Use this URL to cite or link to this record in EThOS:
Title: Machine learning in medical image analysis
Author: Chen, Liang
ISNI:       0000 0004 7659 0668
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Machine learning is playing a pivotal role in medical image analysis. Many algorithms based on machine learning have been applied in medical imaging to solve classification, detection, and segmentation problems. Particularly, with the wide application of deep learning approaches, the performance of medical image analysis has been significantly improved. In this thesis, we investigate machine learning methods for two key challenges in medical image analysis: The first one is segmentation of medical images. The second one is learning with weak supervision in the context of medical imaging. The first main contribution of the thesis is a series of novel approaches for image segmentation. First, we propose a framework based on multi-scale image patches and random forests to segment small vessel disease (SVD) lesions on computed tomography (CT) images. This framework is validated in terms of spatial similarity, estimated lesion volumes, visual score ratings and was compared with human experts. The results showed that the proposed framework performs as well as human experts. Second, we propose a generic convolutional neural network (CNN) architecture called the DRINet for medical image segmentation. The DRINet approach is robust in three different types of segmentation tasks, which are multi-class cerebrospinal fluid (CSF) segmentation on brain CT images, multi-organ segmentation on abdomen CT images, and multi-class tumour segmentation on brain magnetic resonance (MR) images. Finally, we propose a CNN-based framework to segment acute ischemic lesions on diffusion weighted (DW)-MR images, where the lesions are highly variable in terms of position, shape, and size. Promising results were achieved on a large clinical dataset. The second main contribution of the thesis is two novel strategies for learning with weak supervision. First, we propose a novel strategy called context restoration to make use of the images without annotations. The context restoration strategy is a proxy learning process based on the CNN, which extracts semantic features from images without using annotations. It was validated on classification, localization, and segmentation problems and was superior to existing strategies. Second, we propose a patch-based framework using multi-instance learning to distinguish normal and abnormal SVD on CT images, where there are only coarse-grained labels available. Our framework was observed to work better than classic methods and clinical practice.
Supervisor: Rueckert, Daniel ; Bentley, Paul Sponsor: National Institute for Health Research
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral