Classification of Highly Myopic Eyes Based on Visual Function and Fundus Features

Abstract

Purpose: To develop a novel classification of highly myopic eyes using artificial intelligence (AI) and investigate its relationship with contrast sensitivity function (CSF) and fundus features.

Methods: We enrolled 616 highly myopic eyes of 616 patients. CSF was measured using the quantitative CSF method. Myopic macular degeneration (MMD) was graded according to the International META-PM Classification. Thickness of the macula and peripapillary retinal nerve fiber layer (p-RNFL) were assessed by fundus photography and optical coherence tomography, respectively. Classification was performed by combining CSF and fundus features with principal component analysis and k-means clustering.

Results: With 83.35% total variance explained, highly myopic eyes were classified into four AI categories. The percentages of AI categories 1 to 4 were 14.9%, 37.5%, 36.2%, and 11.4%, respectively. Contrast acuity of the eyes in AI category 1 was the highest, which decreased by half in AI category 2. For AI categories 2 to 4, every increase in category led to a decrease of 0.23 logarithm of the minimum angle of resolution in contrast acuity. Compared with those in AI category 1, eyes in AI category 2 presented a higher percentage of MMD2 and thinner temporal p-RNFL. Eyes in AI categories 3 and 4 presented significantly higher percentage of MMD ≥ 3, thinner nasal macular thickness and p-RNFL (P < 0.05). Multivariate regression showed AI category 4 had higher MMD grades and thinner macular compared with AI category 3.

Conclusions: We proposed an AI-based classification of highly myopic eyes with clear relevance to visual function and fundus features.

Translational Relevance: This classification helps to discover the early hidden visual deficits of highly myopic patients, becoming a useful tool to evaluate the disease comprehensively.

Introduction

High myopia has become one of the leading causes of visual impairment worldwide.^1–3 With elongation of axial length (AL) in these eyes, a series of retinal and choroidal pathologies can occur and can be assessed using advanced structural measurements such as ultra-wide fundus photography and optical coherence tomography (OCT).^4,5 

In 2015, Ohno-Matsui et al proposed the now widely used META-analysis for Pathologic Myopia (META-PM) photographic classification system to grade myopic macular degeneration (MMD) in highly myopic eyes based on fundus photographs.^6,7 Significant correlations have been found between classification based the system and vision loss.⁸ Research has reported a higher magnitude of vision deterioration in eyes with MMD grades 3 or 4 than those without.^9,10 Wang et al¹¹ found that the presence of MMD grade 4 was associated with increased risk of visual impairment during 5-years of follow-up. Nevertheless, subtle variations in visual function are difficult to detect from fundus photographs, and eyes with similar fundus morphology can sometimes exhibit very different visual functions. Zheng et al¹² found the association between visual acuity and MMD severity was not significant. Visual difference between eyes with mild maculopathy, such as MMD grades 1 and 2, were not observed in some studies.^9–12 Therefore, a classification system that combines features from both visual function and the retinal structure is necessary to evaluate highly myopic eyes comprehensively. 

Because high myopia is a slowly progressive ocular disease, best-corrected visual acuity (BCVA) is usually not affected until significant pathological changes involving the central macula occur. BCVA changes slowly, and such slow change of BCVA may lead to underestimate of visual disability. Compared with BCVA, contrast sensitivity function (CSF) may be a more sensitive tool to evaluate visual function of highly myopic eyes, because it correlates better with subjectively perceived visual impairment and detects more subtle functional changes at an earlier stage of several ocular disorders, such as age-related macular degeneration, diabetic retinopathy, and glaucoma.^13–16 Yet methodological limitations of CSF testing have hindered its application.¹⁷ Recently, the novel Bayesian adaptive quantitative CSF testing, implemented on the Manifold Contrast Vision Meter platform (Adaptive Sensory Technology, San Diego, CA), can provide a highly precise assessment of the entire CSF in 25 trials within only 3 to 5 minutes.^18,19 

