Analysis of optical imaging data (abstract)

Porrill, J, Stone, J V, Berwick, J, Mayhew, J and Coffey, P (2000)

This chapter will present techniques for the recovery of component spatial and temporal modes from spatio-temporal data sets, in particular from medical imaging data such as that obtained by functional magnetic resonance imaging \index{functional magnetic resonance imaging} (fMRI) and optical imaging \index{optical imaging} (OI) of brain activity. These techniques were developed in order to help address some current issues involving the nature of the haemodynamic response to neural activity (for more details of this research and further references see \cite{mayhew}).

Current analysis techniques can be divided, roughly speaking, into two classes: data-driven and model-driven. An example of a data driven technique is principal component analysis (PCA). This requires no prior knowledge about the form of the components but physical interpretation of the decomposition is difficult and there is no obvious way to test for statistical significance. In contrast model-driven techniques such as general linear model analysis require that rather constrained models of the expected behaviour be prescribed in advance but have the very significant advantage that they allow the use of standard statistical tests.

In our application it is important that the components we recover correspond to underlying physical processes. Since current biophysical knowledge of these processes is limited, the decomposition must be based, as far as is possible, on generic principles rather than on highly constrained models. We therefore define an intermediate class of \emph{weak temporal models}, designed to utilise incomplete prior information about temporal responses (available, for example, from experimental protocols) without having to specify the expected behaviour in detail.

We will show how to define a \emph{weak causal model} for the response to a single stimulus and a \emph{weak periodic model} for the response to a periodic (eg box-car) stimulus, and discuss how the statistical significance of temporal components recovered using such models might be assessed.

The weak models we define here prescribe an optimal temporal response but leave the problem of recovering the associated spatial activity map under-constrained. We show that this ill-posed problem can be regularised effectively using the entropy measure underlying spatial independent component analysis \index{independent component analysis}. We will argue that this regulariser can be regarded as implementing a spatial prior based on knowledge of the spatial statistics of various classes of activity maps.