GPU inverse Adding-Doubling method for analysis of optical spectral images

The Adding-Doubling method (AD) is a general analytical solution of the radiative transfer equation (RTE). AD is significantly faster than the accurate solutions of RTE (e.g. Monte Carlo), and accurate in contrast to the approximate solutions of RTE (e.g., Diffusion Approximation).

Since the existing AD algorithms are implemented on CPU, AD was implemented on GPU to speed-up the method. GPU AD was combined with GPU Levenberg-Marquardt algorithm (LM) to extract imaged object parameters from the spectral data. The performance of the developed GPU AD-LM algorithm was evaluated on simulated and experimental optical spectral images of different SiliGlass tissue phantoms and human skin under various hemodynamic conditions. A good agreement between the determined and the true tissue properties was obtained. The speed-up compared to the GPU Monte Carlo method was between 2,160-170,400. Optimal number of threads was approx. 10,000.

The new GPU AD-LM algorithm results in significantly faster and accurate determination of tissue properties compared to other solutions of RTE. Sole GPU AD can be also used to speed up the machine learning methods for analysis of spectral images.