Abstract
Purpose
To investigate spatially distinctive features in fundus photographs of highly myopic glaucoma (HMG) by integrating radiomics and deep learning.
Design
Cross-sectional study.
Participants
Semi-automated optic disc segmentation was performed on 2000 images sourced from the Retinal Fundus Glaucoma Challenge Edition and Pathologic Myopia Challenge public data sets. We trained models with 628 images, including 217 cases of high myopia (HM), 221 cases of primary open-angle glaucoma (POAG), and 190 cases of HMG. An external validation set of 106 images was collected including 31, 36, and 39 fundus photographs from patients with HM, HMG, and POAG, respectively.
Methods
Semi-automated optic disc segmentation was performed by U-Net combined with manual delineation. Additionally, 5 regions of interest (ROIs) covering the optic disc and its surrounding region were explored. Fundus photography–based radiomics feature selection and optimization were conducted using random forest and support vector machine algorithms, which underwent fivefold cross-validation. Model performance was evaluated for radiomics, clinical, and combined models. An external validation set was used to evaluate the models performance, and we also examined radiomics features varied across glaucoma stages.
Main Outcome Measures
A total of 414 radiomics features were extracted from 5 regions of interest. We evaluated the model performance using accuracy, recall, precision, F1 score, and area under the receiver operating characteristic curve (AUROC). The Youden index was used to determine the cut-off value, and sensitivity and specificity were calculated.
Results
The U-Net–based optic disc segmentation reached a Dice Similarity Coefficient of 0.95. The radiomics model detected HMG from HM and POAG achieved the accuracies of 0.97 and 0.85, respectively. Both results exceeded the performance of the clinical model, which achieved 0.90 and 0.71. The top radiomics features achieved AUROC of 0.984 and 0.855 for HMG versus HM and HMG versus POAG in the external validation, respectively, including intensity features in the extra-optic nerve and textural features in the optic nerve. These features showed strong diagnostic capability when stratified by Youden values and exhibited independence from glaucoma progression.
Conclusions
The study combined U-Net and radiomics to delineate the spatial distribution biomarkers of HMG, establishing a quantitative classification model correlated with anatomical features.
Financial Disclosure(s)
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
Keywords: Radiomics, Deep learning, Highly myopic glaucoma, Fundus photography
Over the past 3 decades, high myopia (HM) has experienced a widespread surge globally, particularly in East Asia.1,2 By 2050, approximately 938 million individuals (9.8% of the global population) will be afflicted with HM.1 The risk of primary open-angle glaucoma (POAG) is nearly sixfold higher than that in the non-HM population, with a substantial 27.2% probability of developing glaucomatous optic neuropathy.3,4 Nontypical optic nerve (ON) tilting and twisting resulting from HM, ON peripapillary atrophy arc, and nontypical glaucomatous visual field (VF) defects collectively contribute to the diagnostic complexity. Abnormal ON morphology in HM poses challenges for objective evaluation with current imaging techniques, potentially delaying glaucoma treatment and leading to irreversible blindness.5, 6, 7
Artificial intelligence has shown promising results in detecting various eye conditions with sensitivities and specificities of up to 90%.8, 9, 10 However, when it comes to highly myopic glaucoma (HMG), the error rate can be as high as 50%. This discrepancy is due to challenges in image recognition, leading to false-negative and false-positive readings caused by issues such as ON tilt and irregular torsion.11
Radiomics is a technology that extracts high-throughput quantitative features from medical images using feature-mining capabilities.12 Radiomics has shown promising efficacy in ophthalmology, with studies across 5 major diseases and 7 imaging modalities reporting areas under the curve mostly ≥0.70.13 The methods of radiomics are preferable when dealing with small sample sizes.14,15 The existing studies have identified radiomics features of fundus photographs that are associated with the severity of HM and choroidal thickness.16,17 But there is still a lack of radiomics studies specifically focusing on HMG. Fundus photography remains a simple, cost-effective imaging modality that enables longitudinal follow-up. The combination of fundus photography and radiomic analysis provides a powerful tool for detecting HMG.
This study aimed to quantitatively detect the interpretable imaging biomarkers of HMG from fundus photographs by integrating deep learning and radiomics. In addition, we developed a segmentation-guided analysis framework to assess the contribution of regional retinal features to HMG detection.
Methods
Semi-Automated Optic Disc Segmentation
Auto Optic Disc Segmentation by U-Net
The Retinal Fundus Glaucoma Challenge Edition and Pathologic Myopia Challenge data sets, comprising 2000 images annotated with optic disc regions, were combined for training (1600 images) and testing (400 images) in optic disc segmentation (Supplementary File 1, available at www.ophthalmologyscience.org). After registration and normalization preprocessing of the original images, segmentation was performed using the U-Net algorithm in Python 3 (https://www.python.org/getit/). The U-Net framework included an encoder and a decoder, as shown in Figure 1A and Supplementary File 2 (available at www.ophthalmologyscience.org).
Figure 1.
(A) Framework of the U-Net algorithm for optic disc segmentation; (B) Semi-automated optic disc segmentation. The blue outline shows the U-Net–generated segmentation, and the green area shows the manual optic disc segmentation; (C) Design of subregional analysis for fundus photographs. HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma; ROI = region of interest.
The segmentation performance was evaluated using the DICE Similarity Coefficient (DSC), a widely used metric for comparing segmentations of different regions of interest (ROIs). The DSC ranges from 0 to 1, where a value of 1 indicates perfect overlap between the predicted segmentation and the ground truth and a value of 0 indicates no overlap. The application of U-Net for optic disc segmentation achieved a DSC of 0.95, demonstrating high accuracy. This outcome underscores the approach’s effectiveness in achieving accurate optic disc segmentation.
Manual Optic Disc Segmentation
The automated optic disc segmentation model was applied to fundus photographs collected from our enrolled patients, and the segmentation results were evaluated by 2 experienced examiners with >5 years of expertise (M.L. and S.F.H.). If both examiners agreed that the deep learning–based automatic segmentation was accurate, the process would proceed to the next phase of the study. If either examiner considered the segmentation incorrect, it was deemed inaccurate. Both examiners performed manual segmentation and calculated the DSC.
The 3D Slicer software (version 5.2.2 r31382, http://www.slicer.org/) was used. A total of 35 images in HM, 30 images in HMG, and 15 images in POAG underwent manual segmentation (Fig 1B). Approximately 79.94% of the U-Net–generated optic disc segmentations were deemed clinically acceptable by experienced ophthalmologists without manual correction. The DSC for manual segmentation was 0.92, indicating good consistency. We randomly selected one examiner’s manual segmentation result for inclusion in the subsequent analysis.
Establishment of Classification Models
Image Collection
We enrolled patients between January 2018 and December 2021 from Zhongshan Ophthalmic Center (ZOC) and collected clinical data from deidentified patient records. We enrolled 628 images from the ZOC data set to build classification models. Supplementary File 3 (available at www.ophthalmologyscience.org) shows the inclusion criteria and the exclusion criteria. This study was approved by the Ethics Committee of ZOC (No. 2022KYPJ105-2) and was performed according to the Declaration of Helsinki. The review board waived the consent requirement based on the retrospective study design and the deidentification of the images. The study (ChiCTR2300067468) was registered at http://www.clinicaltrials.gov.