Furthermore, artificial intelligence (AI) has been applied extensively in the diagnosis of ocular diseases and prediction of visual prognosis based on ophthalmic images. AI approaches to automatically detect and grade myopic maculopathy from color fundus photos and OCT macular images have been developed in the previous literature, showing high sensitivity and perhaps helping doctors to make preliminary diagnoses of different degrees of myopic maculopathy.^20–23 Using data from nine multiethnic cohorts and six regions, Tan et al²⁴ developed a retinal photograph-based deep learning algorithm with good performance in the diagnosis of high myopia and MMD. Moreover, researchers have also made efforts to predict visual loss from fundus photographs or OCT images with various AI models, such as support vector machines, logistic regression, and Resnet.^11,25 Thus, AI can serve as an effective tool for risk stratification and grading of high myopia.²⁶ Previously, early loss of CSF was observed in simple high myopia.²⁷ Developing AI-based classification models integrating imaging features and CSF variables may bring new insights into AI in pathological myopia, but it has never been implemented. 

The purpose of the current study was to propose a novel AI-based classification of highly myopic eyes by integrating features from CSF and fundus features. 

Methods

This observational study was approved by the Institutional Review Board of the Eye and Ear, Nose, Throat (EENT) Hospital of Fudan University, Shanghai, China, and was conducted in accordance with the tenets of the Declaration of Helsinki. It was affiliated with the Shanghai High Myopia Study launched at the EENT Hospital of Fudan University since 2015 (www.clinicaltrials.gov, NCT03062085). Written informed consent was obtained from each patient before participation in the study. 

Patients

The Shanghai High Myopia Study continuously enrolled highly myopic patients (AL ≥ 26.0 mm) aged ≥18 years in the EENT Hospital of Fudan University. The medical records of the patients in the study database were reviewed. Highly myopic eyes with CSF data, clear fundus photographs and good quality OCT scans (quality score of > 30 and well-focused images with no link, motion or doubling artifacts) of the macula and optic nerve were included in this study. The CSF data and fundus and OCT images were collected after cataract surgery. Exclusion criteria were corneal opacity, glaucoma, uveitis, fundus pathology other than high myopia-related changes, lens opacity, multifocal intraocular lens implantation, previous trauma or vitrectomy, and systemic disorders such as diabetes. When both eyes of a patient fit the criteria, we randomly selected one eye from the patient, whereas for patients with only one qualified eye, that eye was selected. Ultimately, a total of 616 highly myopic eyes of 616 patients were used as original dataset. 

Ophthalmic Examinations

All patients underwent a comprehensive ophthalmic examination, including evaluation of BCVA (logarithm of the minimum angle of resolution [logMAR]), a slit-lamp examination, CSF testing (Manifold Contrast Vision Meter, Adaptive Sensory Technology, San Diego, CA), fundus photography, OCT examination (Spectralis, Heidelberg Engineering, Germany), B-scan ultrasound examination, and AL measurements (IOLMaster 700, Carl Zeiss AG, Oberkochen, Germany). 

Acquisition of CSF Data

The CSF testing was used to estimate the two-dimensional CSF curve, which describes contrast sensitivity as a function of spatial frequency.^18,19 Patients wearing best-correcting spectacles with the nontest eye patched were tested at a viewing distance of 3 m with background luminance of approximately 90 cd/m² on the screen. The area under the logarithmic of the CSF (AULCSF) representing CSF integrated between 1.5 to 18.0 circles per degree (cpd), served as a summary metric of the CSF. Other CSF outcomes were also recorded, including contrast sensitivities at 1, 1.5, 3, 6, 12, and 18 cpd, and contrast acuity, which represents the smallest detectable optotype at 100% contrast (the intersection of the CSF curve with the x axis). 

