Development of Deep Learning Models to Screen Posterior Staphylomas in Highly Myopic Eyes Using UWF-OCT Images

Abstract

Purpose: To develop a deep learning (DL) model for screening posterior staphylomas in highly myopic patients using ultra-widefield optical coherence tomography (UWF-OCT) images.

Methods: Our retrospective single-center study collected 1428 qualified UWF-OCT images from 438 highly myopic patients between 2017 and 2019 for model development. An independent test dataset for internal validation included 216 images from 69 highly myopic patients obtained between June 2020 and December 2020. Posterior staphylomas were detected by identifying the staphyloma edges. Seven independent architectures (VGG16, VGG19, ResNet18, ResNet50, ResNet101, DenseNet121, and DenseNet161) were used to train the models and identify staphyloma edges. The area under the receiver operating characteristic (ROC) curve (AUC) was used to evaluate and compare the performance of each model.

Results: The AUCs of seven DL models ranged from 0.794 (95% confidence interval [CI], 0.708–0.875) to 0.903 (95% CI, 0.846–0.953) for staphyloma edge detection in the internal test dataset. VGG19, with the highest AUC, achieved sensitivity (0.871; 95% CI, 0.773–0.931) that was comparable to or better than those of retina specialists. Heatmaps showed that the DL models could precisely identify the region of staphyloma edges.

Conclusions: Our models reliably identified staphyloma edges with high sensitivity and specificity. Given that posterior staphylomas are a key contributor to various fundus complications, the development of DL models holds significant promise for improving the clinical management of highly myopic patients.

Translational Relevance: This effective artificial intelligence system can help ophthalmologists screen posterior staphylomas in highly myopic eyes.

Introduction

Myopia is a growing global epidemic of great concern.¹ The recent rise in the prevalence of myopia worldwide will increase the incidence of pathologic myopia in the future. A newly revised definition of pathologic myopia includes posterior staphyloma, which is described as a posterior outpouching of the ocular wall with a shorter radius of curvature than the surrounding regions.² The growth of posterior staphylomas causes mechanical damage to optically critical structures such as the macular retina and the optic nerve, increasing the risk of myopic complications and ultimately blindness.³ Previous studies have indicated that highly myopic eyes with posterior staphylomas exhibit significantly poorer visual outcomes, including lower best-corrected visual acuity and more frequent visual field defects, compared to eyes without staphylomas.^4–6 

At present, ultra-widefield optical coherence tomography (UWF-OCT) is the best method for diagnosing posterior staphylomas.³ UWF-OCT is equipped with a scanning range of 23 × 20 mm and a depth of 5 mm, allowing for tomographic images of posterior staphylomas with unprecedented resolution and size.⁷ This capability overcomes the constraints of previous approaches based on conventional OCT, which lacked the width required for comprehensive staphyloma inspection. Compared with three-dimensional magnetic resonance imaging, UWF-OCT reduces examination time and enables imaging from the vitreous to the sclera in a single image, facilitating the diagnosis and analysis of posterior staphylomas.^7–11 

Deep learning (DL), a subfield of artificial intelligence (AI), has demonstrated remarkable ocular imaging analysis capabilities and is increasingly used in the field of myopia.^12,13 DL algorithms use multilayered artificial neural networks for feature extraction and transformation, without the need for manual feature engineering. This ability allows them to analyze more complicated inputs, such as entire images, and achieve superior performance, particularly in tasks such as image classification.¹⁴ OCT image evaluation is now one of the most commonly used AI approaches for high myopia because it is non-invasive and repeatable and has few side effects, allowing the collection of many images.¹⁵ DL models using OCT images have shown consistent efficacy in diagnosing pathologic myopia and related complications, such as myopic macular neovascularization,¹⁶ dome-shaped macula, retinoschisis, macular hole, and retinal detachment.^17,18 

Given the irreversible myopic changes and subsequent vision loss, prevention or treatment is highly needed before the occurrence of secondary changes of posterior staphylomas. Nevertheless, identifying posterior staphylomas on UWF-OCT remains difficult, especially for non-retina specialists. Because of insufficient and unevenly distributed medical resources, ophthalmologists have difficulties screening all posterior staphylomas and identifying high-risk individuals who would benefit from early intervention. 

This study aimed to construct an AI system for screening posterior staphylomas in highly myopic patients using UWF-OCT images. We also collected an independent test dataset to assess the practicality of the AI system, and the performance was compared with human ophthalmologists from the Advanced Clinical Center for Myopia, Institute of Science Tokyo (Science Tokyo, Tokyo, Japan). 

