Predicting central 10 degrees visual field from peripapillary optical coherence tomography using deep learning approach

Abstract

Purpose : To develop a deep learning (DL) simulation of standard automated perimetry (SAP) in the central 10^° based on spectral domain optical coherence tomography (SD-OCT) retinal nerve fiber layer thickness (RNFLT).

Methods : This study included 5352 pairs of SD-OCT and 10-2 SAP from 1365 eyes of 724 healthy, glaucoma suspect and glaucoma patients from the Diagnostic Innovations in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES). Each pair of SD-OCT and 10-2 SAP was obtained within a 6-month duration. The dataset was randomly divided into training (65%), validation (15%), and test (20%) sets at the patient level. A unidimensional DL convolutional neural network was used to map all 768 peripapillary RNFLT values of SD-OCT to 68 sensitivity thresholds of 10-2 SAP. Visual field indices were generated using CNN-predicted sensitivity thresholds at 68 10-2 SAP test locations and included Total deviation (TD) values, Pattern deviation (PD) values at each test location, Mean deviation (MD) and pattern standard deviation (PSD). The accuracy of the model at each test location was evaluated by calculating the average mean absolute error (MAE) and the Pearson correlation coefficients (r) of predicted and actual TD values. Global accuracy of the model was evaluated by calculating the MAE and r for MD and PSD predictions.

Results : The DL TD predictions had an average MAE of 4.04 dB and r of 0.59 (P < 0.001) over the 68 10-2 SAP test points. This model was capable of predicting 10-2 SAP MD and PSD with MAE (r) of 2.88 dB (0.74) and 2.30 dB (0.59), respectively (Fig. 1).

Conclusions : An artificial intelligence approach was capable of reconstructing SAP central 10^° using peripapillary SD-OCT RNFLT measurements. Peripapillary RNFLT provides information about central visual field in glaucoma.

This is a 2021 ARVO Annual Meeting abstract.

 

MAEs and correlation coefficients between the predictions from the CNN (Left) and linear regression (LR) (Right) and the total deviation for each of the 68 locations tested in 10-2 VF (Right eye format). The number inside each square represents the value for that specific location and the grayscale (darker colors represent lower MAEs and higher correlations with the measured values; lighter colors represent higher MAEs and lower correlation coefficients) illustrates the performance.