Objective.An automated tool for choroidal segmentation and quantitative analysis under pathological conditions is currently lacking, hindering the exploration of choroidal structural changes in fundus diseases. This study aims to create a fully automated deep learning system for the quantitative analysis of the choroid with pathological changes, and to apply the system in analyzing the correlation between the choroidal structure and the severity of high myopia.Approach. A total of 2590 optical coherence tomography B-scan images of 1424 eyes of 1029 patients of high myopia from 3 hospitals were collected. We developed a curriculum learning-based system, including a two-stage U-net (TSU-net) and a post-process module for segmentation of the choroid, to calculate mean choroidal thickness (MCT) and choroidal vascularity index (CVI). The output of the images was statistically analyzed to explore the associations among MCT, CVI and the clinical characteristics of the patients.Main results. The Dice coefficient and IoU measures of choroid segmentation were 0.9221 and 0.8575, respectively. In a human-machine comparison, the system performed faster and better than a senior ophthalmologist. Statistical analysis demonstrated that, MCT is correlated with age, scan region, axial length, maculopathy type, and CVI, and CVI is correlated with scan region and MCT.Significance. A fully automated choroidal structural quantification system was developed. Clinical evaluation demonstrated that severity of high myopia is closely related to MCT but shows only a low correlation with CVI, suggesting that CVI may have little applicability in eyes with large anatomical structural variations. Future quantitative analysis of choroidal structure of large samples will enable exploration of the pathogenesis of additional fundus diseases.
Keywords: choroidal structure; deep learning; high myopia; optical coherence tomography; segmentation.
© 2022 Institute of Physics and Engineering in Medicine.