Methods

Ethics and Data Protection

This study adhered to the tenets of the Declaration of Helsinki and was authorized by the Ethics Committee of Science Tokyo (application no. M2022-180). All participants completed consent forms, and any private information that could identify them was excluded. 

UWF-OCT Image Acquisition and Labeling

Data from highly myopic patients and their UWF-OCT images were retrospectively collected from the Advanced Clinical Center for Myopia, Science Tokyo, between February 2017 and March 2019. High myopia was defined as a refractive error ≤ –6.0 diopters (D) and/or an axial length ≥ 26.5 mm. The UWF-OCT device was a Xephilio OCT-S1 instrument (Canon, Tokyo, Japan). Multiple examinations of the same eye were considered independent data if the interval between examinations was longer than 1 month. We excluded examinations with poor-quality UWF-OCT scans, as well as eyes with a history of vitreoretinal surgery and patients with systemic or ocular diseases related to staphylomas. The inclusion and exclusion criteria details are presented in Supplementary Figures S1 and S2. Both horizontal and vertical scans of one examination were included. Finally, 1428 qualified UWF-OCT images collected from 701 eyes of 438 patients were chosen for AI development. 

To prevent data leakage, we separated 1428 images into two datasets at the patient level with a ratio of 8:2. The training dataset (80% of the patients) was used for model development, and the validation dataset (20% of the patients) was used for hyperparameter optimization. Another independent test dataset was used to evaluate the AI system in a real clinical situation. Also, 216 qualified images from 108 examinations of 69 highly myopic patients recruited at Science Tokyo between June 2020 and December 2020 were chosen using the same criteria. No patients had previously been utilized in the training or validation datasets. 

The existence of posterior staphylomas in each UWF-OCT image was determined by detecting the staphyloma edges. The study by Shinohara et al.⁷ provided the diagnosis criteria of staphyloma edges—gradual “thick–thin–thick” changes in choroidal thickness from the periphery to the edge of the staphyloma and then to the posterior pole, as well as an inward protrusion of the sclera at the edge. In the previous study, staphyloma edges were determined in horizontal and vertical scans from a single examination. If at least one staphyloma edge was observed, the eye was diagnosed with posterior staphyloma.⁸ The diagnosis of staphyloma edges is illustrated in Supplementary Figure S3. Three human doctors were selected as reviewers, including one senior retina expert and two retina specialists. Two retina specialists (CC, YW) reviewed all images independently. If the label outcomes of the two reviewers coincided, then the outcome was considered the standard diagnosis (ground truth) of the image. The standard diagnosis was validated in a group discussion with the senior retina expert (KO-M) when there was disagreement. 

Development of AI System

Seven independent convolutional neural network (CNN) architectures were evaluated, including VGG16, VGG19, ResNet18, ResNet50, ResNet101, DenseNet121, and DenseNet161. Deeper CNNs (e.g., ResNet, DenseNet) have been shown to enhance performance by increasing the number of layers, using repeated modules with more complex or parallel filters, bottleneck connections, and dropout. Older architectures (e.g., VGGNet) were still selected in this study because they have consistently produced comparable results in medical image analyses.¹² Due to the limited number of images used for training, transfer learning was also used during the training process to improve the performance.¹⁹ In our study, the weights of each architecture pretrained on ImageNet were kept, and only the weights of fully connected layers were updated on the training dataset. Furthermore, before being employed in the model development, raw UWF-OCT images were preprocessed. First, all original images were scaled to the same size, from 2300 × 1480 pixels to 230 × 148 pixels. We downsized the input images without cropping to increase the prediction speed while at the same time keeping all information of the images. Second, image normalization was performed using the mean and variance values of the pretraining dataset. Third, to increase the heterogeneity, the images in the training dataset were randomly augmented by horizontal and vertical flipping. The number of newly generated images in each epoch was the same. This data augmentation method was universally used to increase the robustness, especially for the training process with a small sample size. 

The hyperparameters of each algorithm were fine tuned based on the model performance on the validation dataset. The models with the highest areas under the receiver operating characteristic (ROC) curve (AUCs) were selected as the final ones. The Youden index method²⁰ was used to determine the optimal cutpoint in the ROC curve. Additional information on the image preprocessing and AI training process is provided in Supplementary Methods S3 and S4. 

