Deep Learning Models Based on Unsegmented OCT RNFL Circle Scans Provide Accurate Detection of Glaucoma and High Resolution Prediction of Visual Field Damage

Abstract

Purpose : To apply deep learning (DL) to unsegmented optical coherence tomography (OCT) scans to identify glaucoma, predict global visual field (VF) results, and create high resolution maps of visual function.

Methods : Imaging and VF measurements were collected from a cohort of 2,014 healthy, suspect, and glaucoma participants (3,882 eyes). Imaging data consisted of 68,055 unsegmented Spectralis circle scans. Functional measurements included 15,717 24-2 visual field (VF) tests and 5,155 10-2 VF tests. Participants were randomly divided into independent training (85%), validation (5%), and test (10%) sets. Resnet50 architecture was used to identify healthy vs. glaucoma eyes and predict global VF metrics (MD, PSD, VFI) from unsegmented circle scans. U-Net architecture was used to predict individual 24-2 and 10-2 test points to provide high resolution maps of visual function. Evaluation used area under the receiver operating characteristic curve (AUC), mean absolute error (MAE), and R². For comparison, predictions from mean retinal nerve fiber layer (mRNFL) thickness were also evaluated.

Results : DL outperformed mRNFL thickness in terms of AUC [95% CI] for detecting any glaucoma (0.84 [0.76 – 0.89] vs. 0.79 [0.70 – 0.86]), moderate-to-severe glaucoma (0.97 [0.94 – 0.98] vs. 0.94 [0.90 – 0.97]), and mild glaucoma (0.75 [0.66 – 0.83] vs. 0.70 [0.59 – 0.79]). DL prediction R² [95% CI] was significantly (p<0.001) higher than mRNFL thickness for both 24-2 MD (0.77 [0.66 – 0.86] vs. 0.41 [0.30 – 0.53]) and 10-2 MD (0.83 [0.75 – 0.89] vs. 0.41 [0.27 – 0.55]). DL predictions of 24-2 VF MD, PSD, and VFI achieved MAE [95% CI] of 1.8 dB [1.6 – 2.1], 1.1 dB [0.9 – 1.3], and 3.8 [3.1 – 4.7], respectively. In predicting 10-2 VF MD and PSD, MAE was 1.8 dB [1.5 – 2.1] and 1.0 dB [0.8 – 1.3], respectively. In predicting individual VF test points, the R² ranged from 0.07 dB to 0.71 for 24-2 VFs and from 0.01 to 0.85 for 10-2 VFs.

Conclusions : DL models of unsegmented circle scans significantly improves identification of glaucoma eyes, prediction of global VF metrics, and provide high resolution maps of visual function compared to circle scans. This suggests we can accurately predict disease stage and patterns of functional loss to help reduce reliance on and frequency of VF testing.

This is a 2020 ARVO Annual Meeting abstract.