Robustness of diffuse reflectance spectra analysis by the inverse adding-doubling algorithm*

Extracting skin properties from diffuse reflectance spectra requires light transport modeling within the tissue. Typically, tissue model parameters defining optical properties are fixed to speed up the fitting process and improve the performance.

We assessed the effect of fixing model parameters on the accuracy and robustness of a GPU-accelerated two-layer inverse adding-doubling (IAD) algorithm by calculating RMSE and MAE for estimated parameters and fitted spectra. Specifically, multiple model parameters in 27 fitting cases were fixed for noiseless simulated skin spectra. Two selected fitting cases were considered for simulated skin spectra with three noise levels added and for in-vivo measured reflectance spectra from hyperspectral images of human hands. Finally, a method was proposed to pre-estimate model parameters from experimental spectra to improve the performance.

Fixing multiple model parameters for simulated spectra reduced SD of estimated parameters up to 10 times to 3%. Pre-estimating parameters improved the agreement of fitted and measured spectra (cumulative RMSE decreased from 0.003 to 0.001) and reduced the fitting time by 0.08 s per spectrum.

Results imply that fixing multiple parameters improves the fitting performance for simulated and experimental skin spectra.