Machine-learning-enhanced image reconstruction in optical tomography using the Monte Carlo method for light transport

Significance: The Monte Carlo method for light transport is widely accepted as an accurate method for simulating light propagation in a scattering medium. Its use in optical tomography, however, suffers from inherent stochastic noise. This noise is present in both evaluations of the forward model, as well as in the search direction of the minimization algorithm used for image reconstruction. Aim: We aim to utilize machine learning to compensate for the stochastic Monte Carlo noise in the reconstruction of absorption and scattering in optical tomography. Approach: An iterative image reconstruction algorithm is proposed. The algorithm uses convolutional neural networks in a stochastic Gauss-Newton update when estimating absorption and scattering coefficients. Results: The methodology is evaluated using numerical simulations and compared against the conventional stochastic Gauss-Newton algorithm in optical tomography. It is demonstrated that the methodology can be used to compensate for image reconstruction artifacts caused by the stochastic noise. Conclusions: The proposed machine learning approach can be used to compensate for stochastic noise in Gauss-Newton iterations, and it enables reconstruction of absorption and scattering with a significantly lower number of photons than a conventional stochastic Gauss-Newton algorithm.