In optical diffraction tomography (ODT), the 3D shape of an object is retrieved from multiple holographic recordings of its 2D projections. Machine learning can be used to augment the performance of ODT by correcting distortions in the reconstruction due to the limited number of projections. We will describe approaches for accomplishing this. The first relies on digital phantoms that are used to generate examples and serve as the ground truth for training the neural network. A second approach relies on the constraints imposed by the physics of the problem to train the neural network that produces corrected 3D images.
Copyright. All Rights Reserved.