Learning Optical Scattering through Symmetrical Orthogonality Enforced Independent Components for Unmixing Deep Tissue Photoacoustic Signals

Noninvasive mapping of chromophore distribution in deep tissue is highly desired in numerous biomedical applications. Multispectral photoacoustic (PA) imaging fused with linear spectral unmixing is employed to generate 2-D and 3-D maps of tissue chromophores. However, wavelength and depth dependent attenuation of optical fluence leads to nonuniform variations in the spectral behavior of biomolecules affecting the accuracy of conventional linear unmixing methods. To address this, a modified independent component analysis (ICA) model is proposed to construct adaptive and statistically independent components for each chromophore from the given spectral PA data. The model was trained with custom designed multispectral PA imaging simulations consisting of both endogenous (oxy- and deoxy-hemoglobin) and exogenous (indocyanine green - ICG) chromophores inside 30 mm deep tissue. End-to-end unsupervised nature of the proposed approach made the process independent of human labeling and outperformed the standard linear spectral unmixing method when tested on a set of simulated phantoms, experimental phantoms, and in vivo mouse imaging data.