MIBE Seminar: Enabling Light Field 3D Microscopy and Domain Shift Detection with Normalizing Flows

Titel: Enabling Light Field 3D Microscopy and Domain Shift Detection with Normalizing Flows.

Sprecher: Josue Page Vizcaino, Doctoral Candidate (Lasser)

Datum und Uhrzeit: Dienstag, 03. Mai 2022, 13:00 Uhr

Abstract: Light field microscopy is a scanless imaging technique capable of acquiring angular and spatial information of fluorescent samples at camera frame rate. As a post-processing step, 3D volumes can be reconstructed from the acquired 2D images, by means of a reconstruction method (e.g. deconvolution). However, the available methods produce reconstructions with artifacts, anisotropic resolution which take long times to process.
In recent years, the usage of neural networks for 3D reconstruction has alleviated these setbacks. Unfortunately, most methods still overlook the origin of the input data, fully ignoring the in/out of distribution and domain shift issues, which, can lead to inaccurate 3D reconstructions, which is undesired and even dangerous within the biomedical realm. In my research, I investigate how Normalizing Flows enable real-time 3D reconstructions and allow for exact likelihood computation, thereby domain shift detection, and, in turn, how much re-training is necessary to perform well when working with images from a novel domain.