Depth estimation in turbid media from stack of epi-illuminated microscopy images, using deep learning

Abstract

Optical imaging is the simplest modality for any imaging applications. Also, microscopy has the best resolution among all the imaging modalities used for clinical diagnosis and biomedical research. However, detection of inclusions by optical imaging in a biological tissue might be challenging because of the scattering and absorption inside the tissue. Due to the scattering, an optical image of an object inside a turbid medium yields a scattering-induced blur, which increases with the depth of the object from the surface of the medium. This correlation has been shown in literature to be used for depth estimation from single trans-illumination images. In a similar way, the scattering-blur also changes while varying the focal plane of the imaging system. This gradual change in the contrast can be utilised to determine the actual position of the object from a stack of multiple focus-shifted images. Here, we present a proof of concept of a deep-learning based method for determining the location of objects inside turbid media, from a stack of blurred images. Since trans-illumination is not applicable for large body-parts, epi-illumination setup was used for imaging. For this preliminary study, a U-Net neural network regression model was trained and tested under simplified conditions. It takes 3D image data as input and gives a 2D depth matrix as output. Black PU structures of simple geometrical shapes were used as absorbing objects. An intralipid solution of 0.7% concentration was used as the scattering substrate. The black absorbers, immersed in the substrate, were imaged with microscope by varying the focal plane to obtain the image-stacks. The actual depth profile of the absorbers was measured with a 3D profilometer, which was used as corresponding ground truth for training. The predicted results from the trained model show good agreement with the ground truth for testing data.