Computational Neuroscience with Deep Learning for Brain Imaging Analysis and Behaviour Classification

Abstract

Recent advances of artificial neural networks and deep learning model have produced significant results in problems related to neuroscience. For example, deep learning models have demonstrated superior performance in non-linear, multivariate pattern classification problems such as Alzheimer’s disease classification, brain lesion segmentation, skull stripping and brain age prediction. Deep learning provides unique advantages for high-dimensional data such as MRI data, since it does not require extensive feature engineering. The thesis investigates three problems related to neuroscience and discuss solutions to those scenarios. MRI has been used to analyse the structure of the brain and its pathology. However, for ex- ample, due to the heterogeneity of these scanners, MRI protocol, variation in site thermal and power stability can introduce scanning differences and artefacts for the same individual under- going different scans. Therefore combining images from different sites or even different days can introduce biases that obscure the signal of interest or can produce results that could be driven by these differences. An algorithm, the CycleGAN, will be presented and analysed which uses generative adversarial networks to transform a set of images from a given MRI site into images with characteristics of a different MRI site. Secondly, the MRI scans of the brain can come in the form of different modalities such as T1- weighted and FLAIR which have been used to investigate a wide range of neurological disorders. The acquisition of all of these modalities are expensive, time-consuming, inconvenient and the required modalities are often not available. As a result, these datasets contain large amounts of unpaired data, where examples in the dataset do not contain all modalities. On the other hand, there is a smaller fraction of examples that contain all modalities (paired data). This thesis presents a method to address the issue of translating between two neuroimaging modalities with a dataset of unpaired and paired, in semi-supervised learning framework. Lastly, behavioural modelling will be considered, where it is associated with an impressive range of decision-making tasks that are designed to index sub-components of psychological and neural computations that are distinct across groups of people, including people with an underlying disease. The thesis proposes a method that learns prototypical behaviours of each population in the form of readily interpretable, subsequences of choices, and classifies subjects by finding signatures of these prototypes in their behaviour.