Wavelet-Based Sparse Non-Negative Matrix Factorization for Bioluminescent Imaging Unmixing

Bioluminescence imaging (BLI) is an optical imaging technology used in preclinical oncology to estimate tumor size by spatio-temporal photon counting. It generally requires the biologist to hand-draw the region of interest around tumors to quantify its activity, which can be time-consuming and tricky when the subjects bear several tumors. In the latter context, we here propose to use a non-negative matrix factorization algorithm to separate the kinetics of each tumor, which is furthermore constrained by the prior prior knowledge of the tumors spatial appearance. Such prior is expressed using a sparsity constraint in a transformed wavelet domain. Results are presented on a cohort of 18 real BLIs, showing enhanced estimated spatial repartition of the tumors.