Quantification of Multispectral Photoacoustic Images: Unsupervised Unmixing Methods Comparison

Multispectral photoacoustic imaging is a functional modality based on the detection of ultrasound waves coming from tissues illuminated by laser pulses at different wavelengths. The specific photoacoustic behavior of different media for each illuminating wavelength allows their quantification which is of great interest for various medical applications. The quantification algorithm performances are related to the correct estimation of reference spectra for each medium to quantify. This study aims at comparing three different unsupervised methods to extract these reference spectra (Group Lasso with Unit sum and Positivity constraints, Vertex Component Analysis and Spatio-Spectral Mean-Shift). After the reference extraction, a supervised unmixing method called Fully Constrained Least-Square is used to estimate the medium concentrations. Using Vevo LAZR as acquisition system, the quantification performances are evaluated on a colored 4% agar phantom containing two pure media and a mix of both. The results highlight the suitability of the Spatio-Spectral Mean-Shift to extract reference spectra that allow the assessment of photoacoustic for dataset medium concentration.