Deep Learning of Diffuse Optical Tomography based on Time-Domain Radiative Transfer Equation

Abstract

Near infrared diffuse optical tomography (DOT) provides an imaging modality for the oxygenation of tissue. In this paper, we propose a novel machine learning algorithm based on time-domain radiative transfer equation. We use temporal profiles of absorption measure for a two-dimensional model of target tissue, which are calculated by solving time-domain radiative transfer equation. Applying a longshort-term memory (LSTM) deep learning method, we find that we can specify positions of cancer cells with high accuracy rates. We demonstrate that the present algorithm can also predict multiple or extended cancer cells.

Tutorial on the Use of Deep Learning in Diffuse Optical Tomography

“…In reflectance geometry, the source and the detector are placed on the same side of the sample, while in transmittance geometry, they are on the opposite side of the sample. The time-domain diffuse optical tomography system has been developed by many academic groups [43,[78][79][80][81][82][83].…”

“…However, despite its many advantages, analytical methods suffer from very high computational cost, and in the context of DOT, from an inability to achieve accurate results for realistic 3D problems. Therefore, it has become customary to employ deep learning tools to solve reconstruction problems [6,9,81,86,87]. In the case of diffusive imaging, this refers to the mapping between the optical properties of the media to the measured timeresolved signal registered by the detectors.…”