GAN-derived automated choroid and retinal segmentation using swept-source optical coherence tomography images

Abstract

Purpose : Retinopathy of prematurity (ROP) is traditionally staged using fundus images; however, advancements in portable ultra-widefield (UWF) optical coherence tomography (OCT) now allow for detailed cross-sectional images of various structures, including the choroid and retina, which may provide additional insight into disease progression. Measurements of these potential biomarkers require manual segmentations, which are tedious, time-consuming, and impractical given the large volumes of data. Therefore, we hypothesized that a generative adversarial network (GAN) automated segmentation would perform similarly to manual segmentation.

Methods : We used previously collected UWF-OCT scans from patients with ROP from the Oregon Health & Science University (OHSU) neonatal intensive care unit (NICU). Two volumes of B-scans (779) from a single patient were averaged down to 156 frames, manually segmented by two graders for the choroid and retina using custom-built tools in Napari image viewer, and interpolated up to 776 frames. A GAN, pix2pixHD, was trained on these interpolated scans to perform automated segmentation of the retina and choroid. We then manually segmented 10 B-scans each from 5 additional patients (total of 50-B-scans) and compared the results to the GAN segmentation using an F-score.

Results : Automated GAN segmentation of UWF-OCT images was moderately accurate. The retinal and choroidal segmentation F-Score ± standard deviation was 0.81 ± 0.06 and 0.86 ± 0.06, respectively.

Conclusions : Our results showed that GAN-automated choroid and retinal segmentation of UWF-OCT images performed similarly to manual segmentation, with moderate levels of agreement between the two methods. The results demonstrate that a GAN has the potential to automate the segmentation of a variety of UWF-OCT structures and could aid in defining UWF anatomic features, and eventually growth trajectories, in ROP.

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

 

Figure 1. From left to right, examples A & B depict two UW-OCT B-scan images, manual segmentations, and then GAN-automated segmentations of the retina (red) and choroid (green).