(n0n)stationarity of Temporal Dynamics in Fmri

Estimation of the temporal information encoded in the observed hemodynamic response in functional MRI (iMRI) is of great interest. One assumption that most of the current studies make is that the mean waveform observed is consistently locked with the stimulus. and variably,dis@buted about its mean. Because the noise is hi any ViOlahOnS of a method 8 r filtering and tracking the hemodynamic response using a recursive least squares (RLS) algorithm and probabilistic shift ma s @sm). Our initial results clearly demonstrate an overafchan e in the amplitude from greater to lesser and a latency shift 8om shorter to longer in primary visual cortex. These changes ma indicate fah ue or adaptation of the neuronalpatterns or &ood supply., d e fact *at we observe a change in primary visual cortex is notable since this suggests that changes in brain regions which are specialized for hi her cognihve functions may exhibit even larger changes. h i s suggests the importance of detemuning the degree to which the measured bram region adapts to the faradifn presented in any fMRI experiment. n e , uchon: Functional MRI is a brain imaging technique sensitive to blood oxygenahon and flow, a change m which is coupled to neuronal activation. Most fMRI methods ignore the possibility of receivin a different response with the same stimulus. It is not difficui? to imagine a scenario in which the brain res onds differently over time (e.g. adaptation to stimulus, fatigue). We apply adaptive techniques to observe chan es in the res onse over time. We appl a simple visual stim&s (a flash of light to both eyes) eridcally and model the res onse with a finite number oPFourier terms. This methocfhas been ap lied to EEG data using a LMS filter [l j whrch we then appfy to fMRI data and extend to an RLS ada tation. The maxunum value and latency is calculated for eaci epoch. We then calcula!e the distribution of positive and negative latency and amplitude changes. The histogram of these shifts indicates which voxels demonstrate a significant chan e during the course of the experiment. Simulated fMh Data: A data set was generated, usin a periodic Poisson s i p in additive white Gaussian noise. 'he values for the amp itude and standard deviation were based upon empirical results from the iMRI experiment described later. The RLS filter converged rapidly, as.expected, after @e fust 30 samples.[2] erformance of the RLS filter h shifin the Poisson sign3 gy eee seconds gradually iuring &e course of the expenment. This was done for a value of h = 0.99, thus allowing the RLS filter to slowly forget past data and allow tracking to occur. We also measure the tracking error and detemune that it is sufficiently small for our purposes. Experimental Methodology.: We used a gradient-echo echo-planar pulse sec$mce wt!~ TR=l? TE=39ms, a=9Odeg, FOV=24cm, 128x12 , slice thickness-.lmm. 750 Five-slice this assum tion are not easily observed. % e have developed We then determined the trackm and the results are saved for the ends of epoch 5 (after we are confident convergence has occurred) through epoch 25 to produce 21 30s time courses for each voxel. We use a correlation filter, tuned to the primary visual cortex, to mask non-activated voxels 141. For each epoch, an amplitude and 0-7803-5674-8/99/$10.00