Probabilistic Parametric Atlases
Characterising variations in different groups is a powerful way of finding differences between normal and abnormal connectivity and structure. A basic atlas captures only the mean of a group, while a more sophisticated approach incorporates mean and covariance information for different groups. This project will develop probabilistic parameterized atlases, for normal and individual disease groups, accounting for biometric and clinical factors. Developing such atlases requires large datasets (e.g. HCP, Biobank), with biometric data, and the estimation of very large sets of parameters, especially covariances between multi-modal quantities. Bayesian techniques will be used to provide robust and sparse estimation of key data-driven parameters, ensuring that the derived atlases are robust across scanner variations and to account for missing data (e.g., demographics) by using appropriate priors (e.g., on age). When used with longitudinal data, these atlases will capture the timing/order information, which is hugely beneficial in visualising the progression of a disease across multiple modalities. The ultimate aim is to incorporate these into a clinical tool that will allow a single patient’s multimodal imaging dataset to be evaluated with respect to the different population distributions.