Beyond tessellation: AI-derived choroidal vessel density on multi-channel cSLO as a surrogate biomarker for OCT choroidal parameters.
Authors
Affiliations (6)
Affiliations (6)
- Department of Ophthalmology, Peking University People's Hospital, Beijing, China.
- Beijing Key Laboratory of Ocular Disease and Optometry Science, Beijing, China.
- EVision Technology (Beijing) Co. Ltd., Beijing, China.
- Aerospace Information Research Institute, Chinese Academy of Sciences (CAS), Beijing, China.
- Suzhou Microclear Medical Instruments Co., Ltd., Suzhou, Jiangsu, China.
- Faculty of Biomedical Engineering, Shenzhen University of Advanced Technology, Shenzhen, China.
Abstract
To develop an AI-based method for quantifying choroidal vessel density (CVD) from ultra-widefield confocal scanning laser ophthalmoscopy (UWF-cSLO) and to validate this metric against swept-source optical coherence tomography (SS-OCT) choroidal parameters. In this retrospective cross-sectional study, 633 eyes from 354 participants underwent simultaneous UWF-cSLO and SS-OCT imaging on an integrated platform, ensuring precise spatial co-registration. A TransUNet deep-learning model segmented choroidal vessels from red-channel (R; 665 nm) and infrared-channel (IR; 785 nm) images, yielding channel-specific metrics (R-CVD and IR-CVD) and a combined metric (IR + R CVD). Eight SS-OCT parameters-including choroidal thickness (CT), choroidal vascularity index (CVI), luminal area, stromal area (SA), total choroidal area (TCA), and vessel volume index (VVI)-served as reference measures. Associations were evaluated using Spearman correlations with participant-level clustered bootstrap confidence intervals, multivariable generalised estimating equations (GEE) adjusted for age and sex, and receiver operating characteristic (ROC) analysis. Segmentation accuracy exceeded 0.95 (Dice coefficient) in both channels. R-CVD exhibited the strongest inverse correlations with structural and volumetric OCT parameters (SA: <i>ρ</i> = -0.619; CT: <i>ρ</i> = -0.615; TCA: <i>ρ</i> = -0.615; all <i>q</i> < 0.001), whereas IR + R CVD correlated most closely with proportion-based indices (VVI: <i>ρ</i> = -0.34; CVI: <i>ρ</i> = -0.29). These channel-dependent patterns persisted after adjustment for age and sex in multivariable GEE models. On ROC analysis, R-CVD discriminated thin from thick choroids with an area under the curve (AUC) of 0.789 (sensitivity 72.5%; specificity 74.8%), whereas IR-CVD alone showed near-chance performance (AUC 0.517). AI-derived CVD from UWF-cSLO correlates strongly with SS-OCT choroidal parameters in a wavelength-dependent manner and may serve as an accessible surrogate for choroidal structural assessment in settings without OCT.