Ophthalmic Examination and Color Fundus Photographs
All examinations were independently conducted in a masked manner by 2 experienced examiners (Yh. Zhuo and Yt. Zhu), without access to additional clinical data. Every participant underwent a comprehensive ophthalmic examination, including slit-lamp microscopy, gonioscopy, fundus examination, Goldmann applanation tonometry, axial length measurement (IOL Master, Carl Zeiss Meditec), VF mean deviation by a VF test (standard 30-2 algorithm, Humphrey Field Analyzer, Model 750, Carl-Zeiss Meditec), and color fundus photography using KOWA (Kowa NonmydWX, Kowa Inc).
ROI Definition and Segmentation
Building upon the precise segmentation of the optic disc, 5 ROIs were defined (Fig 1C): OpticOri (Optic Origin): Representing the area of the optic disc; OpticCir (Optic Circle): Denoting the minimum circumscribed circle area around the optic disc; C2: Defined by using the center of the optic disc as the center point and a radius of 1 papillary diameter to draw a circle, excluding the area of OpticCir; C3: Defined by using the center of the optic disc as the center point and a radius of 1.5 papillary diameter to draw a circle, excluding the area of C2; and outer: Encompasses the entire fundus photographs.
Fundus Photography-Based Radiomics Features Extraction
A total of 414 radiomics features were extracted from fundus photography images, comprising first-order statistical features and texture features derived from the original images, wavelet-transformed images, and Laplacian of Gaussian filtered images (Supplementary File 4, available at www.ophthalmologyscience.org). No OCT or OCT angiography images were used at any stage of the analysis, and all radiomics features were therefore based exclusively on fundus photography. Feature extraction was a pivotal step in radiomics analysis. The specific features and their algorithms are elaborated on the website (https://pyradiomics.readthedocs.io/en/latest/).
Manual Interpretation of Clinical Features
Experienced physicians interpreted clinical features from fundus photographs. Image quality assessment was conducted in accordance with the technical standards outlined in the Chinese Fundus Color Photograph Annotation and Quality Control Specification (T/CAQI 166-2020). The definition and diagnostic criteria are shown in Supplementary File 5 (available at www.ophthalmologyscience.org).18,19
Two attending doctors (Yt. Zhu and Z.L., with 10–15 years of experience) would interpret it. If they reached a consensus, the result was adopted; in cases of disagreement among the attending doctors, the final interpretation was made by one senior doctor (Cg. Zuo) with >15 years of experience.
Classification Models Establishment Processing
The radiomics model was created only using fundus photography–based radiomics features, while the clinical model relied on the clinical features provided by the doctor in Supplementary File 5. Additionally, we developed a combined model integrating radiomics and clinical parameters, enabling a more comprehensive comparison of model performance between radiomics and clinical features.
The data set was partitioned into 2 groups: (1) a randomly selected subset comprising 75% designated as the training set for model establishment and (2) the remaining 25% of the images serving as an independent test set.
Within the training set, we employed a fivefold cross-validation approach, with 80% of subjects for training and the remaining 20% for validation to tune model parameters. In the initial phase, we employed a random forest approach to select the most relevant features from radiomics, clinical, and the combined models within the training set. Features with an importance factor >0.1 were then chosen to construct the support vector machine model. Random forest and support vector machine were employed to select variables, simplify the model, reduce training time, and enhance generalization by mitigating overfitting. Detailed hyperparameter settings have been provided in the Table S1 (available at www.ophthalmologyscience.org). We then assessed the model’s performance on the test set.
External Validation
We obtained 31, 36, and 39 fundus photographs from patients with HM, HMG, and POAG, respectively. All images met the inclusion and exclusion criteria, and these patients were enrolled at ZOC between January 2021 and March 2022. These patients underwent the same subregional radiomics feature extraction. The fundus photography–based radiomics features were grouped based on Youden index, and the diagnostic performance of individual features was evaluated.
Image Quality Control for the Color Fundus Photographs
All images used in the study, including those from publicly available data sets and patient images collected from the hospital, underwent the following standard. According to the “Specification for Labeling and Quality Control of Fundus Color Photograph (T/CAQI 166-2020),” if ≥1 item was met, it was considered that the image quality was unqualified: (1) overexposed or underexposed images and (2) poor image clarity.
Data Analysis
The analysis was conducted using the R programming language (http://www.r-project.org/) and Statistical Product and Service Solutions version 25.0 (SPSS Inc). The Shapiro-Wilk test was employed to assess the distribution of variables. Non normally distributed data were represented as medians (interquartile range). The Mann–Whitney U test, Kruskal-Wallis test, and χ2 test were used to examine differences in non normally distributed variables. The correlation between continuous variables was assessed using Spearman correlation, and the correlation between dichotomous and continuous variables was assessed using point-biserial correlation. We evaluated the feature performance using accuracy, recall, precision, F1 score, and area under the receiver operating characteristic curve (AUROC). Statistical comparisons between models were performed using bootstrap-based confidence intervals (CIs) to account for sampling variability. The Youden index was used to determine the cut-off value, and sensitivity and specificity were calculated. Statistical significance was set at P < 0.05.
Results
A spatial classification model was established with 628 images. We enrolled 217 cases of HM, 221 cases of POAG, and 190 cases of HMG in the ZOC data set. The baseline data are presented in Table S2 (available at www.ophthalmologyscience.org). The overall study design is represented in Figure 2. In the HMG & HM binary classification model, 25, 23, 26, 28, and 32 features were selected, including 3 homogeneity features across the 5 subregions and 59 nonoverlapping spatial heterogeneity features. In the HMG & POAG binary classification model, 17, 17, 17, 11, and 22 features were selected, with no homogeneous features across the 5 subregions and 43 indicators of nonoverlapping heterogeneity. Radiomics features in different ROIs are presented in Figure 3.
Figure 2.
Overall study workflow. (A) Construction of the radiomics spatial model, including semi-automated optic disc segmentation, ROI definition, and radiomics feature extraction from fundus photographs. (B) Development of the clinical model based on ophthalmologists’ assessment. (C) Development of the combined model integrating clinical and radiomics information. (D) External validation of the radiomics spatial model for differentiating HMG from HM and POAG. HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma.
Figure 3.
Distribution of radiomics features with an importance factor > 0.1 across the OpticOri, OpticCir, C2, C3, and outer ROIs. (A) HMG versus POAG classification model. (B) HMG versus HM classification model. HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma; ROI = region of interest.
The Correlation Between Radiomics and Clinical Features
The correlation analysis between radiomics and clinical features showed that radiomics features were only weakly correlated with clinical features. The correlation coefficients were between 0.48 and 0.50 for the HMG versus HM model (Fig 4) and between 0.35 and 0.36 for the HMG versus POAG model (Fig 5), with all P values being <0.05.
Figure 4.
Heatmap of correlation analysis between HM and HMG. HM = high myopia; HMG = highly myopic glaucoma. ∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001.
Figure 5.
Heatmap of correlation analysis between POAG and HMG. HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma. ∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001.
