A Deep Autoencoder Model to Denoise Visual Fields in Glaucoma

Abstract

Purpose : To denoise visual fields (VFs) in glaucoma by developing a deep autoencoder model.

Methods : We selected reliable SITA Standard 24-2 VFs from a multi-center dataset excluding Massachusetts Eye and Ear (MEE) data. A 2D convolutional autoencoder was developed to denoise VFs using the multi-center dataset for training. The noise is defined as the difference between the original VFs and the reconstructed (denoised) VFs from our deep autoencoder model. Our autoencoder model was tested on the MEE dataset. To demonstrate the effectiveness of our denoising model, we compared the structure-function relationships between linear regression models using 12 clock-hour retinal nerve fiber layer thickness (RNFLT) from reliable optical coherence tomography (OCT, signal strength at least 6) to predict the 52 TDs of the paired (within 3 months) original VFs and the denoised VFs. Model selection based on Bayesian information criteria (BIC) was applied to remove redundant features from the structure-function linear models. Adjusted r-squared and BIC penalized for linear model complexity were used to measure the structure-function relationship strength.

Results : A total of 419,755 of VFs from 204,619 eyes of 116,288 patients (age: 64.17 ± 15.42 years; mean deviation: -4.93 ± 6.10 dB) in the multi-center dataset excluding MEE data were used to train our deep autoencoder. There were 9,456 OCT-VF pairs from 4,609 eyes of 2,962 patients from MEE (age: 61.76 ± 15.53 years; mean deviation: -5.04 ± 7.01 dB) available to test our denoising autoencoder. Two examples of the original VF and denoised VFs are shown in Figure. 1 (a-d), where we observe extreme values (noise) to be mitigated after being denoised by our autoencoder model. The absolute average VF differences between original and denoised VF were greater in the central sector (Figure. 1 [e]). Smaller absolute denoising amount was correlated with worse TD values and the correlations were greater in inferior paracentral region (Figure. 1 [f]). The denoised VFs (Figure. 2) can be better predicted by the 12 clock-hour RNFLTs than the original VF with R-squared and BIC improvements up to 0.02 and 48 (BIC improvement ≥ 6: strong improvement) in both 3,672 left and 5,784 right eyes, and the improvement is relatively more conspicuous in superior hemifield.

Conclusions : Denoising VFs with our deep autoencoder model can improve structure-function correlation in glaucoma.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1

 

Figure 2