Comparison of experimental vs simulated data to train neural networks for speckle imaging data analysis

Multiple Laser Speckle contrast Imaging (MESI) is an imaging method that provides relative blood flow maps from the statistical analysis of the dynamic speckle patterns observed when a coherent source is used to illuminate a tissue that contains moving scatterers. The gold standard analysis of MESI data is done by pixelwise regression of the experimental images to a theoretical function of the contrast K as a function of the exposure time T and decorrelation time τc. This approach is computer intensive, and the duration required to obtain a single flow map is too long for "real-time" analysis of in vivo hemodynamics. In addition, the mathematical model used relies on assumptions that oversimplify the local flow within the object of study. We have evaluated as an alternative a method based on Convolutional Neural Networks (CNN) to directly infer blood flow maps from MESI data, bypassing the model based fitting procedure. The CNN approach is model-free and delivers blood flow maps several orders of magnitude faster than the classical pixelwise non-linear regression. Here, we have evaluated two different datasets of annotated speckle contrast images to train the neural networks. One is composed of simulated time integrated speckle while the other one is composed of experimental data acquired for microfluidic channels with controlled geometries and flows. The study aims at discussing the assets and limits of both approaches.