High-speed computation of high-quality light fluence distribution from low-photon Monte Carlo using a Fourier neural network.

Abstract

Monte Carlo (MC) simulation is the gold standard for studying light propagation in biological tissues within the field of tissue optics. However, the high computational costs of MC simulation limit its broader use in practice. Although GPU-accelerated MC methods significantly enhance computational efficiency, the time required remains substantial under conditions of high-photon numbers and fine mesh due to limited computational resources. Here, we developed a high-efficiency MC method based on the Fourier neural network (MC-Fourier) for high-speed computation of high-quality light fluence (LF) distributions from low-photon MC simulation results, enabling fast yet robust acquisition of two- and three-dimensional LF distributions on fine meshes (up to 1024²). Tests conducted under different mesh configurations demonstrate that the MC-Fourier achieves comparable LF distribution quality to high-photon MC-GPU while reducing the number of photons used by three orders of magnitude. This strategy significantly reduces computation time without sacrificing fidelity, paving the way for fast biomedical optical imaging based on MC simulation.