Extraction of optical properties from scattering media using a convolutional neural network and diverse exit information

Estimation of the optical properties of scattering media, such as tissue, is important in diagnostics as well as in the development of techniques to image deeper. Inverting scattering patterns to recover optical properties is not simple, and machine learning has been proposed to recover these properties. We train a neural network on simulated data to predict scattering (<i>µ_s</i>), reduced scattering (<i>μ</i>_<i>s</i>^'), and absorption (<i>µ_a</i>) coefficients, as well as anisotropy factor (<i>g</i>). Our best network achieves prediction Mean Absolute Relative Error (MARE) of 3.4% for <i>µ_a</i> and 2.1% for <i>μ</i>_<i>s</i>^'. We show that combining photon exit angle and position improves accuracy by ~34% compared to using exit position alone. To capture angle information practically, we propose a novel, to our knowledge, method using intensity distributions measured at two planes above the sample. We also analyze the conditions when <i>µ_s</i> and <i>g</i> can be separated from <i>μ</i>_<i>s</i>^'. In the non-diffuse regime, <i>g</i> and <i>µ_s</i> are determined with MARE <10% in the majority of cases. In the diffusive regime, <i>g</i> and <i>µ_s</i> cannot be determined well. However, when tissue thickness is <2 transport mean free paths (TMFP), the majority of prediction MARE is <25%.