A deep learning method for Bruch’s membrane segmentation for choroidal analysis in SS-OCT

Abstract

Purpose : Reliable choroidal and choriocapillaris analysis depend on reliable Bruch’s membrane (BM) segmentation for diagnosis of retinal diseases. We propose a novel deep learning (DL) algorithm to segment BM on swept-source optical coherence tomography (SS-OCT) volume scans.

Methods : The multilayer segmentation (MLS) was used to generate the ground truth followed by manual inspection to exclude segmentation errors. Ground truth is defined as the region between inner limiting membrane (ILM) and BM. Patches with size of 3 mm were extracted from five neighboring fast B-scans as the inputs. A total of 14,098 training and 11,360 validation data from 603 volumes acquired using PLEX^® Elite 9000 (ZEISS, Dublin, CA) including 3x3 mm, 6x6 mm, 9x9 mm, 12x12 mm scans, with various eye diseases, were used to train a novel DCT-network architecture (Bagherinia et al., IOVS June 2022, Vol.63, 2060), and generate a model. An uncertainty algorithm detects the scan regions with low segmentation confidence that is then replaced by interpolation to generate the complete 2D BM surface. The choroidal-scleral interface (CSI) segmentation was generated using the method in the same publication. Performance of the algorithm was evaluated using 35 12x12 mm scans from 28 subjects (one or both eyes) with various eye diseases. Choroidal thickness maps using manual and automated segmentation were generated, defined as the distance between the BM and the CSI. Regression analyses of choroidal thickness between manual and automated method were performed for 1mm, 3mm, 6mm, 9mm, and 12mm zones of the extended ETDRS grid (Fig 2). Regression plots were used for 1mm, 3mm, 6mm, 9mm and 12mm zones.

Results : Fig 1 shows an example for five neighboring 3 mm OCT patches and five ground truth patches as input and output, and examples for BM segmentation results overlaid on OCT B-scans. Fig 2 shows the ETDRS and correlation between the manual and automated method in each sector including the regression plots for the central inputs 3mm, 6mm and 12mm zones. Most ETDRS sectors show significant correlation (>0.85) between manual and automated measurements.

Conclusions : We proposed a deep learning solution for BM segmentation. The choroidal thickness maps generated by automated and manual segmentations have a strong correlation. Automated segmentation may be a valuable diagnostic tool for retinal diseases.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.