Estimation of Tissue Perfusion Parameters with Multispectral Images and Machine Learning Regression Models

The study investigated various regression strategies to estimate perfusion characteristic maps during an arterial occlusion protocol using multispectral images. The proposed approach involves a preliminary nonlinear unmixing stage utilizing the Nonlinear Extended Blind End-Members and Abundance Extraction (NEBEAE) method for feature extraction. Subsequently, machine learning regression models are employed to estimate perfusion characteristics, specifically the change in oxygenated hemoglobin
$(\Delta HbO_{2})$
. As a reference, punctual estimation of
$\Delta HbO_{2}$
via a photoplethysmography (PPG) signal was included. The regression results demonstrated determination coefficients reaching up to 0.993 and normalized errors as low as 0.042. The nonlinear unmixing stage utilized only three end-members to generate precise
$\Delta HbO_{2}$
maps. This not only reduces the information needed for estimation but also minimizes spectral decomposition and processing time. Consequently, the results suggest the potential for real-time application in future research.