Evaluation of the AI System

Figure 1 represents the overall process of our AI system. When the AI system was established, each input image underwent preprocessing before being fed into the final models. In the test dataset, the classification result of each image was transformed into the diagnosis outcome of each eye. Each eye included two images, the horizontal and vertical scans. If at least one scan was determined with the staphyloma edge, the eye was diagnosed with staphyloma. This process was more like clinical practice. The performance of the AI system was assessed using sensitivity, specificity, and AUCs. The heatmaps of corresponding input images were generated by gradient-weighted class activation mapping (Grad-CAM),²¹ which computes gradients and uses them as weights to visualize the activation regions. Finally, the other two retina specialists (YW, ZW) reviewed the images of the test dataset, and the diagnosis results were compared with the AI system by the ROC curve. 

Figure 1.

 

The overall process of the AI system. The study had three main stages: (1) data preparation, (2) model development, and (3) model evaluation. (1) All of collected UWF-OCT images were divided into training, validation, and test datasets. The training and validation datasets were used for model training, and the independent test dataset was used to assess model performance. (2) After being scaled to the same size, all images were augmented by horizontal and vertical flipping. In our study, seven different CNN architectures were experimented with. (3) In addition to basic metrics for evaluation, the diagnosis results of CNN models were compared with human doctors and visualized.

Statistical Analysis

Cohen’s kappa values were calculated during data labeling to test interrater reliability between the two retina specialists.²² All seven CNN architectures were trained on the Python 3.10.12 and PyTorch 2.0.1 platforms. ROC curves were plotted with the Python matplotlib 3.8.0 package and analyzed with the Python module scikit-learn 1.5.1. The DeLong method was used to calculate the 95% confidence intervals (CIs) of AUCs.²³ The sensitivity and specificity were calculated according to the confusion matrix, which was generated by scikit-learn 1.5.1. The 95% CIs represented the Wilson score intervals for sensitivity and specificity. 

Results

Demographics and Baseline Characteristics of the Datasets

The dataset for AI system development was comprised of 1428 images obtained from 438 patients (701 eyes, including 340 right eyes and 361 left eyes), of which 559 images (39.15%) showed staphyloma edges. The images were randomly assigned to the training dataset (1138 images, 80%) and the validation dataset (290 images, 20%). The training dataset included 559 eyes (268 right eyes and 291 left eyes), and the validation dataset included 142 eyes (72 right eyes and 70 left eyes). Another independent test dataset included 216 images obtained from 108 eyes (including 49 right eyes and 59 left eyes) of 69 patients. The diagnosis of the test dataset was assigned at the eye level for each examination to reflect the diagnostic approach in real clinical practice, of which 70 examinations (64.81%) were staphylomas. More information on the three datasets is shown in Table 1. 

Table 1.

 

Development and Test Dataset Details

 

Performance of the AI System

In this study, we tried seven independent CNN architectures to detect staphyloma edges. The models were trained based on the training dataset. A significant number of trials were run on each model to optimize hyperparameters, and the models with the best AUCs on the validation dataset were chosen. Supplementary Table S1 shows the threshold and AUC of each CNN architecture in the validation dataset. In the independent test dataset, the AUCs of seven models ranged from 0.794 (95% CI, 0.708–0.875) to 0.903 (95% CI, 0.846–0.953) for staphyloma edge detection at eye level. The VGG19 architecture outperformed other architectures, including VGG16, ResNet18, Resnet50, ResNet101, DenseNet121, and DenseNet161. The VGG19 architecture achieved a sensitivity of 0.871 (95% CI, 0.773–0.931) and a specificity of 0.737 (95% CI, 0.579–0.85), showing good generalization on the test dataset. The details of model performance are listed in Table 2. Our AI system also showed equal or even better performance than the retina specialists (Fig. 2). 

Table 2.

 

Performance of Different CNN Models for the Test Dataset

 

Figure 2.

 

Comparison of the performance of the AI system and ophthalmologists for posterior staphylomas using the ROC curve.

Heatmaps

The generated heatmap demonstrated that the AI system made diagnoses based on the lesions of staphyloma edges (Fig. 3). Warmer colors in the heatmaps indicate areas critical for determining the classification. Similar to the judgment of the retina specialists, the AI system can precisely capture the region of staphyloma edges. In the meantime, our AI system output labels (staphyloma edges/no staphyloma edges) and triggered an informed sentence at the top of the UWF-OCT image when the staphyloma edge was detected (Supplementary Fig. S5). 

Figure 3.

 

Heatmaps for staphyloma edges. (A) An example of a staphyloma edge lesion detected by our AI system. The heatmap highlighted the region most relevant to the diagnosis in the UWF-OCT image. (B) The raw UWF-OCT image.

