Parallel inverse-problem solver for time-domain optical tomography with perfect parallel scaling

TL;DR

This paper introduces a highly efficient parallel inverse-problem solver for time-domain optical tomography using radiative transfer, achieving perfect scaling and significant acceleration through innovative computational strategies.

Contribution

It presents a novel parallel-decomposition strategy and a Multiple Staggered Source method that drastically reduce computational costs and enable large-scale 2D inverse problem solutions.

Findings

Six-fold acceleration with MSS method

Reduced computation time from months to hours

Successful large-scale 2D inverse problem solutions

Abstract

This paper presents an efficient parallel radiative transfer-based inverse-problem solver for time-domain optical tomography. The radiative transfer equation provides a physically accurate model for the transport of photons in biological tissue, but the high computational cost associated with its solution has hindered its use in time-domain optical-tomography and other areas. In this paper this problem is tackled by means of a number of computational and modeling innovations, including 1) A spatial parallel-decomposition strategy with perfect parallel scaling for the forward and inverse problems of optical tomography on parallel computer systems; and, 2) A Multiple Staggered Source method (MSS) that solves the inverse transport problem at a computational cost that is independent of the number of sources employed, and which significantly accelerates the reconstruction of the optical parameters: a six-fold MSS acceleration factor is demonstrated in this paper. Finally, this contribution presents 3) An intuitive derivation of the adjoint-based formulation for evaluation of functional gradients, including the highly-relevant general Fresnel boundary conditions -- thus, in particular, generalizing results previously available for vacuum boundary conditions. Solutions of large and realistic 2D inverse problems are presented in this paper, which were produced on a 256-core computer system. The combined parallel/MSS acceleration approach reduced the required computing times by several orders of magnitude, from months to a few hours.