Test of the Robustness of an OCT-based Method for Identifying Glaucomatous Damage Across Different Instruments

Abstract

Purpose : To test the Columbia University (CU) OCT-based method [1,2] for distinguishing glaucomatous damage from healthy controls by applying it to scans from different OCT instruments.

Methods : OCT glaucoma reports [3] were analyzed from 116 eyes from a prospective, observational, case-control study, included 54 healthy controls (HC), 32 early (EG, 24-2 mean deviation (MD) >-6dB), 12 moderate (MG, MD <-6dB and >-12dB), and 18 advanced glaucoma (AG, MD <-12dB). Eyes were scanned with two commercially available OCT instruments from different manufacturers, A and B. Each instrument uses different technology (swept-source vs. spectral-domain), segmentation algorithms, and normative data. A set of reports was generated for each device. These reports [3] included thickness and probability (p-) maps from both the retinal nerve fiber layer (RNFL) and ganglion cell layer (GCL), as well as a circumpapillary (cp) b-scan image and cpRNFL thickness plot. Four graders, experienced with the CU OCT-based method, categorized each eye as glaucoma (G), not glaucoma (NG), or uncertain (UNC) based upon clearly defined rules (Fig. 1).[1,2] Disagreements were adjudicated. Eyes with grades of UNC were treated as G in calculating specificity and sensitivity, as both grades require additional testing.

Results : For 113 (97%) of the 116 eyes, the grades were identical (Table 1). All 3 disagreements involved an UNC grade for either the A or B report. After combining eyes with UNC and G grades, the specificity for the 54 HC eyes was 98% (A) and 96% (B), while the sensitivity for the 62 G eyes was 100% for both A and B instruments.

Conclusions : The CU OCT-based method showed consistency across instruments from different manufacturers, with different normative groups and different algorithms. The sensitivity and specificity suggested that this method may help physicians distinguish glaucomatous eyes from healthy controls using OCT. 1. Liebmann et al; 2022; 2. Hood et al., PRER, 2022; 3. Hood PRER, 2017.

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