The usefulness of the Deep Learning method of variational autoencoder to reduce measurement noise in glaucomatous visual fields

The aim of the study was to investigate the usefulness of processing visual field (VF) using a variational autoencoder (VAE). The training data consisted of 82,433 VFs from 16,836 eyes. Testing dataset 1 consisted of test-retest VFs from 104 eyes with open angle glaucoma. Testing dataset 2 was series of 10 VFs from 638 eyes with open angle glaucoma. A VAE model to reconstruct VF was developed using the training dataset. VFs in the testing dataset 1 were then reconstructed using the trained VAE and the mean total deviation (mTD) was calculated (mTD(VAE)). In testing dataset 2, the mTD value of the tenth VF was predicted using shorter series of VFs. A similar calculation was carried out using a weighted linear regression where the weights were equal to the absolute difference between mTD and mTD(VAE). In testing dataset 1, there was a significant relationship between the difference between mTD and mTD(VAE) from the first VF and the difference between mTD in the first and second VFs. In testing dataset 2, mean squared prediction errors with the weighted mTD trend analysis were significantly smaller than those form the unweighted mTD trend analysis.