Grading of MMD

MMD was graded independently by two doctors (W.H. and J.M.) based on fundus photographs, according to the International META-PM Classification System,⁶ without access to any clinical information. Any disagreement was resolved by a senior eye specialist (X.Z.). MMD of grade ≥ 3 was regarded as severe MMD. 

Measurement of Beta-Zone Peripapillary Atrophy

The beta-zone peripapillary atrophy (β-PPA), a peripapillary crescent area of chorioretinal atrophy with visible choroidal vessels and sclera, was assessed on fundus photographs with the ImageJ software (V.1.47, NIH, Bethesda, MD). The area of β-PPA was defined as the β-PPA area/optic disc area ratio. The point of maximum radial extent (PMRE) was defined as the point on the temporal β-PPA margin with the maximal radial extent. The angle of point of maximum radial extent was defined as the angle between the reference line from the center of optic disc to fovea and the line from the center of optic disc to the point of maximum radial extent.²⁸ 

Measurement of Macular and p-RNFL Thickness

Macular and p-RNFL thickness were assessed with OCT (Spectralis, Heidelberg Engineering, Germany). High-speed B-scans were acquired in a continuous, automated sequence, covering a 30° × 25° area centered on the fovea. To obtain good-quality images, ≥20 frames were averaged automatically for each scan. A builtin software of the device automatically generated color-coded macular thickness maps by applying the Early Treatment Diabetic Retinopathy Study grid on the fovea, which divided the macula into three concentric rings: center (1 mm in diameter), inner (1–3 mm in diameter), and outer (3–6 mm in diameter). The inner and outer rings were further divided into four quadrants: nasal, temporal, superior, and inferior. Macular thicknesses in all Early Treatment Diabetic Retinopathy Study subfields were recorded. 

p-RNFL thickness was measured on a 12° circular B-scan centered on the optic nerve head. p-RNFL thickness in the six subfields (nasal, nasal–superior, nasal–inferior, temporal, temporal–superior, and temporal–inferior) as well as the global average p-RNFL thickness were recorded. 

Scan quality and automatic segmentation were assessed prior to analysis. Any error of automated layer segmentation was manually corrected by a trained doctor (Y.C.). 

Proposed Classification Based on Principal Component Analysis and Clustering Analysis

To generate the classification system, multidimensional features were selected, including demographic information (age, AL), functional vision features (AULCSF, contrast acuity, CS values at 1–18 cpd, BCVA), and structural features (MMD grade, PPA area, global and temporal RNFL thickness, inner nasal and outer nasal macular thicknesses). After importing and standardizing the data, principal component analysis (PCA) was applied with six principal components. Several clustering algorithms with three to four clusters were evaluated, including k-means clustering, Gaussian mixture model, hierarchical clustering, and density-based spatial clustering of applications with noise. Total variance explained and explained variance ratio of each principal component were examined. Two-dimensional projections of the PCA solutions, with clusters labeled by different colors, were created. The best classification model was selected based on the quality of the clustering pattern and higher variance explained ratio. 

Statistical Analyses

Continuous data are presented as means ± standard deviations. The PCA and clustering analyses were conducted using the Python programs and use of the classification system is available on our website (URL: https://hm-zai.com/). A one-way analysis of variance (ANOVA) with Tukey's post hoc test was used to compare continuous data among groups. A χ² test was used to compare categorical data. Multivariate logistic regression analysis was used to identify the independent influencing indicators of different categories including BCVA, AULCSF, MMD grades, area of β-PPA, central macular thickness and global average p-RNFL thickness after adjusting for age, sex, eye laterality, and AL. Statistical analyses were performed with SPSS version 22 (IBM Corp., New York). Two-sided P value of < 0.05 were considered statistically significant in all analyses. 

Results

Characteristics

Of the 616 patients included, 287 were male and 329 were female, with a mean age of 62.8 ± 9.3 years. The mean AL was 29.08 ± 2.16 mm, (range, 26.01–36.42 mm). 

