Deep-learning-based diagnosis of myopia in children using optical coherence tomography angiography

Abstract

As myopia develops and progresses at an accelerated pace during adolescence, a timely diagnosis is beneficial for preventing its further progression. Optical coherence tomography angiography (OCTA) can visualize distinct layers of retinal microvessels, offering valuable insights into structural changes associated with myopia. This capability facilitates the early detection and monitoring of myopia-related complications, such as choroidal neovascularization and myopic maculopathy. Previous research suggests that alterations in the superficial capillary plexus (SCP) vessel density and deep capillary plexus (DCP) of OCTA images occur in myopic eyes, but few studies have focused on myopia in younger children.¹ A recent study compared retinal microvasculature in the SCP of children and adolescents using OCTA imaging, which indicated that there were no obvious variations in microvessel density, perfusion density (PD), and the size of the foveal avascular zone within the SCP between groups with mild and moderate/high myopia.² Conversely, another study demonstrated a negative correlation between children's myopia diopter and the microvessel density of both the superficial and deep retinal capillary plexus in the macula, as well as retinal thickness.³ Nonetheless, these studies only scrutinized a limited number of OCTA image parameters in the macular region and involved a relatively small number of subjects. Further research is warranted to fully understand the potential of OCTA images in assessing myopia during adolescence.

Deep learning can extract high-dimensional features from images through its multilayer network architecture, leading to improved task performance. However, to our knowledge, there is limited research on the use of artificial intelligence for analyzing OCTA images related to myopia. Our study contributes to addressing this research gap by highlighting the potential of deep learning in OCTA image analysis for myopia assessment. In this study, we aimed to employ end-to-end deep learning models to classify children with mild versus severe myopia. This study aimed to evaluate the potential of deep and superficial blood vessels in the macula and optic disc as indicators of myopia severity in children, utilizing a classification task based on OCTA images.

Initially, we collected four images from both the superficial and deep retinal capillary plexus in the macula, as well as from the optic disc, of children aged 8–16. Exclusion criteria included poor quality OCTA images, patients with other ocular conditions, or those who had undergone eye surgery. Ultimately, a total of 129 children (242 eyes) were included in this study. The subjects were divided into two groups based on their degree of refractive error: emmetropia/mild myopia (177 eyes, with a mean spherical equivalent between −3.00 and ≤0.50 D) and moderate/high myopia (65 eyes, with a mean spherical equivalent of less than −3.00 D). The average age of participants in this study was 11.14 years, including 63 girls. For more detailed information on the data set, please see File S1: Materials and methods section.

Next, we utilized the resnet152 model to analyze the performance of different layer images within the macula and optic disc for the classification task. The workflow is illustrated in Figure 1A, and detailed methods are available in the File S1: Materials and methods section. Our results indicate that the four separate OCTA image models (SCP and DCP of macula and optic disc, respectively) exhibit good performance in distinguishing between the emmetropia/mild and moderate/high myopia groups, with the area under the curves (AUCs) ranging from 0.803 to 0.892 (Supporting Information S1: Table S2). The superior discriminatory effect of these models could be attributed to the changes occurring in the retinal microvessels within the optic disc and macula during the period of myopia development in teenagers. These changes may be a consequence of altered blood vessel morphology arising from the elongation of the eye axis in this period. It is worth highlighting that the models utilizing superficial optic disc microvessel images exhibited the most optimal classification effect. Previous studies have suggested that changes in the shape of the optic disc play a significant role in the development and progression of myopia. The elongation of the axial length of the eye could cause remarkable changes in the size of the optic disc, peripapillary scleral flange, and peripapillary choroidal tissue, which may impact the morphology of the fundus blood vessels.⁴ These changes may be particularly noticeable in the superficial blood vessels of the optic disc. To improve interpretability, we utilized the Grad-CAM technique to visualize the areas of the fundus that the model focuses on, represented through heat maps, as shown in Figure 1B. The maps illustrate that not only the blood vessels but also the structure of the optic disc was taken into consideration in the model, specifically in the SVC image of the optic disc. Previous research using OCT for myopia detection achieved an AUC of 0.813 and an accuracy of 71.4%, with inner retinal layers and steepened curvature as key indicators.⁵ These findings are consistent with our study's outcomes, further validating our model's efficacy in myopia assessment.

Third, we attempted to incorporate the information of both superficial and deep retinal microvessels to investigate whether there exists complementary information across different levels that can enhance the model's performance. However, despite obtaining AUC values exceeding 80 (Figure 1C), the fusion model did not exhibit any superior classification performance compared to the single image models. The training data set size may be insufficient for optimal model training results. Therefore, further improvements in model construction and parameter adjustment are needed to enhance the model's performance. Additionally, a longitudinal observational study in this field would be necessary.

In summary, our study demonstrated that deep learning is a promising tool for evaluating OCTA images in children with myopia. Additionally, there were observable differences between superficial and deep retinal microvessels in the macula and optic disc for varying degrees of myopia, with the superficial OCTA images of the optic disc displaying the most effective performance. Further research is necessary to refine and optimize model construction and parameter adjustments for myopia assessment utilizing OCTA images. Our study will contribute to a better understanding of the mechanism of myopia development in children and aid the development of more effective interventions to prevent and manage myopia.

Conceptualization: Wenchao Xiao. Data curation: Wenchao Xiao and Wenting Zhao. Formal analysis: Wenchao Xiao, Lu Yuxing, Hanpei Miao, Wenting Zhao, and David Schanzlin. Funding acquisition: Wenchao Xiao. Investigation, methodology; project administration, resources, software, supervision, validation, writing—original draft: Wenchao Xiao and Hanpei Miao. Writing—review and editing: Wenchao Xiao, Hanpei Miao, and David Schanzlin. Wenchao Xiao, Lu Yuxing, Hanpei Miao, Wenting Zhao, and David Schanzlin collected and analyzed data. Wenchao Xiao designed the study and wrote the paper with the help of Hanpei Miao and David Schanzlin. All authors discussed the results and commented on the manuscript. All authors have read and approved the final manuscript.

The authors declare no conflict of interest.

Ethical approval for this study was obtained from the institutional ethics committee of the Chinese Academy of Medical Sciences, Zhuhai People's Hospital and adhered to the principles outlined in the Declaration of Helsinki. Written informed consent was obtained from the parents or guardians of all participants. APPROVAL NUMBER: ZY [2019] Number (20).