Measuring Spectral Inconsistency of Multispectral Images for Detection and Segmentation of Retinal Degenerative Changes.

Multispectral imaging (MSI) creates a series of en-face fundus spectral sections by leveraging an extensive range of discrete monochromatic light sources and allows for an examination of the retina's early morphologic changes that are not generally visible with traditional fundus imaging modalities. An Ophthalmologist's interpretation of MSI images is commonly conducted by qualitatively analyzing the spectral consistency between degenerated areas and normal ones, which characterizes the image variation across different spectra. Unfortunately, an ophthalmologist's interpretation is practically difficult considering the fact that human perception is limited to the RGB color space, while an MSI sequence contains typically more than ten spectra. In this paper, we propose a method for measuring the spectral inconsistency of MSI images without supervision, which yields quantitative information indicating the pathological property of the tissue. Specifically, we define mathematically the spectral consistency as an existence of a pixel-specific latent feature vector and a spectrum-specific projection matrix, which can be used to reconstruct the representative features of pixels. The spectral inconsistency is then measured using the number of latent feature vectors required to reconstruct the representative features in practice. Experimental results from 54 MSI sequences show that our spectral inconsistency measurement is potentially invaluable for MSI-based ocular disease diagnosis.