Background and Objective:Quantitative analysis of choroidal vessels is important for the diagnosis of choroidal related diseases and the cognition of the choroid. Previous studies focused on segmenting choroidal vessels in enhanced depth imaging optical coherence tomography (OCT) and swept-source OCT modalities. There is still a blank for the segmentation of 3D choroidal vessels in the spectral-domain OCT (SD-OCT) modality. Quantitative analysis of choroidal vessels is crucial for diagnosing choroidal-related diseases and understanding the choroid. Previous studies have focused on segmenting choroidal vessels in enhanced depth imaging optical coherence tomography (OCT) and swept-source OCT modalities. There is still a gap in the segmentation of 3D choroidal vessels in the spectral-domain OCT (SD-OCT) modality. Method:We propose a novel 3D CNN-based method, called Shape-aware Adversarial Networks (SAN), for segmenting choroidal vessels in SD-OCT images. To learn the 3D mapping of SD-OCT images and labels, we built a 3D U-shape network backbone with anisotropic down-sampling and up-sampling operations. To learn the morphological characteristics of choroidal vessels, we constructed an auxiliary output branch to predict a newly designed choroidal vessel shape representation generated by labels. We further introduced a label-conditional adversarial loss to enforce shape learning in a patch-based pipeline. The code for SAN is publicly available at: https://github.com/nicetomeetu21/SAN. Results:The effectiveness of SAN has been verified in data with myopia. The proposed model reported as measures of segmentation performance a mean dice similarity coefficient of 85.96% and a mean Hausdorff distance of 2.54, outperforming other 2D and 3D segmentation methods. We further applied our method in the statistical analysis of the choroidal feature between 100 eyes with low to moderate myopia and 108 eyes with high myopia. The results show that in all Early Treatment of Diabetic Retinopathy Study subfields, the choroid volumes and vascular volumes of the two groups are significantly different (P<0.01). The choroidal vascular indexes (CVI) of two groups are significantly different in most subfields (P<0.01) except for central and nasal inner subfields. The CVI of the high myopia group is higher than that of the low to moderate myopia group in most subfields except for the nasal inner subfield. Conclusions:The results demonstrate that the proposed method can accurately segment choroidal vessels from SD-OCT volumes. The quantitative analysis of choroidal biomarkers in a total of 208 myopia eyes shows the applicability of the proposed method in helping to study the choroid.