Abstract
Purpose :
Visual field (VF, Humphrey Field Analyzer, Carl Zeiss Meditec) measurement is very useful for the diagnosis and tracking of glaucoma or neurological diseases. But its unique measurement method introduces some noise in the results. This study aims to reduce the noise in the VF measurement by generative models.
Methods :
The generative model generates VF from 10 independent normally distributed variables using Wasserstein GAN. The training dataset includes 85988 visual fields from the electronic health records at Moorfields Eye Hospital in London. With the trained generative model, we generate a new dataset of 94059 visual fields. And then we use a dataset including 9314 visual fields from the United Kingdom Glaucoma Treatment Study (UKGTS) as the validation set. For every VF in the validation set, we search the closest VF in the generated dataset to validate the coverage of the generator. And for every VF in the generated set, we search the closest VF in the validation dataset to validate the correctness of the generator. After we get a stable generative model, we fix the parameters of the neural network, set the real VF from the test-retest dataset as the input, and optimize the latent variable to get the closest VF with the input VF. The converged latent variable is used to generate predicted VF. The variance and expectation of predicted VF for every person are compared with that of the test-retest dataset.
Results :
For every VF in the test-retest dataset, we predict a paired VF. Multiple real VF and predicted VF per person are analyzed with paired-samples t test, and 77.9% of the results believe they are not statistically different, when the significance coefficient is 0.01. The variance of the real VF and paired predicted VF for every person are scattered in Figure 1. 87.6% of the variance of predicted VF is smaller than that of real VF.
Conclusions :
Given a VF, the generative model can be used to predict a new VF. Compared with the origin VF series, the variability of predicted VF series is smaller with acceptable accuracy. The data quality in visual fields is enhanced by the generative model.
This is a 2020 ARVO Annual Meeting abstract.