The Performance of the Classification Model
HMG vs. HM
The fundus photography–based radiomics models for all ROIs demonstrated an exceptional ability to distinguish between HMG and HM. In the internal validation set, the accuracy, precision, recall, and F1 scores ranged from 0.98 to 0.99. In the test set, these metrics ranged from 0.93 to 0.98.
The clinical model achieved an accuracy and F1 score of 0.93 (95% CI: 0.90–0.95), with precision and recall of 0.91 (95% CI: 0.89–0.94) across all ROIs in the internal validation set. In the test set, the accuracy, precision, recall, and F1 score were 0.90 (95% CI: 0.87–0.92), 0.90 (95% CI: 0.87–0.92), 0.90 (95% CI: 0.87–0.92), and 0.91 (95% CI: 0.88–0.93), respectively.
The combined model demonstrated performance similar to that of the radiomics model and better than that of the clinical model. Detailed metrics for accuracy, precision, recall, and F1 scores ranged from 0.98 to 0.99 in the internal validation. In the test set, these metrics ranged from 0.95 to 0.98 (Table 1).
Table 1.
Classification Model Performance for HMG and HM
| Model | Precision |
Recall |
F1 Score |
Accuracy |
||||
|---|---|---|---|---|---|---|---|---|
| Test | Validation | Test | Validation | Test | Validation | Test | Validation | |
| Radiomics | ||||||||
| C2 | 0.97 (0.94–1.00) | 0.99 (0.99–1.00) | 0.98 (0.96–1.00) | 0.99 (0.99–1.00) | 0.97 (0.94–1.00) | 0.99 (0.99–1.00) | 0.97 (0.94–1.00) | 0.99 (0.98–1.00) |
| C3 | 0.96 (0.93–0.99) | 0.99 (0.97–1.00) | 0.98 (0.96–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.98–0.99) |
| OpticCir | 0.95 (0.92–0.99) | 0.98 (0.97–1.00) | 0.94 (0.92–0.98) | 0.98 (0.97–1.00) | 0.94 (0.92–0.98) | 0.98 (0.97–1.00) | 0.94 (0.92–0.98) | 0.99 (0.98–0.99) |
| OpticOri | 0.95 (0.92–0.99) | 0.99 (0.97–1.00) | 0.93 (0.89–0.97) | 0.99 (0.97–1.00) | 0.94 (0.92–0.98) | 0.99 (0.97–1.00) | 0.94 (0.92–0.98) | 0.98 (0.98–0.99) |
| Outer | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.98 (0.98–0.99) |
| Clinical | 0.90 (0.87–0.92) | 0.91 (0.89–0.94) | 0.90 (0.87–0.92) | 0.91 (0.89–0.94) | 0.91 (0.88–0.93) | 0.93 (0.90–0.95) | 0.90 (0.87–0.92) | 0.93 (0.90–0.95) |
| Combined | ||||||||
| C2 | 0.97 (0.94–1.00) | 0.99 (0.99–1.00) | 0.98 (0.98–1.00) | 0.99 (0.99–1.00) | 0.97 (0.94–1.00) | 0.99 (0.99–1.00) | 0.97 (0.94–1.00) | 0.99 (0.98–1.00) |
| C3 | 0.98 (0.96–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) |
| OpticCir | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.98 (0.96–1.00) | 0.98 (0.96–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) | 0.97 (0.94–1.00) | 0.99 (0.97–1.00) |
| OpticOri | 0.95 (0.91–0.99) | 0.98 (0.96–1.00) | 0.97 (0.94–1.00) | 0.98 (0.96–1.00) | 0.96 (0.93–0.99) | 0.98 (0.97–1.00) | 0.96 (0.93–0.99) | 0.98 (0.96–1.00) |
| Outer | 0.97 (0.94–1.00) | 0.98 (0.96–1.00) | 0.98 (0.96–1.00) | 0.98 (0.97–1.00) | 0.97 (0.94–1.00) | 0.98 (0.97–1.00) | 0.97 (0.94–1.00) | 0.98 (0.97–0.99) |
HM = high myopia; HMG = highly myopic glaucoma.
HMG vs. POAG
For the radiomics model, performance varied across different ROIs. The outer zone exhibited superior metrics, achieving the highest accuracy, precision, recall, and F1 scores compared to the other 4 ROIs, with values of 0.87 (95% CI: 0.85–0.90), 0.80 (95% CI: 0.77–0.88), 0.88 (95% CI: 0.85–0.90), and 0.87 (95% CI: 0.85–0.90) in the internal validation set. In the test set, the corresponding scores were 0.85 (95% CI: 0.80–0.90), 0.90 (95% CI: 0.81–0.92), 0.83 (95% CI: 0.80–0.88), and 0.84 (95% CI: 0.80–0.88).
The clinical model’s performance was not satisfactory. The accuracy, precision, recall, and F1 score were 0.73 (95% CI: 0.68–0.79), 0.73 (95% CI: 0.68–0.79), 0.71 (95% CI: 0.65–0.74), and 0.71 (95% CI: 0.65–0.74), respectively, in the intervalidation set. In the test set, the corresponding scores were 0.71 (95% CI: 0.68–0.79), 0.71 (95% CI: 0.68–0.76), 0.70 (95% CI: 0.68–0.77), and 0.70 (95% CI: 0.68–0.77).
The classification capability of the combined model varied across ROIs, with the outer zone exhibiting the best performance. However, the precision in the internal validation set was 0.87 (95% CI: 0.82–0.93) and 0.86 (95% CI: 0.82–0.92) in the test set. The metrics of the radiomics model and combined model were comparable to those of the clinical model (Table 2).
Table 2.
