Improving Reliability of Retinal Vessel Caliber Measurement Using Deep Convolutional Autoencoder

Abstract

Purpose :
Retinal microvasculature can be viewed and measured using color fundus photographs (CFP) and is predictive of systemic microvascular damage. Integrative Vessel Analysis (IVAN) is a semi-automated vessel measurement software that has been used in multiple epidemiological studies. Good focus and clarity of vessels is essential for reliable measurements. Poor image quality accounts for about 10 – 25% data loss using IVAN. This study represents a deep learning autoencoder approach to images and restoring the structural details required for IVAN.

Methods :
The dataset included de-identified CFP (field 1) with vessel measurements in IVAN from the Carotenoids in Age Related Macular Degeneration Study 2 (CAREDS2), an ancillary study of the Women’s Health Initiative. The training data was 511 images with reliable vessel measurements, as established by the IVAN software and approved by a human grader using defined criteria. Training images were intentionally degraded by shrinking their original resolution by a factor of 20 and then bringing back to their original size. A U-net based autoencoder was used to reconstruct the original images in 400x400 segments. The trained network was then applied to 12 images IVAN was originally unable to measure. Outcomes included (1) the percent of ungradable images that were considered gradable after autoencoder enhancement and (2) the mean differences between vessel measurements, and 3) sigma, a reliability index between original and enhanced cohorts, to identify any systematic bias.

Results :
Of the 12 images that were considered unreliable by IVAN, autoencoder could retrieve vessel caliber measures in 9 (75%). Two were ungradable due to intrinsic vascular structure and one continued to have poor focus. In the second dataset of gradable images (n = 10) vessel caliber pre and post autoencoder showed a mean difference of 6.03 microns (p = 0.41) for central retinal artery equivalent (CRAE) and 1.12 (p = 0.90) for central retinal vein equivalent (CRVE). The mean sigma for CRAE and CRVE was 3.26 and 3.83 respectively in the original dataset and 4.13 and 4.76 in the autoencoder datase

Conclusions :
A deep convolutional autoencoder improves the vascular details on retinal images and enables IVAN measurements, reducing data loss due to inferior quality photographs. Vessel diameter is not altered due to autoencoder enhancement and can be reliably used in longitudinal data.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.