AI-Based Classification of Highly Myopic Eyes

Based on both visual function and fundus features, the highly myopic eyes were classified into four categories with PCA and k-means clustering. Principle components 1 to 6 explained 47.2%, 10.6%, 7.8%, 6.2%, 5.9%, and 5.7% of the total variance, respectively. The AI-based classification explained 83.35% of the total variance. Figure 1 shows the two-dimensional projections of the PCA solution with principal component 1 (x axis) and 2 (y axis). The three-dimensional projections with principal component 1 (x axis), 2 (y axis) and 3 (z axis) are shown in Supplementary Figure S1. A description of the principal components is presented in Supplementary Table S1. For evaluation of clustering performance of the classification model, the Silhouette score, Calinski–Harabasz index, and Davies–Bouldin index were 0.26, 243.00, and 1.38, respectively. 

Figure 1.

 

Two-dimensional projection of the PCA solutions with PC 1 and PC 2. Highly myopic eyes of the AI category 1 to 4 were labeled as green, yellow, red, and blue circles, respectively.

Of the highly myopic eyes, 14.9% (92/616), 37.5% (231/616), 36.2% (223/616), and 11.4% (70/616) were classified into AI categories 1 to 4, respectively. Table 1 displays the characteristics of the four AI categories. Patients in AI category 1 were significantly younger, whereas those in AI category 4 were significantly older, compared with the other two categories (one-way ANOVA with Tukey's post hoc analysis, all P < 0.05). No significant difference was found in sex or eye laterality among groups (one-way ANOVA with Tukey's post hoc analysis; both P > 0.05). Eyes in AI categories 1 and 2 presented significantly shorter ALs and better BCVA compared with those in AI category 3, and those in AI category 4 presented the longest AL and worst BCVA (one-way ANOVA with Tukey's post hoc analysis; all P < 0.05). 

Table 1.

 

Characteristics of the Highly Myopic Eyes in Different AI Categories

 

Associations Between the Classification and Visual Function

Eyes in AI category 1 presented the AULCSF and contrast acuity, followed by those in AI category 2, and then AI category 3. The worst CSF was identified in AI category 4 (one-way ANOVA with Tukey's post hoc analysis; all P < 0.05) (Table 1). The AULCSF loss was 0.38 log unit between AI categories 1 and 2, 0.56 log unit between AI categories 2 and 3, and 0.19 log unit between AI categories 3 and 4. No significant difference was found between AI categories 1 and 2 in BCVA, whereas AI category 2 showed significantly worse contrast acuity than that of AI category 1 (P < 0.05) (Table 1). Compared with AI category 2, AI category 3 showed a decrease in BCVA by 0.13 logMAR, whereas in contrast acuity by 0.23 logMAR. Compared with AI category 3, AI category 4 showed a decrease in BCVA by 0.66 logMAR and in contrast acuity by 0.23 logMAR. Compared with BCVA, CSF differentiated visual change between either two of the four AI categories, and also exhibited better performance in eyes with mild maculopathy. 

Figure 2 shows the average CSF curves of the eyes in the four AI categories. For eyes in AI category 1, the curve started with a high contrast sensitivity at 1 cpd, increased slightly, peaked at 3 cpd, and then decreased gradually at higher spatial frequencies. Eyes in AI category 2 had significantly lower contrast sensitivities than those in AI category 1 when the spatial frequency was >1 cpd (one-way ANOVA with Tukey's post hoc analysis; all P < 0.05). Eyes in both AI categories 3 and 4 exhibited significantly lower contrast sensitivities than AI category 2 at all spatial frequencies except 18 cpd (one-way ANOVA with Tukey's post hoc analysis; all P < 0.05). Eyes in AI category 4 also had significantly lower contrast sensitivities than those in AI category 3 at spatial frequencies of 1 to 6 cpd (one-way ANOVA with Tukey's post hoc analysis; all P < 0.05). 

Figure 2.

