Representation and Reconstruction of Image-Based Structural Patterns of Glaucomatous Defects Using only Two Latent Variables from a Variational Autoencoder

• Published in: Ophthalmic Medical Image Analysis Vol. 12970; pp. 159 - 167
• Main Authors: Wang, Jui-Kai, Kardon, Randy H., Garvin, Mona K.
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Ganglion cell-inner plexiform layer (GCIPL); Glaucoma; Optical coherence tomography (OCT); Variational autoencoder (VAE)
• ISBN: 9783030869991; 3030869997
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-87000-3_17

Glaucoma can result in both diffuse and regional patterns of retinal neuron loss due to damage to their axons at the optic nerve head. However, most quantitative estimates of glaucomatous progression use a global average and do not capture underlying spatial patterns. Motivated by the need for quantitative methods for describing and visualizing the spatial patterns of neuron loss in glaucoma, we evaluate the feasibility of spatial modeling of macular ganglion cell plus inner plexiform layer (mGCIPL) thickness maps using a deep learning variational autoencoder (VAE). More specifically, after training from optical coherence tomography based mGCIPL thickness maps of glaucoma and normal subjects, our VAE model was able to (1) succinctly represent the pattern of mGCIPL thickness maps with only two latent variables (using the encoder part of the VAE), and (2) reconstruct individual mGCIPL thickness maps given just two latent variable values. Based on evaluation of reconstruction errors on the mGCIPL thickness maps from an independent testing set of glaucoma and normal eyes, our results demonstrate the promise of the VAE model for a succinct representation of patterns of glaucomatous damage as well as use of the latent space for visualizing these patterns.