Optimising convolutional neural networks for large-scale neuroimaging studies

Ageing has a pronounced effect on the human brain, leading to cognitive decline and an increased risk of neurodegenerative diseases. Thus, the ageing population presents a significant challenge for healthcare. The use of MRI and the availability of computational methods for analysing the MRI data is increasingly contributing to the understanding of healthy and diseased structural brain maturation and ageing. Increasingly, large cross sectional and longitudinal neuroimaging studies are becoming available, presenting opportunities for the application of deep learning to neuroimage analysis. There are however, many domain specific problems for applying deep learning to neuroimaging problems, which currently limit their wider applicability. This thesis explores three distinct problems. First, a model is developed to explore brain ageing. Both normal ageing and neurodegenerative disease cause morphological changes to the brain, and deep learning models are well suited to capturing these patterns. A 3D CNN architecture is developed to predict chronological age, using T1- weighted MRI from the UK Biobank dataset. The proposed method shows competitive performance on age prediction, but, most importantly, the CNN prediction errors correlated significantly with many clinical measurements from the UK Biobank in the female and male groups. In addition, having used images from only one imaging modality in this experiment, the relationships between ∆_BrainAge and the image-derived phenotypes (IDPs) from all of the other imaging modalities in the UK Biobank are explored, showing correlations consistent with known patterns of ageing. The effect of the pre-processing is also explored. Specifically, it is shown that the use of non-linearly registered images to train the CNNs can lead to the network being driven by artefacts of the registration process, and therefore to miss subtle indicators of ageing, which would limit the clinical relevance of the model. Increasingly large MRI neuroimaging datasets are becoming available but many of these are highly "multi-site, multi-scanner", which leads to an increase in variance due to nonbiological factors when these are combined. This increase in variance is due to factors such as differences in acquisition protocols and hardware, and can mask the signals of interest, and this is known as the harmonisation problem. A deep learning based scheme, developed from domain adaptation techniques, is used to create harmonised outputs. An iterative update approach is used to create scanner-invariant features, whilst simultaneously maintaining performance on the main task of interest, thus reducing the influence of the acquisition scanner on the network predictions. The framework is demonstrated for regression, classification, and segmentation tasks with two different network architectures. It is shown that not only can the framework harmonise multi-site datasets, but it can also adapt to many data scenarios, including biased datasets and limited training labels. Finally, it is shown that the framework can be extended for the removal of other known confounds in addition to scanner or data source. The overall framework is therefore flexible and should be applicable to a wide range of neuroimaging studies. Finally, the parameterisation of neural networks is considered: the vast number of parameters means that networks need large numbers of training examples and labels. Especially for medical image segmentation, large labelled datasets are rarely available. A method to train and prune UNet architectures simultaneously is developed, alongside an adaptive targeted dropout scheme that makes the network robust to the pruning - that is, the removal of filters from the model. It is shown that the pruned models outperform the standard UNet models, especially when working in very low data regimes, across medical imaging tasks. The framework is then applied to multisite MRI data: the standard UNet models trained on the data from one site suffer significant performance degradation when applied to the data from the other sites, but this is reduced when using the pruned models, due to a reduction in model overfitting. The generalisability is systematically explored, and it is shown that the pruned models have increased robustness, compared to the standard UNet models.