Classification Model Performance for HMG and POAG
| Model | Precision |
Recall |
F1 Score |
Accuracy |
||||
|---|---|---|---|---|---|---|---|---|
| Test | Validation | Test | Validation | Test | Validation | Test | Validation | |
| Radiomics model | ||||||||
| C2 | 0.79 (0.71–0.86) | 0.86 (0.80–0.91) | 0.74 (0.70–0.77) | 0.84 (0.79–0.88) | 0.75 (0.71–0.77) | 0.83 (0.78–0.89) | 0.77 (0.72–0.80) | 0.84 (0.79–0.89) |
| C3 | 0.78 (0.75–0.81) | 0.73 (0.66–0.81) | 0.79 (0.71–0.86) | 0.73 (0.66–0.81) | 0.79 (0.72–0.86) | 0.73 (0.66–0.81) | 0.79 (0.72–0.86) | 0.73 (0.65–0.80) |
| OpticCir | 0.88 (0.82–0.92) | 0.78 (0.75–0.81) | 0.73 (0.68–0.81) | 0.76 (0.75–0.78) | 0.73 (0.68–0.81) | 0.75 (0.73–0.77) | 0.77 (0.72–0.80) | 0.76 (0.75–0.78) |
| OpticOri | 0.79 (0.75–0.86) | 0.81 (0.76–0.87) | 0.71 (0.68–0.79) | 0.80 (0.75–0.84) | 0.72 (0.69–0.79) | 0.79 (0.74–0.84) | 0.76 (0.75–0.78) | 0.80 (0.75–0.84) |
| Outer | 0.90 (0.81–0.92) | 0.80 (0.77–0.88) | 0.83 (0.80–0.88) | 0.88 (0.85–0.90) | 0.84 (0.80–0.88) | 0.87 (0.85–0.90) | 0.85 (0.80–0.90) | 0.87 (0.85–0.90) |
| Clinical model | 0.71 (0.68–0.76) | 0.73 (0.68–0.79) | 0.70 (0.68–0.77) | 0.71 (0.65–0.74) | 0.70 (0.68–0.77) | 0.71 (0.65–0.74) | 0.71 (0.68–0.79) | 0.73 (0.68–0.79) |
| Combined model | ||||||||
| C2 | 0.76 (0.75–0.78) | 0.84 (0.80–0.90) | 0.73 (0.68–0.79) | 0.83 (0.77–0.88) | 0.73 (0.68–0.79) | 0.83 (0.77–0.88) | 0.77 (0.72–0.80) | 0.83 (0.77–0.88) |
| C3 | 0.78 (0.76–0.88) | 0.79 (0.71–0.86) | 0.77 (0.72–0.80) | 0.78 (0.71–0.85) | 0.77 (0.72–0.80) | 0.78 (0.71–0.85) | 0.79 (0.71–0.86) | 0.78 (0.72–0.85) |
| OpticCir | 0.76 (0.72–0.80) | 0.77 (0.70–0.82) | 0.75 (0.72–0.80) | 0.76 (0.70–0.81) | 0.76 (0.72–0.80) | 0.77 (0.69–0.80) | 0.75 (0.72–0.80) | 0.77 (0.70–0.80) |
| OpticOri | 0.75 (0.70–0.78) | 0.79 (0.74–0.85) | 0.75 (0.70–0.78) | 0.78 (0.73–0.83) | 0.74 (0.68–0.78) | 0.78 (0.72–0.83) | 0.76 (0.75–0.78) | 0.78 (0.73–0.84) |
| Outer | 0.86 (0.82–0.92) | 0.87 (0.82–0.93) | 0.83 (0.80–0.88) | 0.87 (0.82–0.92) | 0.84 (0.81–0.90) | 0.87 (0.81–0.92) | 0.85 (0.82–0.92) | 0.87 (0.82–0.92) |
HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma.
Spatial Mapping of Radiomics Features to Fundus Subregions
The cut-off values were established in the training set by maximizing the Youden index on the AUROC curve. We then calculated the sensitivity and specificity of radiomics features with an important factor >0.10 (Fig 6, Fig 7, Tables S3–S5, available at www.ophthalmologyscience.org). Single features with AUROC >0.950, sensitivity >0.90, and specificity >0.90 were selected for further analysis in HMG versus HM across all ROIs. Meanwhile, features with AUROC >0.850, sensitivity >0.80, and specificity >0.80 were selected for further analysis in the outer zone between HMG and POAG.
Figure 6.
The ROC of radiomics features in ROIs. (A–E) The ROC of OpticOri, OpticCir, C2, C3, and outer ROIs of HMG vs. HM; (F) The ROC of outer ROI of HMG vs. POAG. HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma; ROC = receiver operating characteristic curve; ROI = region of interest.
Figure 7.
Violin plot of variance analysis. Variance in radiomics features between HMG (yellow) andHM (blue) or POAG (green). HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma.
HMG vs. HM
We summarized 12 radiomics indicators with diagnostic areas under the curve >0.950, including 3 from the OpticOri region, 2 from OpticCir, one each from C2 and C3, and 5 from the outer region. By integrating overlapping and region-specific features corresponding to anatomical locations in fundus photographs, we obtained the summarized results presented in Table 3 and Figure 8.
Table 3.
The Performance of Radiomics Features in HMG vs. HM and HMG vs. POAG
| Model | ROI | Radiomics Features | AUROC (95% CI) | Youden | Sen (95% CI) | Spe (95% CI) | HMG | HM/POAG | P Value |
|---|---|---|---|---|---|---|---|---|---|
| HMG vs. HM | OpticOri | wavelet-H_glcm_Correlation | 0.960 (0.938–0.982) | –0.07 | 0.93 (0.91–0.96) | 0.96 (0.94–0.99) | –0.25 (0.12) | 0.02 (0.04) | 0.000 |
| wavelet-H_glrlm_RunPercentage | 0.953 (0.928–0.978) | 0.55 | 0.96 (0.94–0.99) | 0.90 (0.87–0.92) | 0.65 (0.08) | 0.49 (0.03) | 0.000 | ||
| wavelet-H_glrlm_LongRunEmphasis | 0.952 (0.927–0.977) | 4.62 | 0.90 (0.87–0.92) | 0.96 (0.94–0.99) | 3.03 (0.85) | 6.03 (1.35) | 0.000 | ||
| OpticCir | wavelet-H_glcm_Correlation | 0.958 (0.934–0.981) | –0.07 | 0.94 (0.92–0.96) | 0.95 (0.92–0.97) | –0.26 (0.11) | 0.01 (0.04) | 0.000 | |
| wavelet-H_glrlm_LongRunEmphasis | 0.955 (0.931–0.979) | 4.53 | 0.90 (0.87–0.92) | 0.96 (0.94–0.99) | 2.85 (0.85) | 5.75 (1.03) | 0.000 | ||
| C2 | wavelet-H_glcm_Correlation | 0.961 (0.939–0.983) | –0.09 | 0.92 (0.89–0.94) | 0.98 (0.96–1.00) | –0.32 (0.06) | 0.00 (0.05) | 0.000 | |
| C3 | wavelet-H_glcm_Correlation | 0.964 (0.942–0.986) | –0.06 | 0.92 (0.89–0.94) | 1.00 (0.97–1.00) | –0.33 (0.06) | 0.01 (0.04) | 0.000 | |
| Outer | wavelet-L_firstorder_RootMeanSquared | 0.976 (0.959–0.993) | 366.75 | 1.00 (0.97–1.00) | 0.93 (0.91–0.95) | 377.64 (2.77) | 355.03 (1.26) | 0.000 | |
| original_firstorder_RootMeanSquared | 0.975 (0.958–0.992) | 344.49 | 0.99 (0.97–1.00) | 0.93 (0.91–0.95) | 352.11 (3.39) | 335.35 (2.10) | 0.000 | ||
| wavelet-L_firstorder_Mean | 0.971 (0.953–0.989) | 47.85 | 0.98 (0.96–1.00) | 0.92 (0.89–0.94) | 61.89 (10.95) | 35.66 (7.19) | 0.000 | ||
| original_firstorder_Mean | 0.971 (0.953–0.989) | 33.84 | 0.98 (0.96–0.99) | 0.92 (0.89–0.94) | 43.77 (7.74) | 25.21 (5.08) | 0.000 | ||
| wavelet-H_glcm_Correlation | 0.963 (0.940–0.985) | –0.14 | 0.94 (0.92–0.96) | 0.97 (0.95–0.99) | –0.34 (0.05) | 0.00 (0.04) | 0.000 | ||
| HMG vs. POAG | Outer | log-sigma-5-0-mm-3D_firstorder_Mean | 0.868 (0.826–0.911) | –0.33 | 0.81 (0.79–0.85) | 0.84 (0.81–0.86) | –0.25 (0.13) | –0.44 (0.14) | 0.000 |
| log-sigma–1–0–mm-3D_firstorder_Mean | 0.856 (0.811–0.902) | –0.05 | 0.80 (0.77–0.82) | 0.83 (0.80–0.86) | –0.04 (0.03) | –0.07 (0.02) | 0.000 |
AUROC = area under the receiver operating characteristic curve; CI = confidence interval; HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma; ROI = region of interest; Sen = sensitivity; Spe = specificity.
