Deep generative models for medical image synthesis and strategies to utilise them

Medical imaging has revolutionised the diagnosis and treatments of diseases since the first medical image was taken using X-rays in 1895. As medical imaging became an essential tool in a modern healthcare system, more medical imaging techniques have been invented, such as Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET), Computed Tomography (CT), Ultrasound, etc. With the advance of medical imaging techniques, the demand for processing and analysing these complex medical images is increasing rapidly. Efforts have been put on developing approaches that can automatically analyse medical images. With the recent success of deep learning (DL) in computer vision, researchers have applied and proposed many DL-based methods in the field of medical image analysis. However, one problem with data-driven DL-based methods is the lack of data. Unlike natural images, medical images are more expensive to acquire and label. One way to alleviate the lack of medical data is medical image synthesis. In this thesis, I first start with pseudo healthy synthesis, which is to create a ‘healthy’ looking medical image from a pathological one. The synthesised pseudo healthy images can be used for the detection of pathology, segmentation, etc. Several challenges exist with this task. The first challenge is the lack of ground-truth data, as a subject cannot be healthy and diseased at the same time. The second challenge is how to evaluate the generated images. In this thesis, I propose a deep learning method to learn to generate pseudo healthy images with adversarial and cycle consistency losses to overcome the lack of ground-truth data. I also propose several metrics to evaluate the quality of synthetic ‘healthy’ images. Pseudo healthy synthesis can be viewed as transforming images between discrete domains, e.g. from pathological domain to healthy domain. However, there are some changes in medical data that are continuous, e.g. brain ageing progression. Brain changes as age increases. With the ageing global population, research on brain ageing has attracted increasing attention. In this thesis, I propose a deep learning method that can simulate such brain ageing progression. Specifically, longitudinal brain data are not easy to acquire; if some exist, they only cover several years. Thus, the proposed method focuses on learning subject-specific brain ageing progression without training on longitudinal data. As there are other factors, such as neurodegenerative diseases, that can affect brain ageing, the proposed model also considers health status, i.e. the existence of Alzheimer’s Disease (AD). Furthermore, to evaluate the quality of synthetic aged images, I define several metrics and conducted a series of experiments. Suppose we have a pre-trained deep generative model and a downstream tasks model, say a classifier. One question is how to make the best of the generative model to improve the performance of the classifier. In this thesis, I propose a simple procedure that can discover the ‘weakness’ of the classifier and guide the generator to synthesise counterfactuals (synthetic data) that are hard for the classifier. The proposed procedure constructs an adversarial game between generative factors of the generator and the classifier. We demonstrate the effectiveness of this proposed procedure through a series of experiments. Furthermore, we consider the application of generative models in a continual learning context and investigate the usefulness of them to alleviate spurious correlation. This thesis creates new avenues for further research in the area of medical image synthesis and how to utilise the medical generative models, which we believe could be important for future studies in medical image analysis with deep learning