Novel deep learning approaches for imaging, localisation and data post-processing in single molecule localisation microscopy
File(s)
Author(s)
Boland, Miguel
Type
Thesis
Abstract
This thesis contributes to the field of Single Molecule Localisation Microscopy (SMLM) through three interconnected projects. Firstly, a novel method for 3D localisation is developed utilizing intrinsic optical aberrations, eliminating the need for specialized optical components while providing fast and computationally efficient z-localisation. We demonstrate the method is capable of reconstructing nuclear pores complexes with a performance comparable to alternative methodologies, but at a greatly reduced computational cost. Building on this, an improved autofocus system is developed which leverages the aforementioned 3D localisation tool, potentially surpassing existing solutions in accuracy and reliability. Lastly, the thesis explores an unsupervised clustering algorithm using Graph Neural Networks (GNN) for SMLM data clustering. Existing algorithms are shown to outperform the GNN in speed and ease of tuning within the targeted domain, and the work concludes by suggesting future research directions to adapt the GNN if a more complex domain could be found, thereby contributing to the ongoing advancement of SMLM techniques and their applications in biological imaging.
Version
Open Access
Date Issued
2024-10-22
Date Awarded
01/02/2025
Advisor
Cohen, Edward
Sponsor
Wellcome Trust (London, England)
Publisher Department
Mathematics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)