Figure 8.
Diagram of the correspondence between radiomics features and the ocular structures in fundus images.
Wavelet-H_glcm_Correlation was significantly lower in HMG than in HM in all ROIs (all P = 0.000), with AUROC values ranging from 0.958 to 0.964. The determined cut-off value effectively classified HMG and HM, achieving sensitivities of 0.92 to 0.94 and specificities of 0.95 to 1.00.
Wavelet-H_glrlm_RunPercentage demonstrated an AUROC of 0.953 (95% CI: 0.928–0.978), where HMG exhibited higher values than HM (0.65 [0.08] vs. 0.49 [0.03]; P = 0.000) in the OpticOri zone. The cut-off point for wavelet-H_glrlm_RunPercentage was 0.55, with a sensitivity and specificity of 0.96 (95% CI: 0.94–0.99) and 0.90 (95% CI: 0.87–0.92), respectively.
Wavelet-H_glrlm_LongRunEmphasis was significantly lower in HMG compared to HM in OpticOri and OpticCir (both P = 0.000), with AUROC values of 0.952 (95% CI: 0.927–0.977) and 0.955 (95% CI: 0.931–0.979), respectively. Employing cut-off points of 4.62 and 4.53 for OpticOri and OpticCir, respectively, yield sensitivities and specificities of 0.90 (95% CI: 0.87–0.92) and 0.96 (95% CI: 0.94–0.99), respectively.
Wavelet-L_firstorder_RootMeanSquared, original_firstorder_RootMeanSquared, wavelet-L_firstorder_Mean, and original_firstorder_Mean in the outer zone exhibited AUROC values ranging from 0.971 to 0.976. These 4 features were significantly higher in HMG than in HM, with cut-off values achieving 0.98 to 1.00 sensitivity and 0.92 to 0.93 specificity for classifying HMG and HM.
HMG vs. POAG
Log-sigma-5-0-mm-3D_firstorder_Mean in the outer zone was significantly higher compared to HMG with POAG (–0.25 [0.13] and –0.44 [0.14]; P = 0.000). The AUROC was 0.868 (95% CI: 0.826–0.911), and the cut-off value was –0.33 with a sensitivity of 0.81 (95% CI: 0.79–0.85) and a specificity of 0.84 (95% CI: 0.81–0.86).
Log-sigma–1–0–mm-3D_firstorder_Mean in the outer zone was also significantly higher in HMG than in POAG (–0.04 [0.03] and –0.07 [0.02]; P = 0.000), and AUROC was 0.856 (95% CI: 0.811–0.902). The cut-off value was –0.05 with a sensitivity of 0.80 (95% CI: 0.77–0.82) and a specificity of 0.83 (95% CI: 0.80–0.86) (Table 3).
Radiomics Model External Validation
In the newly included HM, HMG, and POAG external validation set, the cut-off values based on Youden index in Table 3 demonstrated excellent diagnostic capability (Table 4). In the HMG versus HM classification model, the AUROC of individual radiomic features ranged from 0.919 to 0.984, with accuracy between 0.93 and 0.99. In the HMG versus POAG classification model, the AUROC of individual radiomics features was 0.855 (95% CI: 0.762–0.947) and 0.827 (95% CI: 0.727–0.926), with accuracy of 0.85 (95% CI: 0.77–0.91) and 0.82 (95% CI: 0.73–0.88), respectively.
Table 4.
Radiomics Features Performance Grouped by Youden Index Values
| Model | ROI | Radiomics Features | AUROC (95% CI) | Acc (95% CI) | Pre (95% CI) | Recall (95% CI) | F1 Score (95% CI) |
|---|---|---|---|---|---|---|---|
| HMG vs. HM | OpticOri | wavelet-H_glcm_Correlation | 0.956 (0.899–1.000) | 0.96 (0.90–0.98) | 0.97 (0.88–1.00) | 0.94 (0.82–0.98) | 0.96 (0.90–0.98) |
| wavelet-H_glrlm_RunPercentage | 0.952 (0.890–1.000) | 0.96 (0.90–0.98) | 0.92 (0.81–0.97) | 1.00 (0.92–1.00) | 0.96 (0.90–0.98) | ||
| wavelet-H_glrlm_LongRunEmphasis | 0.919 (0.841–0.998) | 0.93 (0.86–0.97) | 0.88 (0.77–0.94) | 1.00 (0.93–1.00) | 0.94 (0.82–0.98) | ||
| OpticCir | wavelet-H_glcm_Correlation | 0.956 (0.899–1.000) | 0.96 (0.90–0.98) | 0.97 (0.88–1.00) | 0.94 (0.82–0.98) | 0.96 (0.90–0.98) | |
| wavelet-H_glrlm_LongRunEmphasis | 0.924 (0.849–0.999) | 0.93 (0.86–0.97) | 0.92 (0.81–0.97) | 0.94 (0.84–0.98) | 0.93 (0.86–0.97) | ||
| C2 | wavelet-H_glcm_Correlation | 0.956 (0.899–1.000) | 0.96 (0.90–0.98) | 0.97 (0.88–1.00) | 0.94 (0.82–0.98) | 0.96 (0.90–0.98) | |
| C3 | wavelet-H_glcm_Correlation | 0.968 (0.917–1.000) | 0.97 (0.90–0.98) | 0.95 (0.83–0.99) | 0.95 (0.79–0.98) | 0.97 (0.88–1.00) | |
| Outer | wavelet-L_firstorder_RootMeanSquared | 0.984 (0.947–1.000) | 0.99 (0.95–1.00) | 0.97 (0.85–1.00) | 1.00 (0.89–1.00) | 0.99 (0.95–1.00) | |
| original_firstorder_RootMeanSquared | 0.984 (0.947–1.000) | 0.99 (0.95–1.00) | 0.97 (0.85–1.00) | 1.00 (0.89–1.00) | 0.99 (0.95–1.00) | ||
| wavelet-L_firstorder_Mean | 0.956 (0.899–1.000) | 0.96 (0.90–0.98) | 0.97 (0.88–1.00) | 0.94 (0.82–0.98) | 0.96 (0.90–0.98) | ||
| original_firstorder_Mean | 0.968 (0.917–1.000) | 0.97 (0.92–0.99) | 0.95 (0.86–0.98) | 1.00 (0.94–1.00) | 0.97 (0.88–1.00) | ||
| wavelet-H_glcm_Correlation | 0.942 (0.878–1.000) | 0.94 (0.88–0.97) | 0.97 (0.87–0.99) | 0.92 (0.83–0.97) | 0.94 (0.82–0.98) | ||
| HMG vs. POAG | Outer | log-sigma-5-0-mm-3D_firstorder_Mean | 0.855 (0.762–0.947) | 0.85 (0.77–0.91) | 0.82 (0.69–0.90) | 0.89 (0.78–0.96) | 0.85 (0.77–0.91) |
| log-sigma–1–0–mm-3D_firstorder_Mean | 0.827 (0.727–0.926) | 0.82 (0.73–0.88) | 0.81 (0.68–0.89) | 0.83 (0.71–0.92) | 0.82 (0.73–0.88) |
Acc = accuracy; AUROC = area under receiver operating characteristic curve; CI = confidence interval; HM = high myopia; HMG = highly myopic glaucoma; POAG = primary open-angle glaucoma; Pre = precision; ROI = region of interest.
