Multispectral imaging for pathogen identification using a filter wheel and smartphone: A frugal innovation approach

Abstract

To create an affordable clinical test for use in more decentralized settings, we are developing a multispectral imaging system based on a filter wheel and a smartphone. Our application is the label-free identification of uropathogens from images of bacterial colonies directly on a non-chromogenic culture medium. Using feature selection techniques, we have previously shown through calculations, the possibility to move from hyper- to multi- spectral imaging by exploiting less than 8 spectral bands. Here, we confirm our findings by testing a database of true multispectral images acquired using a filter wheel holding up to 22 dichroic bandpass filters with a 10 nm bandwidth. Performance is reported for 5 species using only 6 filters. Probabilistic SVM algorithms were implemented to allow to reject species other than the targeted most prevalent uropathogens as it is crucial to keep the false-positive rate low. Evolution of specificity and sensitivity with probability threshold are discussed in the light of probability frequency distributions.