Unsupervised deep learning for removing structured noise in microscopy and flow cytometry

Salmon, Benjamin Raymond ORCID: 0000-0002-5919-0158 (2025). Unsupervised deep learning for removing structured noise in microscopy and flow cytometry. University of Birmingham. Ph.D.

Preview

Salmon2025PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (18MB) | Preview

Abstract

Microscopy and flow cytometry are pillars of life science research. They use agents such as light or electrons to analyse the structure and composition of biological samples at micrometre and nanometre scales. However, the samples can be damaged by this very process, resulting in
observations that do not capture them in their most natural state. Illumination intensity must therefore be minimised, but this reduces the ratio of signal to noise. When signal intensity is limited by practical constraints, and noise cannot be mitigated at the time of observation, denoisers must be employed to estimate the signal underlying a noisy observation post hoc. Currently, the most accurate estimates are made by deep learning-based denoisers. Their strength comes from utilising the information contained in training data, but this is also one of their greatest limitations.

To reliably remove all forms of noise, existing deep learning-based denoisers require paired training data, typically consisting of noisy observations and corresponding clean signals. In the life sciences, paired data and clean signals may be unobtainable. Techniques exist to train denoisers with unpaired noisy observations, i.e., the very data that is to be denoised, but these face another challenge: structured noise. Structured noise is prevalent in both microscopy and flow cytometry, and it is defined as noise that is correlated over pixels or time points. Currently, no deep learning-based denoiser can reliably remove it without either paired training data or by making sacrifices in the quality of the output.

In this thesis, we develop unsupervised deep learning-based denoisers for structured noise as it commonly occurs in microscopy and flow cytometry. These methods are trained without paired observations or clean signals, and have quality approaching and sometimes exceeding that
of denoisers trained with paired data. We also identify another limitation of unsupervised deep learning-based denoisers – their slow inference time – and present a method to reduce their inference time by three orders of magnitude. We believe that the methods presented here will
remove some of the biggest barriers to applying deep learning denoisers to life science data.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):