Discussion

In this study, we compared seven independent CNN models (VGG16, VGG19, ResNet18, ResNet50, ResNet101, DenseNet121, and DenseNet161) for the detection of staphyloma edges in highly myopic eyes using UWF-OCT images. VGG19, which achieved the highest AUC, demonstrated a correct answer rate equivalent to that of the human retina specialists. To the best of our knowledge, this is the first study to develop and verify an effective AI system for screening staphylomas in highly myopic eyes. 

Pathologic myopia is a global public health concern, and its complications rank among the top causes of blindness among the general population of Asia.^24–26 Posterior staphylomas, as hallmarks of pathologic myopia, are associated with other irreversible myopic changes and vision loss.³ Our AI system, which provides diagnoses along with corresponding heatmaps, offers a potential cost-efficient solution for screening and monitoring highly myopic patients in eyecare settings. The use of this AI system can assist ophthalmologists in identifying patients with the potential risk of myopia-related blindness and improve their quality of life. In future studies, this system will be used as the groundwork for further transfer training and evaluated in other cohorts, especially for areas with limited data. 

Previous studies have reported that the OCT image is a useful tool for the development of CNN algorithms to identify pathologic myopia and related lesions.^{16–18,27,28} To detect vision-threatening conditions in high myopia, including retinoschisis, macular hole, retinal detachment, and myopic choroidal neovascularization, Li et al.¹⁸ used 5505 qualified OCT macular images to develop and evaluate a DL program. Their AI system achieved high sensitivities (0.952–1) and AUCs (0.961–0.999) for all conditions. Ye et al.¹⁷ built an AI system to screen five myopic maculopathies using 2342 qualified OCT macular images. The AUCs ranged from 0.927 to 0.974. In our previous study,¹⁶ we succeeded in detecting myopic macular neovascularization, myopic traction maculopathy, and dome-shaped macula using 9176 OCT images, achieving AUCs of 0.985, 0.946, and 0.978, respectively. 

Although our AI system achieved good performance (AUC, 0.903; sensitivity, 0.871), implementing AI systems for posterior staphylomas remains challenging compared to other studies using OCT images. One possible reason is the relatively low number of available UWF-OCT images (n = 1428) for AI system development. UWF-OCT is a relatively new examination technique, first reported for diagnosing staphylomas in 2017.⁷ Several strategies were employed to increase the number of images in our study. First, both horizontal and vertical scans of each eye were used for training. Second, during preprocessing, random augmentation was applied, including horizontal and vertical flipping to increase the image heterogeneity. Further attempts with the collection of additional UWF-OCT data and using advanced AI techniques, such as few-shot learning²⁹ and generative adversarial networks³⁰ for small data, could be beneficial. 

The highlighted regions on the heatmaps showed that our AI system can precisely identify UWF-OCT features of staphyloma edges. As shown in the ROC curve (Fig. 2), the AI system exhibited equal or even better sensitivities than retina specialists. By carefully checking misdiagnosed cases by the AI system and human doctors in the test dataset, edges with slighter curvature or minor deformity and edges near the image boundary may be potential contributors to these errors. Typical examples and corresponding heatmaps are shown in Supplementary Figure S6. For challenging cases, a combined analysis based on 12 sections and three-dimensional reconstructed images of UWF-OCT images is recommended. 

Limitations

Our study has a few limitations. First, the data used for AI system development and validation were obtained from Science Tokyo, and the results were not validated externally. Future research should include multicenter validation to enhance the reliability of our AI system. Second, all patients used for training and testing data were Japanese, and transfer learning may be required when applied to Caucasians and individuals of other ethnicities. Third, the input UWF-OCT images must be high quality. Blurred images due to severe cataracts, vitreous hemorrhage, or severely poor fixation cannot be used for diagnosis. 

Conclusions

This is the first study, to our knowledge, that reports the development of an AI system based on CNN architectures to screen posterior staphylomas in highly myopic eyes using UWF-OCT images. Our AI system has demonstrated good performance, which can be valuable in reducing workload and preventing irreversible damage caused by posterior staphylomas. 

Acknowledgments

Supported by a research grant from the Institute of Science Tokyo (91AA191450). 

Disclosure: Y. Wang, None; C. Chen, None; Z. Wang, None; Y. Wu, None; H. Lu, None; J. Xiong, None; K. Sugisawa, None; K. Kamoi, None; K. Ohno-Matsui, None