AI-based fully automatic analysis of retinal vascular morphology in pediatric high myopia

Abstract

Purpose: To investigate the changes in retinal vascular structures associated with various stages of myopia by designing automated software based on an artificial intelligence model.

Methods: The study involved 1324 pediatric participants from the National Children's Medical Center in China, and 2366 high-quality retinal images and corresponding refractive parameters were obtained and analyzed. Spherical equivalent refraction (SER) degree was calculated. We proposed a data analysis model based on a combination of the Convolutional Neural Networks (CNN) model and the attention module to classify images, segment vascular structures, and measure vascular parameters, such as main angle (MA), branching angle (BA), bifurcation edge angle (BEA) and bifurcation edge coefficient (BEC). One-way ANOVA compared parameter measurements between the normal fundus, low myopia, moderate myopia, and high myopia groups.

Results: The mean age was 9.85 ± 2.60 years, with an average SER of -1.49 ± 3.16D in the right eye and - 1.48 ± 3.13D in the left eye. There were 279 (12.38%) images in the normal group and 384 (16.23%) images in the high myopia group. Compared with normal fundus, the MA of fundus vessels in different myopic refractive groups was significantly reduced (P = 0.006, P = 0.004, P = 0.019, respectively), and the performance of the venous system was particularly obvious (P < 0.001). At the same time, the BEC decreased disproportionately (P < 0.001). Further analysis of fundus vascular parameters at different degrees of myopia showed that there were also significant differences in BA and branching coefficient (BC). The arterial BA value of the fundus vessel in the high myopia group was lower than that of other groups (P = 0.032, 95% confidence interval [CI], 0.22-4.86), while the venous BA values increased (P = 0.026). The BEC values of high myopia were higher than those of low and moderate myopia groups. When the loss function of our data classification model converged to 0.09, the model accuracy reached 94.19%.

Conclusion: The progression of myopia is associated with a series of quantitative retinal vascular parameters, particularly the vascular angles. As the degree of myopia increases, the diversity of vascular characteristics represented by these parameters also increases.

Keywords: Artificial intelligence; Automated analysis; High myopia; Retinal vessels.

Fig. 1

Different refractive stages of retinal image feature extraction maps. Heatmaps were generated using the Attention module of the Transformer model. A represents original retinal images and various subsequent heatmaps for normal individuals, while B represents representative images for the low myopia group, and C, D represent the moderate and high myopia groups, respectively. The first column of images consisted of original high-resolution retinal images for four groups with different refractive statuses, including the retina, optic disc, macula, and posterior pole. Columns 2 through 6 displayed heatmaps generated by the Attention module, capturing the regions of interest

Fig. 2

Model workflow diagram. A primarily involves data quality analysis and classification. B carries out image segmentation and arteriovenous skeletonization based on the results from A. Building upon B, the model further determines vascular order (C) and measurement parameters

Fig. 3

Diagram illustrating vascular measurement parameters. Arteriole (or venule) vessels corresponding to the temporal side of the optic disc in the image are used as reference points. Numbers represent vascular orders. The angle formed by the last vessel with the highest vascular order on both upper and lower vessels intersecting at the optic cup is defined as MA. BA is formed by vessels with different vascular orders, while BEA is formed by vessels with the same vascular order. The calculation method for the branching coefficient is related to the diameter of the branching vessel and the diameter of the parent vessel

Fig. 4

The standard deviation bar chart displayed the differences in vascular parameters as arterial main angle (A), arterial branching angle (B), arterial bifurcation edge coefficient (C), venous main angle (D), venous branching angle (E), and venous bifurcation edge coefficient (F) between the four groups: normal, low myopia (LM), moderate myopia (MM), and high myopia (HM). Special markers (*) were considered to demonstrate a statistical difference between two sets of data