Bridging generative models and Convolutional Neural Networks for domain-agnostic segmentation of brain MRI

Advanced search
Browse by:

Department | Year

UCL Theses | Latest

Deposit your research

Bridging generative models and Convolutional Neural Networks for domain-agnostic segmentation of brain MRI

Billot, Benjamin Pierre Maurice; (2022) Bridging generative models and Convolutional Neural Networks for domain-agnostic segmentation of brain MRI. Doctoral thesis (Ph.D), UCL (University College London). Green open access

Preview

Text
PhD_thesis.pdf - Accepted Version
Download (40MB) | Preview

Abstract

Segmentation of brain MRI scans is paramount in neuroimaging, as it is a prerequisite for many subsequent analyses. Although manual segmentation is considered the gold standard, it suffers from severe reproducibility issues, and is extremely tedious, which limits its application to large datasets. Therefore, there is a clear need for automated tools that enable fast and accurate segmentation of brain MRI scans. Recent methods rely on convolutional neural networks (CNNs). While CNNs obtain accurate results on their training domain, they are highly sensitive to changes in resolution and MRI contrast. Although data augmentation and domain adaptation techniques can increase the generalisability of CNNs, these methods still need to be retrained for every new domain, which requires costly labelling of images. Here, we present a learning strategy to make CNNs agnostic to MRI contrast, resolution, and numerous artefacts. Specifically, we train a network with synthetic data sampled from a generative model conditioned on segmentations. Crucially, we adopt a domain randomisation approach where all generation parameters are drawn for each example from uniform priors. As a result, the network is forced to learn domain-agnostic features, and can segment real test scans without retraining. The proposed method almost achieves the accuracy of supervised CNNs on their training domain, and substantially outperforms state-of-the-art domain adaptation methods. Finally, based on this learning strategy, we present a segmentation suite for robust analysis of heterogeneous clinical scans. Overall, our approach unlocks the development of morphometry on millions of clinical scans, which ultimately has the potential to improve the diagnosis and characterisation of neurological disorders.

Type:	Thesis (Doctoral)
Qualification:	Ph.D
Title:	Bridging generative models and Convolutional Neural Networks for domain-agnostic segmentation of brain MRI
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/10156888