Abstract:
Simulation-based Machine Learning (ML) algorithms have proven successful in the realm of microscopy and enabled significant speedups in Single Molecule Localization Microscopy (SMLM) compared to conventional algorithms. These routines, however, did not work for all SMLM modalities and were difficult to use for individuals without dedicated hardware and computational experience. This thesis addresses these challenges.
SMLM, an inverse problem, is a strong candidate for simulation-based ML. In SMLM, fluorophores of one or more kinds are stochastically activated, resulting in sparse events (emitters) imaged over tens to hundreds of thousands of frames. These emitters are localized and rendered to compute the final superresolution image. SMLM is inherently slow due to its need for sparse emitters. Recent ML approaches tackled this problem by enabling high-density imaging. However, the field lacks a dedicated algorithm for high-density, multi-channeled SMLM applications.
ML algorithms in microscopy are applied to various tasks including denoising images, improving their resolution, segmenting biological sites, or detecting biological events. Yet, these methods often rely on dedicated hard- and software setups and have not found their way into daily scientific routines for that reason. The field lacks approaches for repetitive large-scale workflows without complicated manual intervention.
We present DECODE-Plex, a new framework for high-density multi-channeled localization in SMLM, addressing multi-color and biplane imaging. DECODE-Plex is trained on-the-fly by a simulated training procedure. We demonstrate DECODE-Plex’s performance across various densities for simulated and experimental data.
Furthermore, we present DECODE-OpenCloud, a cloud-backed solution for ML algorithms in microscopy. DECODE-OpenCloud abstracts away hardware and maintenance concerns providing researchers and developers with a user-friendly yet production-ready API. It encourages the integration of unused local computing power, benefiting from centralized maintenance and computational power. DECODE-OpenCloud’s design allows for integration without significant additional effort, and we present three algorithms for SMLM as reference implementations: Localization with DECODE, DECODE-Plex and drift correction with COMET.