Deep learning for image processing in optical super-resolution microscopy

Optical microscopy is fundamentally governed by a trade-off between image quality, imaging speed and duration. The quality can be considered a function of the signal-to-noise ratio, contrast and image resolution, which are all limited by the amount of light that can be acquired within a set exposure time. Many applications in live-cell imaging have specific requirements for illumination power and exposure time, thus necessitating a compromise with quality. In recent years, this fundamental limitation in optical microscopy has been shifted with the aid of deep learning methods. In this thesis, I propose methods that improve robustness to noise in image processing while making greater use of the available signal in the data. Applications include denoising for improved electron tomography when using cryogenic electron microscopy; image segmentation facilitating quantitative analysis of dynamics in endoplasmic reticulum (ERnet); and versatile reconstruction of super-resolved images from raw data acquired with structured illumination microscopy (ML-SIM). The deep learning methods that are presented are compared to classical image processing alternatives and tested on real experimental data acquired by collaborators in different departments of the university. The overall finding of the thesis is that deep learning techniques offer a highly effective approach to many problems in bioimaging. With ERnet, it is possible to obtain a segmentation method that is reliable, fast and functional across different experiments without the need for retraining guided by further manual annotations. As for ML-SIM, I show that the reconstruction of structured illumination microscopy data can be treated as the inverse problem of a forward modelling process. This relies on an approximative image formation model that takes uncertainties and noise into account. By training a deep neural network to invert the forward modelled SIM data, a highly generalised reconstruction model can be obtained, which can handle SIM data from multiple microscopes while providing a high reconstruction quality. The thesis is concluded with a reflective section on where the field is headed and which future applications may be enabled by the advancement of deep learning techniques.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2022

Author dc:creator

• Christensen, Charles N

Advisors dc:contributor.advisor

• Kaminski, Clemens
• Lio, Pietro

• Artificial intelligence
• Computer vision
• Deep learning
• Image processing
• Image segmentation
• Machine learning
• Optical microscopy
• Structured illumination microscopy
• Super-resolution microscopy

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/61cfdb48-221b-4e79-8105-662fb024f6a3/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.101919

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/358096