 

Curves of CSF in highly myopic eyes of different AI categories. cpd, cycle per degree.

Associations Between the Classification and Fundus Features

Figure 3 shows the distribution of MMD grades of the eyes in AI different categories. Eyes in AI category 1 presented significantly higher percentage of MMD grade 0 and lower percentage of MMD grade 2 than those in AI category 2 (χ² test; both P < 0.05). Compared with eyes in AI category 2, eyes in AI category 3 presented significantly lower percentage of MMD grades 0 and 1 and higher percentage of MMD grade 2 (χ² test; all P < 0.05). In addition, there was no eye with MMD grades 3 to 4 found in AI categories 1 and 2. Significantly higher percentage of MMD grade 4 was identified in eyes in AI category 4 compared with those in AI category 3 (χ² test; P < 0.05). Disagreements between MMD grades and AI-based categories were noted. Among patients with MMD grade 1, eyes in AI category 2 (168/234 [72%]) presented significantly worse AULCSF compared with those in AI category 1 (66/234 [28%]), showing a difference of 0.34 log unit in AULCSF. Among patients with MMD grade 2, eyes in AI category 3 (126/174 [72%]) presented significantly worse AULCSF and contrast acuity compared with those in AI category 2 (48/174 [28%]). The AULCSF difference was 0.44 log unit, and contrast acuity difference was 0.16 logMAR. Among patients with MMD grade 3, eyes in AI category 3 (84/120 [70%]) presented significantly better AULCSF compared with those in AI category 4 (36/120 [30%]), showing a difference of 0.11 log unit in AULCSF. 

Figure 3.

 

The percentages of eyes with each MMD grade in different AI categories.

The β-PPA features among eyes of different AI categories are compared in Table 2. Eyes in AI categories 3 and 4 had significantly larger area of β-PPA than those in AI categories 1 and 2 (one-way ANOVA with Tukey's post hoc test; P < 0.05). The representative fundus photographs of highly myopic eyes in different AI categories from the novel classification are shown in Figure 4. 

Table 2.

 

The β-PPA Features in Highly Myopic Eyes of Different AI Categories

 

Figure 4.

 

Representative fundus photographs and the corresponding CSF data of highly myopic eyes in different AI categories. β-PPA was outlined manually (yellow line) and calculated as the β-PPA area/optic disc area ratio. The point of point of maximum radial extent (PMRE) was the point on the temporal β-PPA margin at which the radial extent of β-PPA was greatest. The angle of PMRE (blue) was the angle between the red line (connecting the center of the optic disk and PMRE) and the black line (connecting the center of the optic disc and fovea). PMRE, point of maximum radial extent.

Figure 5 displays the thickness of macula and p-RNFL of eyes in different AI categories. Compared with eyes in AI categories 1 and 2, eyes in AI category 3 presented significantly thinner nasal macular thickness (one-way ANOVA with Tukey's post hoc test; all P < 0.05). Eyes in AI category 4 presented significantly thinner central, inner, and outer nasal macular thickness compared with the other AI categories (one-way ANOVA with Tukey's post hoc test; all P < 0.05). In terms of p-RNFL thickness, eyes in AI category 2 showed significantly thinner p-RNFL thickness at the temporal subfield than those of AI category 1 (one-way ANOVA with Tukey's post hoc test; P < 0.05). Eyes in AI categories 3 and 4 showed significantly thinner p-RNFL thickness at all subfields except the temporal-superior subfield, compared with those in AI category 1 (one-way ANOVA with Tukey's post hoc test; all P < 0.05). The global average p-RNFL thickness and p-RNFL thickness at temporal, temporal–inferior, and nasal subfields were even lower in eyes in AI category 4 compared with those in AI category 3 (one-way ANOVA with Tukey's post hoc test; all P < 0.05). 

Figure 5.

