Choroidal vascular fingerprints from indocyanine green angiography unveil chorioretinal disease state

Abstract

Purpose: To develop an annotation-efficient deep learning algorithm for extracting multi-dimensional features of choroidal vasculature on indocyanine green angiography (ICGA) images via a human-in-the-loop (HITL) strategy and explore their relationship with multiple chorioretinal diseases. Methods: The segmentation model was trained on a multi-source dataset that included both 55° ICGA and 200° ultra-widefield ICGA (UWF-ICGA) images, using a HITL strategy. Choroidal vascular fingerprints were generated from the segmentation maps, quantifying diameter, density, complexity, tortuosity, and branching angle. Reliability was assessed using intraclass correlation coefficients (ICC), and normal ranges for each measurement were estimated. The study retrospectively included 243 eyes diagnosed with central serous chorioretinopathy (CSC), polypoidal choroidal vasculopathy (PCV), or pathological myopia (PM), as well as 151 normal control eyes, to investigate their association with choroidal vascular fingerprints. Multivariable logistic regression models were used for the analysis. Results: The model achieved high segmentation accuracy, with the area under the receiver operating characteristic curve being 0.975 (95% confidence interval [CI, 0.967–0.983) for 55° view ICGA images and 0.937 (95% CI, 0.914–0.960) for UWF-ICGA images. Twenty-six, 28, and 29 multidimensional measurements were significantly associated with CSC, PCV, and PM, respectively (P value < 0.05). The ICC values for 74 choroidal vascular measurements ranged from 0.71 (95% CI, 0.51–0.84) to 0.97 (95% CI, 0.95–0.99). Conclusions: This pioneering study revealed choroidal vascular fingerprints and validated their associations with various chorioretinal diseases. These findings pave the way for future exploration of the pathological mechanisms underlying these conditions.