Radiomics Features Across Glaucoma Stages
Spatial radiomics features exhibited significant differences between HM and HMG across early, moderate, and late stages of glaucoma (all P = 0.000). Similarly, significant differences were observed between POAG and HMG at corresponding stages from early to late (P values ranging from 0.000 to 0.005). Interestingly, no significant differences were found in HMG and POAG across the early, moderate, and late stages (P values ranged from 0.176 to 0.977). This stability indicated their potential as objective, early imaging markers for distinguishing HMG from other HM or POAG. Detailed results are summarized in Table 5 and Table S6 (available at www.ophthalmologyscience.org).
Table 5.
Radiomics Features for Glaucoma Staging Observation in HMG vs. HM
| ROI | Radiomics Features | HM | HMG |
∗P Value | aP Value | bP Value | cP Value | ||
|---|---|---|---|---|---|---|---|---|---|
| Early (n = 48) | Moderate (n = 45) | Late (n = 97) | |||||||
| OpticOri | wavelet-H_glcm_Correlation | 0.02 (0.04) | –0.23 (0.11) | –0.25 (0.14) | –0.22 (0.13) | 0.321 | 0.000 | 0.000 | 0.000 |
| wavelet-H_glrlm_RunPercentage | 0.49 (0.03) | 0.64 (0.13) | 0.66 (0.11) | 0.64 (0.07) | 0.528 | 0.000 | 0.000 | 0.000 | |
| wavelet-H_glrlm_LongRunEmphasis | 6.03 (1.35) | 3.19 (1.74) | 2.93 (1.04) | 3.12 (1.02) | 0.524 | 0.000 | 0.000 | 0.000 | |
| OpticCir | wavelet-H_glcm_Correlation | 0.01 (0.04) | –0.25 (0.09) | –0.26 (0.16) | –0.25 (0.13) | 0.515 | 0.000 | 0.000 | 0.000 |
| wavelet-H_glrlm_LongRunEmphasis | 5.75 (1.03) | 2.94 (1.42) | 2.76 (0.93) | 2.91 (1.04) | 0.561 | 0.000 | 0.000 | 0.000 | |
| C2 | wavelet-H_glcm_Correlation | 0.00 (0.05) | –0.30 (0.10) | –0.33 (0.07) | –0.31 (0.06) | 0.809 | 0.000 | 0.000 | 0.000 |
| C3 | wavelet-H_glcm_Correlation | 0.01 (0.04) | –0.32 (0.13) | –0.33 (0.06) | –0.33 (0.07) | 0.806 | 0.000 | 0.000 | 0.000 |
| Outer | wavelet-L_firstorder_RootMeanSquared | 355.03 (1.26) | 378.01 (4.31) | 377.80 (3.84) | 377.84 (2.37) | 0.782 | 0.000 | 0.000 | 0.000 |
| original_firstorder_RootMeanSquared | 335.35 (2.10) | 352.53 (4.66) | 352.62 (4.44) | 352.36 (2.96) | 0.956 | 0.000 | 0.000 | 0.000 | |
| wavelet-L_firstorder_Mean | 35.66 (7.19) | 62.90 (14.92) | 63.15 (13.41) | 63.88 (9.65) | 0.977 | 0.000 | 0.000 | 0.000 | |
| original_firstorder_Mean | 25.21 (5.08) | 44.48 (10.55) | 44.66 (9.48) | 45.17 (6.83) | 0.977 | 0.000 | 0.000 | 0.000 | |
| wavelet-H_glcm_Correlation | 0.00 (0.04) | –0.35 (0.12) | –0.34 (0.06) | –0.34 (0.06) | 0.896 | 0.000 | 0.000 | 0.000 | |
HM = high myopia; HMG = highly myopic glaucoma; ROI = region of interest.
Values shown as median (interquartile range).
P values were used to compare HM and HMG across the early, moderate, and late stages of glaucoma.
P value was intergroup class in HMG.
Discussion
We first performed semi-automated optic disc segmentation and built quantitative radiomics models to distinguish HMG from HM and POAG. The fundus photography–based radiomics classification models incorporating spatial information achieved accuracies of 97.0% and 85.0%, outperforming clinical models (90.0% and 71.0%) and matching the performance of combined models (up to 97.0% and 85.0%). The top-performing radiomics features achieved AUROC values of 0.984 and 0.855 for the HMG versus HM and HMG versus POAG models, respectively. These spatial radiomics features showed strong diagnostic capability when stratified by cut-off values and demonstrated potential as biomarkers for early differentiation of HMG.
Research on segmentation models tailored for optic disc delineation in HMG remains limited. Axial elongation in HM leads to a 10.8% glaucoma-like VF damage and the morphological abnormalities, which complicate the identification of the optic disc margin.6,20,21 As a result, uncertainties in HM may lead to overtreatment or undertreatment, ultimately impacting disease progression. In our study, we enrolled 190 HMG fundus images, and to ensure accuracy, we used a semi-automated method combining the U-Net algorithm and manual confirmation. The workflow maintains efficiency and high accuracy. Precise delineation of the optic disc is essential for all subsequent radiomics feature extraction and modeling. With reliable segmentation as the basis, the stability and credibility of the downstream analyses can be markedly improved.
Radiomics can extract quantitative and objective features through computer vision and reveal disease patterns beyond human visual assessment.22 Fundus photography is a noninvasive and cost-effective imaging technique that reliably monitors disease progression during long-term follow-up.23 To date, this is the first work to use radiomics on fundus photographs to define quantitative imaging markers of HMG. The fundus photography–based radiomics model showed better diagnostic performance than the subjective assessments of clinicians. Our findings, together with previous work, indicate that radiomics features from fundus images have weak correlations with clinical parameters.24 Unlike traditional clinical features, which rely heavily on the ophthalmologist’s experience, radiomics features can be extracted automatically. Machine learning–based features could capture color and texture information in addition to morphology. These newly derived image features offer a more comprehensive view of the disease’s characteristics.
The shallow depth of the cup and the reduced spatial contrast between the rim and the cup pose challenges for HMG detection, whereas indicators in the macular region and other retinal areas are clinically significant.22 Our subregional radiomics approach captured quantitative features of the optic disc and surrounding retina, enabling a broader assessment of HMG. Currently, no single structural test stands out as the best tool for detecting HMG.25,26 We analyzed 9 high-performance indicators; 7 came from extra-ON regions and 2 from the optic disc and peripapillary areas. In particular, the HMG versus HM classification model achieved an AUROC of 0.952 to 0.976, with 0.919 to 0.984 in external validation using cut-off–based grouping. This regional segmentation enables clinicians to evaluate the entire fundus while focusing on key subregions, thereby improving monitoring precision.
