Adaptive holographic image reconstruction based on physics integrated machine learning

Abstract

Holographic image reconstruction is an ill-posed inverse problem and there is a trade-off between the complexity of the imaging system and the quality of the reconstructed image, so there is no rule of thumb to use. Recently, the image-to-image translation task in the computer vision field has been applied to holographic image reconstruction and shown unprecedented performance. However, they utilized paired information between complex amplitude and hologram intensity and prior information such as sample-to-sensor distance. In this thesis, we demonstrate a novel approach for holographic image reconstruction by embedding a distance parameterized forward imaging model into the cycle-consistency generative adversarial network(cycleGAN). We showed that the proposed network reliably reconstructs the complex field of polystyrene beads even in the presence of strong perturbations on the imaging system. Also, by utilizing the property of an unsupervised learning approach, relaxation of data constraints is proved using tissue samples that have complex structures. Lastly, application for the proposed model is presented using a dynamic imaging sample, red blood cell.