Deep learning prediction of 24-2 visual field map using en-face OCT-Angiography microvascular images in glaucoma

Abstract

Purpose : To develop deep learning models to predict the severity and pointwise patterns of 24-2 visual field (VF) damage in glaucoma using optic nerve head (ONH) and macular OCT-Angiography (OCTA) images.

Methods : This study included 3990 OCTA (1850 ONH, 2140 macula)-VF pairs from 465 Diagnostic Innovations in Glaucoma Study (DIGS) participants (842 eyes). Deep learning models were developed to predict individual 24-2 sensitivity thresholds using en-face ONH and macular OCTA images. Visual field indices, including mean deviation (MD), pattern standard deviation (PSD), and pointwise total deviation values, were then generated using the predicted sensitivity thresholds. The models' accuracy was evaluated using mean absolute error (MAE) and Pearson's correlation coefficient (r) and compared to linear regression (LR) models. An external independent dataset of 1394 clinic patients (2436 eyes) including 6227 OCTA (2033 ONH, 4194 macula)-VF pairs was used to evaluate the models' generalizability.

Results : The ONH deep learning models’ predictions had the respective MAE of 2.23 dB and r of 0.83 for MD and 2.59 dB and 0.78 for PSD. The MAE and r for the clinic dataset were 2.69 dB and 0.78 for MD and 2.50 dB, and 0.74 for PSD. The macular deep learning models had an MAE of 2.15 dB, an r of 0.78 for MD and 2.29 dB, and 0.68 for PSD. The respective MAE and r for the clinic dataset were 2.85 dB and 0.69 for MD and 2.62 dB and 0.69 for PSD (Figure 1). All investigated models significantly improved global and pointwise estimates of VF prediction compared to their respective LR models (P < 0.05).

Conclusions : Deep learning models can predict the severity and patterns of functional loss based on OCTA images. Artificial intelligence methods that accurately estimate the severity and patterns of glaucomatous functional loss from microvascular structure show promise in reducing the time and costs of VF testing, individualizing the frequency and test strategies of VF assessment, and preventing the development of irreversible blindness associated with the disease.

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

 

Figure 1. Models’ performance for the prediction of 24-2 visual field maps based on different image inputs (top: optic nerve head [ONH], bottom: macula) and test sets (left: DIGS, right: Clinic). Scatterplots demonstrate the relationship between models’ predictions and actual mean deviation values. DL: deep learning, LR: linear regression.