Comparison of the deep learning-based choroid segmentation with and without optical attenuation corrected inputs

Abstract

Purpose : Segmentation of the choroid layer in optical coherence tomography (OCT) continues to be valuable in the analysis of retinal diseases. A novel approach using deep learning (DL) to identify the choroidal-scleral junction (CSJ) was previously described by Bagherinia [IOVS 2022; 63(7):2060] where the structural OCT and its optical attenuation correction (OAC) inputs were used. A similar approach was developed to segment Bruch's membrane (BM) [Bagherinia et al. IOVS 2023; 64(8):1115]. This study utilizes a DL-based CSJ segmentation algorithm with the structural OCT inputs only to compare performance of two DL choroidal segmentations using choroidal thickness measurements.

Methods : Patients were imaged by PLEX^® Elite 9000 (ZEISS, Dublin, CA) swept-source OCT (SS-OCT) using the Angio 6 mm x 15 mm x 15 mm (3072 x 834 x 834 pixels) scan pattern. 25 scans from 14 patients (one or both eyes) were used; various disease types including age-related macular degeneration (AMD) were included. Choroidal thickness maps (CTM) were generated and defined as the distance between the BM and the CSJ (Fig. 1). The CTM based on both CSJ segmentation methods were compared with those generated by the manual segmentation. The performance of the algorithms was reported using correlation between manual and automated methods for each subfield of the ETDRS grid (Fig. 2). Regression plots of each ETDRS zone were reported.

Results : Figure 2 shows the ETDRS grid and the choroidal thickness correlations for 25 datasets between manual and each DL-based method. Significant correlation (≥0.85) between manual and automated measurements can be found in all individual ETDRS subfields for the method with structural OCT inputs only. Overall, both CSJ segmentation methods perform well across the entire field compared to manual segmentation as shown in the regression plots.

Conclusions : In this study we compared manual and two DL-based CSJ algorithms for segmenting the choroid. The method with structural OCT inputs only performed well across the entire field, when compared to reference manual segmentations. Both automatic choroidal thickness methods may be valuable diagnostic tools for retinal diseases.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.