Abstract
Purpose :
Dimensionality reduction techniques are used to compress otherwise overwhelming quantities of data into more interpretable information. The purpose of our study was to determine if we could use t-distributed stochastic neighbor embedding (t-SNE) to automatically classify eyes based on their spectral signature on hyperspectral fundus imaging. Furthermore, we wanted to explore if these clusters would identify eyes with similar features of age-related macular degeneration (AMD) in an automated manner.
Methods :
We performed a pilot analysis in four eyes of four patients with intermediate AMD (drusen) who underwent imaging with a prototype hyperspectral fundus camera. We reshaped the original retinal image to the hyperspectral cube (16 wavelength channels: 460 to 630 nm) and implemented t-SNE for dimensionality reduction. Portions of the hyperspectral cube were cropped to create a 256-dimensional space with 8192 points per patient in that space as an input to t-SNE. t-SNE maps reduced these points to two dimensions, preserving clustering present in the high dimensional space (though unobservable given the high number of dimensions). We compared t-SNE output between eyes.
Results :
t-SNE successfully reduced the multi-dimensional hyperspectral data into two dimensional graphs for interpretation. We found that eyes of different patients naturally clustered together in the t-SNE output. Figure 1 shows clustering of different eyes.
Conclusions :
The use of hyperspectral imaging with t-SNE dimensionality reduction post-processing may provide an automatic method for differentiating retinal phenotypes in patients with AMD. We are continuing to analyze data using these methods in a total of 34 eyes with AMD. Further analysis is underway to refine clusters of t-SNE output and identify specific spectra that are associated with drusen types or AMD features, which may provide a new tool for classification and an improved understanding of the pathophysiology of AMD.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.