Our spatial radiomics model improved the biological relevance of radiomics features, facilitating the understanding of HMG pathogenesis. Compared with HM, HMG images showed more uniform textures in the optic disc and peripapillary areas, with bright, elongated structures appearing more continuous. Outside the disc, HMG displayed higher grayscale levels and more complex textures, especially in the macula. Glaucomatous eyes have been reported to show changes such as an increased cup-to-disc ratio, nasal shifting of central vessels, and expanded peripapillary beta zones.6 These changes support the imaging features that correspond to the uniform bright structural alterations observed in HMG optic disc images. In the macular region, previous studies have found significant reductions in macular retinal nerve fiber layer, ganglion cell complex, and ganglion cell-inner plexiform layer thickness in glaucomatous eyes (adjusted AUC >0.800).6,27,28 The retinal thinning and impairment reveal underlying choroidal vasculature and other structures, leading to greater grayscale variation in fundus photographs. Radiomics analysis can sensitively capture these subtle alterations, beyond the resolution of the human eye. This method provides deeper insight into HMG pathophysiology and enhances diagnostic accuracy.
Efforts to develop new structural parameters for the early detection of glaucomatous and/or myopia-related changes are underway. Most of the published work relies on OCT angiography or VF examinations.29, 30, 31 Jiang et al demonstrated that rapid temporal peri-retinal nerve fiber layer thinning may represent a valuable marker for detecting glaucoma in HM eyes.28 Akiyama et al reported that in early glaucomatous patients with HM, circumpapillary vessel density showed minimal influence from AL and maintained strong structure–function correlations.31 However, limited researches explore the use of fundus photographs for identifying early glaucoma with HM. In our study, the quantitative radiomics features reflected spatial information could distinguish early glaucoma from HM with an AUC of >0.952 and 0.919 in the internal data set and external validation, respectively. An interesting finding was that these radiomics features did not vary much across different glaucoma stages. This consistency suggests that the radiomics features could serve as stable imaging markers and be helpful in evaluating early glaucomatous changes in HM eyes.
This study has several limitations: (1) the training data set was derived from a single institution and all images were obtained from medical centers in China, which may introduce potential selection bias and limit the generalizability of the model across different populations and health care environments. Variations in imaging protocols, patient characteristics, and referral patterns across institutions may also influence model performance. Future studies incorporating multicenter data sets from broader regions and diverse populations will be essential to further validate the robustness and clinical applicability of the proposed radiomics model; and (2) the model showed slightly lower performance when separating HMG from POAG, although the AUC remained >0.827. These challenges may partly stem from the overlapping structural features between HMG and POAG, as both conditions can present with glaucomatous optic disc changes such as neuroretinal rim thinning and cup-to-disc ratio enlargement. In highly myopic eyes, additional fundus alterations including peripapillary atrophy and tessellated fundus may further obscure disease-specific imaging patterns. Refractive status and axial length can help clinicians distinguish these 2 conditions and partly compensate for the limits of fundus photography alone. Notably, retinal changes in HM can obscure early glaucomatous signs, underscoring the need for reliable artificial intelligence–assisted tools in this area.
Conclusions
We developed a semi-automated deep learning method for optic disc segmentation, and we further built fundus photography–based radiomics models that quantify image features from fundus photographs to distinguish HMG from HM and POAG. This approach demonstrated satisfactory diagnostic performance and considerable value for biomarkers for early differentiation of HMG. Location-specific imaging features provided valuable clues for identifying HMG and deeper insight into HMG pathogenesis.
Manuscript no. XOPS-D-25-01065.
Footnotes
Supplemental material available atwww.ophthalmologyscience.org.
Disclosure(s):
All authors have completed and submitted the ICMJE disclosures form.
The authors made the following disclosures:
All authors declare no potential conflicts of interest.
This study was supported by the National Natural Science Foundation of China (82471074, 62202318), Guangdong Basic and Applied Basic Research Foundation (2024A1515013058), the Guangdong Basic Research Center of Excellence for Major Blinding Eye Diseases Prevention and Treatment (2024-PIZC-022), and the R&D Program of Beijing Municipal Education Commission (KM202310025015).
Data Availability: The Zhongshan Ophthalmic Center data are not publicly available due to institutional restrictions on data sharing. The REFUGE2 data are available from https://refuge.grand-challenge.org/Home2020/. The PALM data are available from https://palm.grand-challenge.org/. Additional information can be obtained upon reasonable request from the corresponding author.
HUMAN SUBJECTS: No human subjects were included in this study. This study was approved by the Ethics Committee of ZOC (No. 2022KYPJ105-2) and was performed according to the Declaration of Helsinki. The review board waived the consent requirement based on the retrospective study design and the deidentification of the images. The study (ChiCTR2300067468) was registered at http://www.clinicaltrials.gov.
No animal subjects were included in this study.
Author Contributions:
Conception and design: Han, Zhou, Chen, Zuo, Xiao, Li, Huang, Chen, Zhao, Lu.
Analysis and interpretation: Zhu, Chen, Zuo, Xiao, Li, Huang, Chen, Zhao, Lu, Zhuo, Wang.
Data collection: Luo, Han, Hu, Zhou, Zhu, Chen, Zhuo, Wang.
Obtained funding: Luo, Han, Hu, Zhou, Chen, Zuo, Xiao, Li, Huang, Chen, Zhao, Lu.
Overall responsibility: Zhou, Zhu, Zuo, Xiao, Huang, Li, Chen, Zhao, Lu, Zhuo, Wang.
Contributor Information
Yehong Zhuo, Email: zhuoyh@mail.sysu.edu.cn.
Yizhou Wang, Email: Yizhou.wang@pku.edu.cn.
