Fast 3D image reconstruction in optical tomography using a coarse-gain parallelization strategy

Abstract

Absolute dual-parameter image reconstruction in optical tomography (OT) is a nonlinear and ill-posed problem, requiring a model-based iterative approach and an accurate and fully three-dimensional light transport forward model. These factors make OT a computationally expensive problem, resulting in reconstruction times that range from several minutes to hours, which is not acceptable in most clinical applications. In order to reduce reconstruction times we propose a coarse-grain parallel implementation of the inverse algorithm using a shared-memory threaded approach. Unlike most conventional parallelization strategies which operate on the level of the linear algebra subsystem our approach uses a parallelization at the application level, thereby leaving the underlying linear matrix solution routines untouched. This allows to exploit the inherently parallel structure of the problem, while at the same time utilizing fast direct or efficiently preconditioned iterative serial linear solvers, which in most cases can not be efficiently parallelized. We compare the performance of the coarse-grain parallel approach with a low-level parallelization of the linear conjugate gradient solver and show that the proposed method achieves significant performance improvements, thereby bringing reconstruction times within a clinically acceptable range.