Clustering Spectral-Domain Optical Coherence Tomography Images using a Deep Variational Auto-encoder

Abstract

Purpose : To learn a low-dimensional representation of spectral-domain optical coherence tomography (SD OCT) peripapillary images that can be used to classify images into glaucoma versus healthy eyes.

Methods : The study included 23,992 Spectralis SD OCT images from 1,336 eyes, of which 30% were healthy and the remaining were glaucoma or glaucoma suspects. The definition of groups was based on visual fields and inspection of the optic nerve. In order to learn a low-dimensional representation of the high-dimensional SD OCT images a variational auto-encoder (VAE) was used, an unsupervised deep learning technique. The encoder and decoder of the VAE were artificial neural networks (ANN), with fully connected hidden layers; and were trained on 80% of the data, with the remaining images used in a test set. Randomization was performed at the patient level. Clustering from the VAE was performed using the learned latent representation and was compared to retinal nerve fiber layer (RNFL) thickness values, both global and by sectors. To account for within eye dependencies, generalized estimating equations was used. Comparisons were made using area under the receiver operating curve (AUC) values (type 1 error of 0.05).

Results : The number of latent dimensions in the VAE was allowed to vary from 1 to 100, and the latent space proved to be clinically valuable (Figure 1). The AUC for classifying eyes into healthy versus glaucoma and glaucoma suspects was 0.84 (0.82, 0.85) for global RNFL thickness and 0.88 (0.86, 0.90) for a model that also included sectoral measures. The model that included the VAE latent representation had an AUC of 0.93 (0.91, 0.94), a statistically significant improvement (Figure 2).

Conclusions : The VAE was shown to effectively learn a clinically relevant latent representation for SD OCT images that was valuable for clustering images into healthy eyes versus glaucoma and glaucoma suspect eyes. The improved performance of the VAE indicates that there is important information encoded in the raw SD OCT image that is missed in standard RNFL thickness measures.

This is a 2020 Imaging in the Eye Conference abstract.

 

Latent representation from the variational auto-encoder using two dimensions with the global retinal nerve fiber layer thickness overlaid.

 

Area under the receiver operating curve (AUC) values for standard retinal nerve fiber layer (RNFL) thickness measures versus the variational auto-encoder across latent dimensions.