 

The thickness of macula and p-RNFL and the corresponding CSF data among highly myopic eyes of different AI categories. (A) The thickness of macula at the nine subfields of ETDRS grids in eyes of different categories. (B) The p-RNFL thickness at the six subfields around the optic disk and the global average p-RNFL thickness in eyes of different categories. (C) The corresponding CSF curves in eyes of different categories. ETDRS, Early Treatment Diabetic Retinopathy Study; G, global; I, inferior; N, nasal; S, superior; T, temporal.

Multivariate Logistic Regression of Different AI Categories

Supplementary Table S2 shows the details of multivariate logistic regression analysis on different AI categories. Eyes in AI category 2 were associated with lower AULCSF compared with those in AI category 1. Compared with those in AI category 2, eyes in AI category 3 had significantly longer AL, lower AULCSF, larger β-PPA area, and thinner global RNFL. Compared with those in AI category 3, eyes in AI category 4 were significantly associated with lower AULCSF, worse BCVA (odds ratio, 1013.95; P = 0.001), more severe MMD grades, and thinner macular thickness (all P < 0.05). 

Discussion

In 2015, Ohno-Matsui et al proposed the META-PM photographic classification system based on fundus morphology, serving as a very useful tool to trace the natural course and describe the epidemiologic features in different populations, such as Chinese, German, and Japanese cohorts. It also facilitates communication and comparison of findings from clinical trials and genetic analyses.^6–8 Researchers have developed deep learning models with fundus images to assist in the detection of the META-PM grades and subsequent protection of patients against visual impairments owing to maculopathy. The correlation between MMD grade 4 and long-term risk of visual loss has been reported.¹¹ Nevertheless, subtle variations in visual function may be observed in eyes with similar fundus morphology and cannot be detected easily from fundus photos. In the current study, we proposed a novel AI-based classification for highly myopic eyes by integrating features from CSF and fundus features, which in clinical practice may assist to distinguish different degrees of function–structure combinations for pathological myopia comprehensively. 

Thanks to the advances in imaging and AI methods, computational efforts have been made to improve the grading of pathological myopia and further to monitor disease progression or guide clinical therapies. Some studies trained AI systems with fundus images or OCT scans to achieve automatic detection and classification of high myopia.^20–25 Other investigators discovered new MMD-related features through AI methods for the better prediction of the risk of maculopathy progression.²⁹ In contrast, structure–function analysis for pathological myopia has drawn attention owing to its translational importance. Studies found visual acuity performs poor in differentiate mild MMD, although the more sensitive index CSF has already been impaired when visual acuity is normal. In the study, CSF differed significantly between AI categories 1 and 2 but BCVA did not. Also, greater change in CSF was noted across the categories compared with that in BCVA, suggesting the valuable contribution of quantitative CSF to structure–function integration. 

Our classification combined multidimensional parameters including demographic characteristics, functional indexes, and structural features from fundus and OCT images. MMD severity is one of the structural parameters. More early stage structural parameters, such as macular and p-RNFL thickness, are also taken into consideration. In clinical practice, visual impairment or OCT alterations may occur before significant MMD on fundus photos. Conversely, some patients may still maintain a certain degree of vision though their fundus images are graded into severe MMD. Those cases may be redefined and regraded by our classification. Interestingly, disagreements between MMD grades and AI categories were noted, accounting for 30% in eyes with MMD grade 1, grade 2, or grade 3. Even if patients were classified into the same MMD grade, those with a higher level of AI categories tend to exhibit more risk of visual impairment than the others. These findings also imply that the classification may have the potential to serve as a useful complement for the previous structural classification system of high myopia. 

