Multispectral imaging via feature selection: A frugal innovation approach for pathogen identification

Abstract

In order to develop an affordable clinical diagnostic instrument for use in more decentralized settings, we have assessed the feasibility to move from hyperspectral to multispectral imaging via parsimonious feature selection. The targeted application is the label-free identification at the species-level of uropathogens from images of bacterial colonies on their growth support. We show that the number of predictors (i.e., discrete spectral channels), can be dramatically reduced from 240 to less than 10 channels with limited performance loss. The impact of bandwidth is also investigated to consider the high degree of redundancy of raster images obtained by diffuse reflectance and propose a suitable design for a simple filter wheel based solution. Targeting the 8 most prevalent bacterial species responsible for > 80% of urinary tract infections, up to 94% of correct identification rate was reached using only 4 spectral windows.