Deep Learning Estimation of 10-2 Visual Field Map Based on Macular Optical Coherence Tomography Angiography Measurements

Abstract

Purpose : To develop deep learning (DL) models estimating central visual field (VF) from optical coherence tomography angiography (OCTA) vessel density measurements.

Methods : This study included 3405 macula OCTA (1872 DIGS (Diagnostic Innovations in Glaucoma Study), and 1533 VO (Viterbi Ophthalmology )) from 1176 participants (1950 eyes). Data was split on the patient level into training (80%), validation (10%), and test sets (10%). DL ResNet50 models were trained on enface macula vessel density OCTA images to estimate 10-2 mean deviation (MD), pattern standard deviation (PSD), 68 total deviation (TD), and pattern deviation (PD) values and compared with linear regression (LR) model with the same input. Accuracy of the models was evaluated by calculating the average mean absolute error (MAE) and the R (Pearson correlation coefficients) of the estimated and actual VF values.

Results : DL models predicting 10-2 MD achieved R of 0.92 (95% confidence interval [CI], 86–0.96) for 10-2 MD and 0.94 (95% confidence interval [CI], 90–0.97) for 10-2 PSD and MAEs of 1.76 dB (95% CI, 1.39–2.17 dB) for MD and 0.79 dB (95% CI, 0.55–1.05 dB) for PSD. This was significantly better than mean linear estimates for 10-2 MD ([R: 0.58 95% CI, 0.42–0.71] and [MAE: 3.36 dB 95% CI, 2.58–4.22 dB]) and 10-2 PSD ([R: 0.54 95% CI, 0.37–0.68] and [MAE: 2.57 dB 95% CI, 2.05–3.10 dB]). The DL model outperformed the LR model for the estimation of pointwise TD values with an average MAE of 2.48 dB (95% CI: 1.99, 3.02) vs. 4.28 dB (95% CI: 3.46, 5.15) and R of 0.83 (95% CI: 0.76, 0.87) vs. 0.40 (95% CI: 0.26, 0.52) over all test points (Figure1). Similarly, the DL model outperformed the LR model for the estimation of pointwise PD values (Figure2). The DL model outperformed the LR model for the estimation of all sectors in 10-2 VF. Sector predictions ranged from MAE of 1.45(1.16, 1.77) dB (R=0.91 (0.80, 0.96) for inferotemporal sector to MAE of 2.99(2.21, 3.78) dB (R= 0.86 (0.76, 0.94) in superiornasal sector of VF.

Conclusions : DL models estimates functional loss from OCTA enface scans with high accuracy. Applying DL to the OCTA images may possibly lead to new biomarkers for clinical decision making and may lead to improved individualized patient care and risk stratification of patients who are at risk for central VF damage.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.