Artifact Correction in Retinal Nerve Fiber Layer Thickness Maps Using Deep Learning in Glaucoma

Abstract

Purpose : To develop a deep learning model to correct retinal nerve fiber layer thickness (RNFLT) artifacts in glaucoma and evaluate its clinical utility.

Methods : We developed a deep artifact correction model termed RNFLTCorrect. The artifacts are defined as RNFLT less than the literature's known floor value of 50 microns. We constructed the training (22,953 RNFLT maps) and testing (27,316 RNFLT maps) data for developing RNFLTCorrect and evaluating correction accuracy in the following way (Figure 1): ground truth output: using 50,269 high-quality RNFLT maps with artifact ratio < 0.02%; model input: masking the ground truth RNFLT maps with least artifacts with imaging artifact shapes sampled from 21,722 low-quality RNFLT maps with artifact ratio > 0.05%. We further evaluated the impact of artifact correction on visual field prediction and progression prediction using 24,257 patients with 111,966 cross-sectional OCT-VF pairs tested within 30 days with reliable VFs and 3,233 patients with 19,070 reliable longitudinal VF series with at least five measurements over at least four years from Massachusetts Eye and Ear Glaucoma service. T-test with bootstrapping sampling was used to compare the prediction accuracy with and without artifact correction measured by R-squared (R²) the area under the receiver operating characteristic curve (AUC) with cross-validation.

Results : The average mean absolute error (MAE) and Pearson correlation coefficient (R) of artifact correction for the RNFLT map were 9.89 microns and 0.90 (p < 0.001), respectively. Figure 2 (a) shows an artifact correction example that our model can accurately correct large artificially generated artifacts (artifact ratio: 62.8%) with high accuracy of R² 0.92 and MAE 10.6 microns. Artifact correction improved R² for VF prediction (Figure 2 [b]) in RNFLT maps with artifact ratio > 10% and artifact ratio > 20% up to 0.03 and 0.04 (p < 0.001), respectively. Artifact correction generally improved (p < 0.05) AUC for progression prediction in RNFLT maps with artifact ratio ≤ 10%, > 10%, and > 20%: (1) TD pointwise progression (at least 3 locations with TD slope ≤ -1 dB/year and p-value < 0.05): 0.68 to 0.69; 0.62 to 0.63; 0.62 to 0.64; (2) MD fast progression (MD slope ≤ -1 dB/year and p-value < 0.05): 0.67 to 0.68; 0.54 to 0.60; 0.45 to 0.56.

Conclusions : Artifact correction in RNFLT maps can improve VF and progression prediction in glaucoma.

This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.

 

Figure 1

 

Figure 2