Textural Analysis of OCT En-Face Images Predicts Spatial Patterns of Retinal Neuronal Loss at the Macula

Abstract

Purpose : Retinal neuron loss can develop within a month of optic nerve damage and can be quantified by thinning of the inner retinal layers using optical coherence tomography (OCT). Our aim was to determine whether textural information from en-face OCT images provides a means to characterize retinal neuron loss independent of retinal thickness and even before thinning occurs. We developed a variational autoencoder (VAE) model that simultaneously models en-face texture and retinal nerve fiber layer (RNFL) thickness to derive the spatial pattern of the retinal neuron loss as a deviation from the normal map.

Methods : A VAE model (Fig. 1) was designed to extract retinal textural information from the macular en-face images of the RNFL, RNFL-GCL, and retinal pigment epithelium complex and to combine with the RNFL thickness map to encode spatial patterns of neuron loss at different stages of disease severity, using 36 latent variables. The VAE model was trained/tested using 2162/422 scans from 376/50 non-artertic anterior ischemic optic neuropathy (NAION) patients collected longitudinally from onset to six months.

Results : The VAE model successfully captured the spatial patterns of the en-face images and the thickness map, enabling the latent space (Fig. 2) to visualize the spatial patterns of retinal nerve loss. In the test dataset, comparing the estimated RNFL probability deviation map and the reference, the VAE model achieved a Spearman correlation coefficient of 0.8, an accuracy rate of 0.93 in identifying the thinning/normal regions, and a mean difference of -0.10 ± 13.20 percentiles.

Conclusions : The reported VAE model successfully utilized textural features to derive the spatial pattern of neuron loss at the macula. This VAE approach will be used to determine whether textural information can predict disease progression before thickness changes in the retina.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

The proposed VAE architecture. The estimator (highlighted in purple) used 36 latent variables to estimate the RNFL thickness map and probability deviation map.

 

Latent spaces of the channel of (a) the RNFL en-face images (red/blue indicating bright/dark intensity), and (b) the RNFL thickness maps (red/blue indicating thick/thin thickness).