Gilbert Hausmann Award - 2022 - Joowon Lim
Learning approaches to high-fidelity optical diffraction tomography
EPFL thesis n°7954
Thesis director: Prof. D. Psaltis
For his work on optical tomography. His research led to new methods and algorithms for imaging three dimensional objects including cells and tissues.
Optical diffraction tomography (ODT) is an imaging technique that provides us with 3D refractive index (RI) distributions of transparent samples. Since RI values differ across materials, they serve as endogenous contrasts without labeling. The fundamental principle of ODT reconstruction is to recover the 3D information from multiple 2D measurements. However, there exist fundamental challenges in ODT reconstruction. The first challenge is that there are inaccessible measurements due to the limited numerical apertures of the optical system. The second challenge is to model the nonlinear relationship between a sample and the measurements. In this thesis, we aim to address these challenges to improve the accuracy of the reconstructed RI distributions. The first approach uses iterative reconstruction schemes based on the modeling of optical propagation, and we call it learning tomography because the light is propagated through a multilayer structure of transmission matrices. The second approach is based on statistical learning of the artifacts present in the final reconstructions and uses a deep neural network trained on a large dataset. In this contribution, we propose solutions to the main challenges of ODT reconstruction and achieve accurate and faithful reconstructions, which confirms the power of ODT as an imaging technique and allows the extension of its application domain.