Abstract
Purpose :
To assess the repeatability of estimates of mean deviation (MD) and visual field index (VFI) obtained from an automated deep-learning approach that analysed raw OCT volumes.
Methods :
OCT scans were acquired from both eyes of 138 healthy, 743 glaucoma suspects and 941 glaucoma patients (Cirrus HD-OCT scanner, 200x200 ONH Cubes, Zeiss, Dublin CA). The scans were acquired at multiple visits, with two or more scans acquired at each visit. Scans with signal strength < 7 were discarded, giving us a total of 19,208 OCT scans. A subset of 5207 eyes (total of 10,414 scans) had repeat scans of that met the inclusion criteria. 24-2 Humphrey visual field (VF) tests were administered at each visit. A single convolutional neural network was trained to estimate the MD and VFI (dual outputs) from downsampled OCT volumes (50x50x128 voxels). The network consisted of 5 convolutional layers, followed by a global average pooling layer and dual outputs to enable the simultaneous estimation of MD and VFI. A mean squared error loss was used to train the network using an Adam optimiser over a total of 200 epochs. A 10-fold cross-validation scheme was used, where the dataset was divided into 10 non-overlapping folds (~182 subjects per fold) – trained on 8-folds, validated on one and tested on one. Each subject was limited to a unique fold. The performance of the method was assessed by computing the median error and interquartile range. The repeatability was assessed using a set of 5207 OCT scans that had repeats available.
Results :
The median absolute error (Q1, Q3) for the estimates of MD and VFI were 1.66 (0.79, 2.99) dB and 3.01 (1.48, 6.63) %, respectively. In the reproducibility test, the Pearson’s correlation coefficient was 0.91 (CI: [0.91, 0.92]) and 0.91 (CI: [0.90, 0.92]), for MD and VFI, respectively. The median absolute difference between the repeated estimates for MD and VFI were 0.53 (0.21, 0.51) dB and 1.17 (0.45, 1.14)%, respectively.
Conclusions :
The deep-learning based approach for estimating visual field test parameters shows repeatability better than expected test-to-test variability.
This is a 2021 ARVO Annual Meeting abstract.