Spatio-temporal analysis of auditory cortex activation as detected with silent event related fMRI

Functional magnetic resonance imaging (fMRI) allows neuroscientists to assess brain function by evaluating haemodynamic activity (blood flow) when a stimulus is present or absent. In clinical practice, the hearing levels of individuals are determined using an audiometer that allows presentation of a pure-tone of specific intensity and frequency. Functional images of the auditory nervous system have been obtained using stimuli such as pure-tone, speech, noise, etc. However, the observed activation evoked by the stimulus is confounded with the neuronal response evoked by scanner noise generated during imaging. Hence, researchers have been developing fMRI techniques to overcome the inadvertent effect of scanner noise on fMRI studies of the auditory cortex. Silent event related fMRI is a recently reported fMRI technique diminishing the confounding effects of background scanner noise. A drawback of sfMRI is that it requires long acquisition times (30-40 min) to achieve statistically significant activation. An additional complication associated with all fMRI data is that measurements obtained at consecutive times tend to exhibit substantial temporal correlation. Such correlation structure complicates the identification of brain locations (voxels) demonstrating statistically significant activation. We propose an approach for detecting activation with high statistical power and low false-positive rate. To accomplish these goals of high power and low type I error rate in sfMRI with shorter acquisition times, we describe a statistical model that accounts for the spatial and temporal correlation structure of the haemodynamic response. Temporal dependence within each voxel's measurements is modelled, and a regional measurement-error-free kriging predictor is used to combine information from neighbouring voxels when assessing voxel activation. Instead of simply applying a post hoc smoothing to thevoxelwise test statistics (e.g. t statistics), we attempt to make optimal use of information in the locality of each voxel when estimating the voxel's mean, variance, and temporal dependence parameters. The primary advantage to this spatial modelling approach is that the degree to which voxel parameters are smoothed is driven by the data. Thus, we are not subjectively smoothing noisy data, but objectively estimating the noise-free version of the spatial processes associated with the response. The resulting voxel activation maps exhibit substantially more spatial continuity than other currently used approaches, while exhibiting desirable inferential properties including a lower false-positive rate and high power for detection of activated regions. Minimal computational resources are necessary to carry out the approach, which yielded voxel activation maps for our experiment in only minutes.