Abstract
Purpose :
1) To develop a deep learning algorithm capable of examining the three-dimensional (3D) morphology of the optic nerve head (ONH) extracted from optical coherence tomography (OCT) scans, 2) To differentiate between eyes categorized as healthy (H), highly myopic (HM), affected by glaucoma (G), and those presenting a combination of high myopia and glaucoma (HMG).
Methods :
A total of 685 scans from 754 participants (256 H, 94 HM, 227 G, and 108 HMG) in a cross-sectional comparative study, excluding those with pathological myopia. Three retinal and four connective tissue layers of the ONH were delineated from OCT scans using Reflectivity (Abyss Processing Pte Ltd), and the layer boundaries were transformed into 3D point clouds (PC). OCT raster scans were converted into radial scans (RScans), and six RScans were extracted at 300 intervals from each OCT volume. A novel ensemble PointNet network was devised to simultaneously analyze 3D PCs and segmented RScans, exclusively utilizing the structural information of the ONH for 4-class classifications (H, HM, G, and HMG). Three tests (Test1: classification solely with 3D PCs, Test2: classification solely with RScans, Test3: classification using both 3D PCs and RScans) were performed and the network's performance was assessed by reporting the area under the receiver operating characteristic curves (AUCs) for each class (one-vs-all).
Results :
Our network achieved a very good distinction among the four classes, exhibiting a micro average AUC of 0.93±0.03 (Test 3). The one-vs-all AUC values were 0.95±0.02 for H, 0.90±0.02 for HM, 0.90±0.04 for G, and 0.89±0.02 for HMG. When solely utilizing PCs, the network showcased an AUC of 0.91±0.03 (Test 1), and with six segmented RScans, it demonstrated an AUC of 0.90±0.02 (Test 2). Notably, the classification AUC for Test 3 was significantly higher (p<0.001) compared to other cases.
Conclusions :
The devised ensemble PointNet effectively distinguished between myopic and glaucomatous eyes with high accuracy. While our classification performance was outstanding, it's crucial to highlight that validation in a considerably larger population is essential for clinical acceptance. Our approach holds promise in potentially discerning high myopic eyes from glaucomatous ones, relying solely on the 3D morphology of the ONH.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.