"Unsupervised Inverse approach for the 3D reconstruction in Tomographic Diffractive Microscopy" by Laurence Denneulin

Seminar by Laurence Denneulin

Abstract

Tomographic Diffractive Microscopy is a high resolved imaging technique, enabling the reconstruction of the 3D refraction index of a sample without any labeling. A tomographic dataset is composed of several holographic acquisitions of the sample obtained for different orientations of the illumination wave. Using a numerical model of the hologram formation, as a function of the sample, the 3D reconstruction is achieved using an inversion method (direct or iterative). In the context of iterative inverse approach, the reconstruction is estimated as the minimum of a cost function composed of a data-fidelity term and a regularization term. The data-fidelity term is based on the noise statistics of the data and corresponds to its co-log-likelihood. The regularization term is based on prior knowledge on the object (sparsity, smoothness, sharp edges). The contribution of the regularization is weighted with some hyperparameters which need to be perfectly adjusted to ensure a good trade-off between data-fidelity and prior. In this presentation, we present an unsupervised inverse approach to reconstruct the sample using two different hologram formation models. In this method, the hyperparameter is estimated by minimizing the generalized Stein’s unbiased risk estimator (GSURE). We show, on simulated and experimental data, that this method is able to find a the appropriate trade-off between the data-fidelity and the prior. Moreover, in the context of limited data, it outperforms traditional methods.

This seminar will be held in English.