U-Net-derived automated ultra-widefield retinal vessel segmentation using swept-source optical coherence tomography images

Abstract

Purpose : In retinopathy of prematurity (ROP), plus disease, which is determined by the dilation and tortuosity of the retinal vessels is a major determinant in staging and treatment. Human grading of plus disease, typically based on indirect ophthalmoscopy or widefield digital fundus photographs, can have significant inter-rater variation, affecting management. Development of machine learning (ML) based metrics for plus disease has the potential to minimize this inter-rater variation. ML models will require a large number of accurate retinal vessel segmentations for quality training. Portable ultra-widefield (UWF) optical coherence tomography (OCT) can provide detailed, high resolution en face imaging of the retinal vasculature. In this study we hypothesize that a U-Net developed on fundus photographs can be adapted to produce retinal vessel segmentations from UWF OCT scans.

Methods : En face images were generated from previously collected 140 degree UWF-OCT scans from patients over the course of routine ROP screening from the Oregon Health & Science University (OHSU) neonatal intensive care unit (NICU). 23 en face images generated from a single patient across multiple examinations were segmented for retinal vasculature by a single grader. A U-Net with a pre-trained EfficientNet-B5 backbone was trained on these images to perform automated segmentation. A further 3 patients en face images were generated and manually segmented for validation and testing of the network (6 additional en face images). These manual segmentations were compared to the results of the trained U-Net using an F-score.

Results : Automated U-Net segmentation of UWF-OCT images captured core features of the retinal vasculature as evidenced by an F-Score ± SD was 0.57 ± 0.02.

Conclusions : The results demonstrate that U-Net automated UWF retinal vessel segmentation of OCT images is a promising approach for aiding analysis of ROP patient data. We obtain statistically significant agreement with manual segmentation given a minimal set of training data. With an adequately large dataset, the U-Net approach has the potential to automate the segmentation of a variety of biomarkers in UWF-OCT images. Moving forward ML algorithms for assessing vessel tortuosity and dilation can be applied to aid in the assessment of plus diseases in UWF-OCT images.

This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.