Supplementary Data
References
- 1.Holden B.A., Fricke T.R., Wilson D.A., et al. Global prevalence of myopia and high myopia and temporal trends from 2000 through 2050. Ophthalmology. 2016;123:1036–1042. doi: 10.1016/j.ophtha.2016.01.006. [DOI] [PubMed] [Google Scholar]
- 2.Jonas J.B., Ang M., Cho P., et al. IMI prevention of myopia and its progression. Invest Ophthalmol Vis Sci. 2021;62:6. doi: 10.1167/iovs.62.5.6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Marcus M.W., de Vries M.M., Junoy Montolio F.G., Jansonius N.M. Myopia as a risk factor for open-angle glaucoma: a systematic review and meta-analysis. Ophthalmology. 2011;118:1989–1994.e2. doi: 10.1016/j.ophtha.2011.03.012. [DOI] [PubMed] [Google Scholar]
- 4.Xu L., Wang Y., Wang S., et al. High myopia and glaucoma susceptibility the Beijing Eye Study. Ophthalmology. 2007;114:216–220. doi: 10.1016/j.ophtha.2006.06.050. [DOI] [PubMed] [Google Scholar]
- 5.Hsu C.H., Chen R.I., Lin S.C. Myopia and glaucoma: sorting out the difference. Curr Opin Ophthalmol. 2015;26:90–95. doi: 10.1097/ICU.0000000000000124. [DOI] [PubMed] [Google Scholar]
- 6.Jonas J.B., Wang Y.X., Dong L., Panda-Jonas S. High myopia and glaucoma-like optic neuropathy. Asia Pac J Ophthalmol (Phila) 2020;9:234–238. doi: 10.1097/APO.0000000000000288. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Sun M.T., Tran M., Singh K., et al. Glaucoma and myopia: diagnostic challenges. Biomolecules. 2023;13:562. doi: 10.3390/biom13030562. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Coan L.J., Williams B.M., Krishna Adithya V., et al. Automatic detection of glaucoma via fundus imaging and artificial intelligence: a review. Surv Ophthalmol. 2023;68:17–41. doi: 10.1016/j.survophthal.2022.08.005. [DOI] [PubMed] [Google Scholar]
- 9.Gulshan V., Peng L., Coram M., et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA. 2016;316:2402–2410. doi: 10.1001/jama.2016.17216. [DOI] [PubMed] [Google Scholar]
- 10.Li Z., He Y., Keel S., et al. Efficacy of a deep learning system for detecting glaucomatous optic neuropathy based on color fundus photographs. Ophthalmology. 2018;125:1199–1206. doi: 10.1016/j.ophtha.2018.01.023. [DOI] [PubMed] [Google Scholar]
- 11.Liu H., Li L., Wormstone I.M., et al. Development and validation of a deep learning system to detect glaucomatous optic neuropathy using fundus photographs. JAMA Ophthalmol. 2019;137:1353–1360. doi: 10.1001/jamaophthalmol.2019.3501. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Meng Z., Chen Y., Li H., et al. Machine learning and optical coherence tomography-derived radiomics analysis to predict persistent diabetic macular edema in patients undergoing anti-VEGF intravitreal therapy. J Transl Med. 2024;22:358. doi: 10.1186/s12967-024-05141-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Zhang H., Zhang H., Jiang M., et al. Radiomics in ophthalmology: a systematic review. Eur Radiol. 2025;35:542–557. doi: 10.1007/s00330-024-10911-4. [DOI] [PubMed] [Google Scholar]
- 14.Afshar P., Mohammadi A., Plataniotis K.N., et al. From handcrafted to deep-learning-based cancer radiomics: challenges and opportunities. IEEE Signal Process Mag. 2019;36:132–160. doi: 10.1109/MSP.2018.2877123. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Tan H.Q., Cai J., Tay S.H., et al. Cluster-based radiomics reveal spatial heterogeneity of bevacizumab response for treatment of radiotherapy-induced cerebral necrosis. Comput Struct Biotechnol J. 2024;23:43–51. doi: 10.1016/j.csbj.2023.11.040. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Du R., Ohno-Matsui K. Novel uses and challenges of artificial intelligence in diagnosing and managing eyes with high myopia and pathologic myopia. Diagnostics (Basel) 2022;12:1210. doi: 10.3390/diagnostics12051210. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Sun D., Du Y., Chen Q., et al. Imaging features by machine learning for quantification of optic disc changes and impact on choroidal thickness in young myopic patients. Front Med (Lausanne) 2021;8 doi: 10.3389/fmed.2021.657566. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Dong L., He W., Zhang R., et al. Artificial intelligence for screening of multiple retinal and optic nerve diseases. JAMA Netw Open. 2022;5 doi: 10.1001/jamanetworkopen.2022.9960. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Lin D., Xiong J., Liu C., et al. Application of Comprehensive Artificial intelligence Retinal Expert (CARE) system: a national real-world evidence study. Lancet Digit Health. 2021;3:e486–e495. doi: 10.1016/S2589-7500(21)00086-8. [DOI] [PubMed] [Google Scholar]
- 20.Lin F., Chen S., Song Y., et al. Classification of visual field abnormalities in highly myopic eyes without pathologic change. Ophthalmology. 2022;129:803–812. doi: 10.1016/j.ophtha.2022.03.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Wang Y.X., Panda-Jonas S., Jonas J.B. Optic nerve head anatomy in myopia and glaucoma, including parapapillary zones alpha, beta, gamma and delta: histology and clinical features. Prog Retin Eye Res. 2021;83 doi: 10.1016/j.preteyeres.2020.100933. [DOI] [PubMed] [Google Scholar]
- 22.Jiravarnsirikul A., Belghith A., Rezapour J., et al. Evaluating glaucoma in myopic eyes: challenges and opportunities. Surv Ophthalmol. 2025;70:563–582. doi: 10.1016/j.survophthal.2024.12.003. [DOI] [PubMed] [Google Scholar]
- 23.Tuulonen A., Kim T.W., Chauhan B.C., et al. The usefulness of routine photography in a myopic patient treated for glaucoma: 36-year follow-up. Acta Ophthalmol. 2024;102(4):e646–e650. doi: 10.1111/aos.16633. [DOI] [PubMed] [Google Scholar]
- 24.Du Y., Chen Q., Fan Y., et al. Automatic identification of myopic maculopathy related imaging features in optic disc region via machine learning methods. J Transl Med. 2021;19:167. doi: 10.1186/s12967-021-02818-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Hemelings R., Elen B., Barbosa-Breda J., et al. Deep learning on fundus images detects glaucoma beyond the optic disc. Sci Rep. 2021;11 doi: 10.1038/s41598-021-99605-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Jeong Y., Kim Y.K., Jeoung J.W., Park K.H. Comparison of optical coherence tomography structural parameters for diagnosis of glaucoma in high myopia. JAMA Ophthalmol. 2023;141:631–639. doi: 10.1001/jamaophthalmol.2023.1717. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Rezapour J., Walker E., Belghith A., et al. Diagnostic accuracy of optic nerve head and macula OCT parameters for detecting glaucoma in eyes with and without high axial myopia. Am J Ophthalmol. 2024;266:77–91. doi: 10.1016/j.ajo.2024.04.022. [DOI] [PubMed] [Google Scholar]
- 28.Jiang J., Kong K., Lin F., et al. Longitudinal changes of retinal nerve fiber layer and ganglion cell-inner plexiform layer in highly myopic glaucoma: a 3-Year cohort study. Ophthalmology. 2025;132:644–653. doi: 10.1016/j.ophtha.2025.01.014. [DOI] [PubMed] [Google Scholar]
- 29.Yang J., Li Y., Zhang Q., et al. Optic nerve morphology influences structure-function relationship in early glaucoma with and without high myopia. Invest Ophthalmol Vis Sci. 2025;66:18. doi: 10.1167/iovs.66.4.18. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Akiyama K., Saito H., Shirato S., et al. Structure function relationships of vessel density and retinal nerve fiber layer thickness in early glaucomatous eyes with high myopia. Invest Ophthalmol Vis Sci. 2023;64:14. doi: 10.1167/iovs.64.4.14. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Chang P.Y., Wang J.Y., Wang J.K. Optical coherence tomography angiography compared with optical coherence tomography for detection of glaucoma progression with high myopia. Sci Rep. 2025;15:9762. doi: 10.1038/s41598-025-91880-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.