Abstract
Purpose :
Optical coherence tomography (OCT) provides a structural quantification of the retinal nerve fiber layer (RNFL) in several diseases including glaucoma. Standard OCT segmentation algorithm works accurately in normal eyes but fails in disease conditions and requires manual correction which is labor intensive. The purpose of this study is to develop a deep learning (DL) based algorithm to automatically segment and measure the thickness of RNFL from SD-OCT images and compare it to manual segmentation in experimental glaucoma model of nonhuman primates (NHP).
Methods :
SD- OCT (Spectralis HRA + OCT) images from 24 cynomolgus monkeys induced with unilateral experimental glaucoma were dosed with either of the three test articles (n=8): Sham, Brimonidine DDS high dose (200 ug), and Brimonidine DDS low dose (67 ug). Each animal had 3 circumpappillary scans of 3.5 mm, 4.1 mm and 4.7 mm diameters centered on the optic nerve head. Animals were tracked weekly for the first 12 weeks, and then every 2 weeks for up to 24 weeks. Images were annotated and trained using Visopharam software (Visiopharm, Denmark). Automated segmentation of RNFL was achieved using a U-Net architecture. SD-OCT images (n= 1152) were divided into training (n=200) and test (n=952) sets. RNFL thickness was calculated every 50 pixels and performance of the model was evaluated by dice similarity coefficient (DSC), Bland-Altman plot and correlation analysis.
Results :
The DL model demonstrated accurate segmentation across all groups with average DSC of 0.98, 0.99, and 0.99 for sham and treated animals with 200 ug and 67 ug brimonidine, respectively. Global RNFL thickness measurements quantified using DL showed 93.5% accuracy and good correlation (0.971) when compared with manual segmentation. Bland - Altman analyses showed the two methods were in good agreement at 95% confidence level (mean ± 1.96 SD). The limits of the agreement are 4.15 µm and -12. 38 µm with mean bias -4.11 µm. The AI model took 2 hrs to segment RNFL compared to 20 days using manual segmentation.
Conclusions :
Our DL model demonstrated accurate segmentation and quantification automatically thereby reducing hundreds of hours in manual correction of segmentation errors. Future studies will focus on quantifying clinical images using DL and investigate disease progression and treatment outcome.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.