CNN-Based Device-Agnostic Feature Extraction From ONH OCT Scans

• DOI: 10.1167/tvst.13.12.5
• Publication Date: 2024-12-03T15:52:51Z
• Authors: Sjoerd J. Driessen; Karin A. van Gard...; Danilo Andrade De...; Luisa Sanchez Brea; João Barbosa-Breda; Bart Liefers; Hans G. Lemij; Doreen Nelson-Ayifah; Angelina Ampong; Pieter W. M. Bonn...; Alberta A. H. J. ...; Caroline C. W. Kl...
• Keywords: Cirrus

Purpose: Optical coherence tomography (OCT)-derived measurements of the optic nerve head (ONH) from different devices are not interchangeable. This poses challenges to patient follow-up and collaborative studies. Here, we present a device-agnostic method for extraction OCT biomarkers using artificial intelligence. Methods: ONH-centered volumes Heidelberg SPECTRALIS, ZEISS CIRRUS HD-OCT 5000, Topcon 3D OCT-1000 Mark I/II OCT-2000 were annotated by trained graders. A convolutional neural network (CNN) was on these segmented B-scans utilized obtain several ONH biomarkers, such as retinal fiber layer (RNFL) minimal rim width (MRW). The CNN results compared between manufacturer-reported values an independent test set. Results: intraclass correlation coefficient (ICC) circumpapillary (cpRNFL) at 3.4 mm reported 0.590 (95% confidence interval [CI], –0.079 0.901), our resulted in cpRNFL ICC 0.667 CI, –0.035 0.939). 3.5 CIRRUS, OCT-2000, SPECTRALIS generated 0.656 0.055–0.922). Comparing global mean MRWs manufacturer yielded 0.983 0.917–0.997). MRW among 0.917 0.947–0.981). Conclusions: Our feature scans showed higher reliability than measures manufacturers cpRNFL. very well manufacturers. Translational Relevance: open-source software can robustly extract wide range any device, removing dependency manufacturer-specific algorithms, which has significant implications research.