Deep-learning enface image classifier analysis of optical coherence tomography angiography images improves classification of healthy and glaucoma eyes.

Abstract

Purpose : To compare gradient boosting classifier model (GBM) machine learning analysis of instrument-based OCTA vessel density measurements with deep learning convolutional neural network (CNN) classifier analysis of radial peripapillary capillary (RPC) enface vessel density images for classifying healthy and glaucomatous eyes.

Methods : 100 eyes from 51 healthy participants and 264 glaucomatous eyes from 185 patients from the Diagnostic Innovations in Glaucoma Study who underwent OCTA imaging of the optic nerve head (ONH) were enrolled in this observational cross-sectional study. Classification performance of a GBM trained and tested on standard OCTA ONH measurements was compared to performance of a pre-trained ResNet-50 CNN trained and tested on enface OCTA high density (HD) and non-HD optic nerve head 4.5 mm x 4.5 mm images. Images from 69 healthy eyes (of 35 subject) and 188 glaucomatous eyes (of 130 subjects) were used to train both models and an independent test set containing images from 31 healthy eyes (of 16 subjects) and 76 glaucomatous eyes (of 55 subjects), with no patient overlap between training and test sets, were used to compare model performance. Areas under the receiver operating characteristic (AUROC) curves adjusted for age and the inclusion of both eyes per subject were calculated.

Results : The glaucoma test subjects were significantly older (mean [95% CI], 71.9 [68.7, 75.1] years and 59.5 [53.4, 65.6] years; P < 0.001) and glaucoma test eyes had a significantly worse visual field MD (-6.9 [-8.5, -5.2] dB and -0.53 [-0.2, 0.8] dB; P < 0.001) compared to the healthy test eyes. The adjusted AUROC using GBM was 0.82 (0.77, 0.84) for RPC vessel density measurements, 0.85 (0.80, 0.87) for RPC capillary density measurements and 0.86 (0.81, 0.87) for combined vessel and capillary RPC density measurements. The adjusted AUROC using CNN analysis of RPC enface vessel density images was 0.91 (0.88, 0.94) resulting in improved classification compared to all GBM results (P < 0.05 for all comparisons). Two randomly selected images from the test set with associated probabilities of glaucoma by CNN analysis are shown below .

Conclusions : Deep learning enface OCTA measured vessel density image analysis can improve on instrument-based GBM models for classifying healthy and glaucoma eyes.

This is a 2021 ARVO Annual Meeting abstract.

 

Probability of glaucoma = 0.03

 

Probability of glaucoma = 0.68