Jenkinson, Luke Ryan;
(2024)
Big Data, Small Labels: Contrastive Learning for Medical Image Analysis.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
PhDThesisFinal.pdf - Accepted Version Download (6MB) | Preview |
Abstract
Deep neural networks have become the de facto standard for many computer vision tasks. Despite this, the uptake of these state-of-the-art methods to medical imaging tasks has been lacking. One possible reason for this is the scarcity of large, labelled medical image datasets for these models to train on. To combat this, numerous semi-supervised methods have been suggested for improving performance when faced with limited labelled training data. One such class of methods is contrastive learning: these methods aim to learn powerful features from unlabelled data, which can then be used, along with a small amount of labelled training data to produce higher performance than could be achieved with the labelled data alone. This thesis examines two distinct contrastive methods, Contrastive Predictive Coding and SimCLR, across multiple dimensions. In the first part of this thesis, the ability of contrastive methods to increase performance on medical imaging tasks is validated, exploring how the size of the labelled dataset changes the performance of the downstream task compared with a powerful baseline. Additional work is undertaken to understand how this improvement in accuracy may affect the robustness of the model. In the second section of this thesis, design choices of the contrastive training protocol are examined to understand how to achieve the greatest performance. Some contrastive methods, most notably SimCLR, make heavy use of augmentation in their training protocol, and the impact of this has been under studied. This thesis examines the impact of both the type and the magnitude of these augmentations. Finally, a large study of the impact of unlabelled dataset on the downstream classification performance is presented, giving novel recommendations for how to improve performance on a wide variety of tasks.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Big Data, Small Labels: Contrastive Learning for Medical Image Analysis |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) Licence (https://creativecommons.org/licenses/by-nc-nd/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 UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science |
| URI: | https://discovery-pp.ucl.ac.uk/id/eprint/10198112 |
Archive Staff Only
 |
View Item |
