Spectral Band Subset Selection for Discrimination of Healthy Skin and Cutaneous Leishmanial Ulcers

Abstract

Leishmaniasis is a parasitic disease, transmitted by the bite of an insect that has previously fed on an infected host. One of its clinical forms is Cutaneous Leishmaniasis - CL and due to its increasing incidence, it is necessary to create effective and easy-use diagnostic methods. In this paper, we assess two unsupervised band-selection algorithms that allow the dimensional reduction of hyperspectral data taken from CL ulcers, maintaining a high classification accuracy. This is an important task for the development of an non-invasive system based on multispectral imaging, that support the diagnosis and treatment follow-up of cutaneous ulcer caused by Leishmaniasis. Spectral data was obtained in golden hamsters subjected to varying conditions of infection. Two algorithms, one based on similarity and the other based on singular values decomposition, are implemented using MATLAB functions and are applied to the spectral data. The selected subsets of bands are used to classify the spectra into healthy skin, border and ulcer centers using support vector machines - SVM and neural networks - NN. The obtained results are represented in precision tables and allow to observe that both methods achieve an appropriate dimensional reduction of multispectral data without losing key information for their subsequent classification. At the end, we show that it is possible to obtain a subset of spectral bands to discriminate between healthy skin and cutaneous ulcers caused by Leishmaniasis.