Shaw, Richard;
(2022)
MRI Artefact Augmentation: Robust Deep Learning Systems and Automated Quality Control.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
PhD_thesis_final.pdf - Accepted Version Download (68MB) | Preview |
Abstract
Quality control (QC) of magnetic resonance imaging (MRI) is essential to establish whether a scan or dataset meets a required set of standards. In MRI, many potential artefacts must be identified so that problematic images can either be excluded or accounted for in further image processing or analysis. To date, the gold standard for the identification of these issues is visual inspection by experts. A primary source of MRI artefacts is caused by patient movement, which can affect clinical diagnosis and impact the accuracy of Deep Learning systems. In this thesis, I present a method to simulate motion artefacts from artefact-free images to augment convolutional neural networks (CNNs), increasing training appearance variability and robustness to motion artefacts. I show that models trained with artefact augmentation generalise better and are more robust to real-world artefacts, with negligible cost to performance on clean data. I argue that it is often better to optimise frameworks end-to-end with artefact augmentation rather than learning to retrospectively remove artefacts, thus enforcing robustness to artefacts at the feature level representation of the data. The labour-intensive and subjective nature of QC has increased interest in automated methods. To address this, I approach MRI quality estimation as the uncertainty in performing a downstream task, using probabilistic CNNs to predict segmentation uncertainty as a function of the input data. Extending this framework, I introduce a novel decoupled uncertainty model, enabling separate uncertainty predictions for different types of image degradation. Training with an extended k-space artefact augmentation pipeline, the model provides informative measures of uncertainty on problematic real-world scans classified by QC raters and enables sources of segmentation uncertainty to be identified. Suitable quality for algorithmic processing may differ from an image's perceptual quality. Exploring this, I pose MRI visual quality assessment as an image restoration task. Using Bayesian CNNs to recover clean images from noisy data, I show that the uncertainty indicates the possible recoverability of an image. A multi-task network combining uncertainty-aware artefact recovery with tissue segmentation highlights the distinction between visual and algorithmic quality, which has the impact that, depending on the downstream task, less data should be discarded for purely visual quality reasons.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | MRI Artefact Augmentation: Robust Deep Learning Systems and Automated Quality Control |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2022. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Med Phys and Biomedical Eng UCL > Provost and Vice Provost Offices > UCL BEAMS UCL |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10147496 |
Loading...
Loading...Loading...Loading...Archive Staff Only
 |
View Item |
