Abstract
Purpose :
Central serous chorioretinopathy (CSCR) is a vision-related disease characterized by the accumulation of subretinal fluid, leading to visual disturbances and potential vision loss. Recurrence of CSCR poses a significant challenge as it can have irreversible consequences on visual acuity. Therefore, there is a critical need to identify and understand the factors underlying CSCR recurrence. In this study, we aim to identify parameters derived from optical coherence tomography (OCT) linked to CSCR recurrence using statistics and machine learning.
Methods :
We collected data for 255 consenting patients (344 eyes, 5211 visits) diagnosed with CSCR at the Jules-Gonin Eye Hospital. We labeled the patients either recurrent or non-recurrent based on control performed by an expert ophthalmologist.
We automatically extracted the following parameters from OCT cube scans using the proprietary software Discovery® by RetinAI: thicknesses of outer nuclear layer (ONL), photoreceptor and retinal pigment epithelium layers (PR_RPE), choriocapillaris layer (CC_CS), retinal nerve fiber layer (RNFL), ganglion cell layer and inner plexiform layers (GCL_ILP), inner nuclear and outer plexiform layers (INL_OPL) and volumes of subretinal fluid (SRF), intraretinal fluid (IRF), pigment epithelium detachment (PED). Additional parameters were extracted from line scans using computer vision techniques: choroidal vascularity index (CVI), choroidal pachy vessel area (PV_AREA), disruption of the PR_RPE (DSCORE).
We performed statistical tests to determine whether any of these parameters were differently distributed between the recurrent and non-recurrent groups. Finally, we trained a logistic regression model to predict recurrence and we leveraged feature importance analysis methods to identify relevant parameters for the model prediction.
Results :
178 eyes were assigned to the recurrent group, and 109 eyes to the non-recurrent group. The following parameters were significantly differently distributed (i.e. p-values below 0.05) between the two groups: ONL, SRF, IRF, PED, DSCORE; the following were borderline (i.e p-values below 0.06): PR_RPE, CVI, age. Feature importance analysis of the logistic regression reported similar parameters in addition with RNFL, GCL_IPL, INL_OPL.
Conclusions :
We identified several predictive biomarkers for the recurrence of CSCR: age, layer thicknesses, volumes of SRF, IRF and PED, CVI, DSCORE.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.