Understanding the functional architecture of the brain has long been a challenge in neuroscience with a variety of techniques having been developed to explore this structure-function relationship. However, in order to be able to accurately identify the underlying system we require techniques that have the capabilities of describing the complexities therein. In order to perform lesion-function studies a cohort of brain scans with the location of the lesion identified must be collected. Utilising diffusion weighted magnetic resonance imaging, normally collected in the clinical setting, I propose a new unsupervised lesion segmentation routine. The cohort of brain scans also need to be spatially normalised such that homologous regions of the brain are brought into register with each other. However, this process can be perturbed by the presence of a lesion within the scan. Though a series of simulations I evaluate the performance of 12 different spatial normalisation routines on brains scans that possess a lesion. Historically lesion-function mapping studies have tended to use a univariate statistical approach, where different locations within the brain are treated as being spatially independent from each other. Here I show that biases within the structure of the data have the potential to distort the lesion-function inferences we draw. Though a series of simulations, I show that a mass univariate technique is vulnerable to these biases and assess three different multivariate methods (Support Vector Machines, Relevance Vector Machines and Flexible Bayesian Modelling) as potential solutions to this problem. Asides from making lesion-function inferences, these multivariate models can be used to predict future events. Using a data set of paired admission diffusion weighted magnetic resonance imaging scans and functional outcome scores I apply these techniques to the clinical scenario of predicting the functional outcome of patients after a cerebral vascular event.