Compared with visual acuity, CSF, describing contrast sensitivity in a wide range of spatial frequencies, can act as a more sensitive tool to capture the hidden losses of vision.^14–16,27 The recently developed quantitative CSF testing, with both high sensitivity and robust test–retest reliability,^18,19 has become a useful method for detecting early stages of several fundus pathologies, including age-related macular degeneration,¹⁴ diabetic retinopathy,¹⁵ glaucoma,¹⁶ and a functional model to classify amblyopia.¹⁸ In the current study, the CSF patterns in different AI categories can also help to evaluate the degree of high myopia at an early stage. Myopia was found to have CSF deficits at high spatial frequencies.³⁰ Our study found that higher AI category was associated with worse CSF. Notably, the US Food and Drug Administration has previously mentioned 0.45 logCS being a meaningful three-step change, whereas we identified the two-step change between AI categories 1 and 3 was 0.94 logCS, suggesting that visual function may be highly discriminable by the AI classification. 

Our classification also showed close relevance to the fundus features, including MMD severity, β-PPA area, and retinal thickness, particularly manifested in higher AI categories. Macular atrophy as defined as MMD grade 4, was most common in AI category 4 with a nearly 50% incident rate. Eyes in AI category 4 tended to show more severe MMD and worse BCVA, which also indicates an association between poor visual acuity and severe maculopathy.²⁷ Changes in the peripapillary retina, including β-PPA enlargement and RNFL thinning occur with the progression of high myopia.^31–33 We demonstrated that a large β-PPA and thinner p-RNFL might be key characteristics for eyes in AI categories 3. Furthermore, the nerve fiber bundle from the macula to optic disc, also called the papillomacular bundle, plays a vital role in visual function. We noted that the area related to the papillomacular bundle, including the nasal area of the macula and the temporal part of p-RNFL, were thinner in eyes of a higher AI category, suggesting that the invisible fundus changes were consistent with those of visual function and could be well-reflected by the novel classification. 

CSF data may provide early indications of visual impairment, even before significant changes in fundus photographs. The novel classification combining CSF and fundus features could distinguish subtle variations in visual function and different degrees of fundus abnormalities, which might become a useful and elaborate tool to reflect the severity of high myopia. By applying this classification, doctors may discover hidden visual deficits at an earlier stage, which is of importance to the prevention, timely treatment, and improved follow-up for highly myopic patients. The novelty of this classification is its structure–function integration through clustering models. For the generation of structure–function clusters, to use fundus images and deep learning models can be of more convenient than to input the MMD grades. Nevertheless, large training, internal, and external testing datasets are required, and importing images from different devices with varied field into the AI algorithms requires further verification in clinical practice. We are working on adding a deep learning algorithm to the classification system which can automatically measure the grade of MMD, β-PPA area, and retinal thickness from fundus and OCT images, hoping to refresh it in future literature. Although clustering performance was verified, larger dataset for model retest and improvement will be collected. Future longitudinal studies may be needed to further explore its role in the prediction of visual prognosis. 

In conclusion, we proposed a novel AI-based classification of highly myopic eyes by integrating features from CSF and fundus. The application of the method to a large sample of patients demonstrated clear associations between the classification results and both visual function and fundus morphology. The new classification may serve as a useful tool to comprehensively evaluate the highly myopic eyes. 

Acknowledgments

Supported by research grants from the National Natural Science Foundation of China (82122017, 82271069, 82371040, 81870642, and 82301188), Science and Technology Innovation Action Plan of Shanghai Science and Technology Commission (23Y11909800 and 21S31904900), Outstanding Youth Medical Talents of Shanghai “Rising Stars of Medical Talents” Youth Development Program. 

Author Contributions: Conception and design: XJZ. Analysis and interpretation: XJZ, JQM, YXS, WWH, ZLL, YL. Data collection: JQM, WWH, YXC, KKZ, LW, JQ, YD, YL, XJZ. Obtained funding: XJZ, YL. Overall responsibility: XJZ, YL. 

Disclosure: J. Meng, None; Y. Song, None; W. He, None; Z.-L. Lu, Adaptive Sensory (P), Technology, Inc. (F); Y. Cheng, None; L. Wei, None; K. Zhang, None; J. Qi, None; Y. Du, None; Y. Lu, None; X. Zhu, None