Poster No.: 51
Separation of Motor Preparation and Execution Regions Using Meta-K-Means Clustering on fMRI Single Trial Data
A. Purushotham*, F.A. Neilsen+, L.K. Hansen+, S.G. Kim*

*Ctr for Magn Reson Res, Depts of Radiology and Biomedical Eng, Univ of Minnesota, USA +Inst of Math Modeling, Danish Tech Univ, Denmark

Objective
This study was an attempt to disgiinguish the motor regions of the human brain involved in planning motor movement from those involved in its execution, using single-trial fMRI. Since the behavioural response during such a motor experiment is inconstant from trial to trail, K-means clustering (1) following finite impulse response (FIR) filtering (2), was employed to extract the above information from fMRI data.

Materials and Methods
The paradigm consisted of a delayed, cured joystick movement task. The subject was required to move a cursor from the centre of the screen to the memorized location of a target, after a delay period following target presentation. Each run consisted to 16 such repetitions, or trails, with variable delays (0-3 sec) and target-locations. The intertrial interval was 20 sec. Behavioural data including reaction and movement times, were recorded. The study was conducted on normal adult human subjects using a 4 Telsa MRI system. 10 coronal sections of thickness 5 mm each were imaged, including the primary, supplementary and pre-motor areas. 64X64 size single-shot EPI image wer acquired every 100ms (1 volume/sec).

Analysis and Results
Data analysis was done using the Lyngby toolbox(http://hendrix.imm.dtu.dk/software.software.html). The means of each trial in a run was zeroed. Two paradigm functions were constructed for analysis based on the behavioural data: one for motor preparation, the other for execution. FIR filtering was done on the time-courses using these two reference functions, and K-means clustering was performed on these filters. The non-noise clusters had FIR filters that in the majority were of the form of a single smooth, positive wave, differing primarily in the latency and amplitude of the response.
Fig. 1: Map of a cluster of voxels
Fig. 2: Mean FIR filter of cluster from fig. 1
Fig. 3: Mean FIR filter of a different cluster

Figure 1 is a map of voxels that belong to one such cluster, while fig. 2 shows the centre of this cluster, i.e., the mean FIR filter of the member voxels. The paradigm used in this case was of motor execution. The FIR filter waveform can be considered to represent the BOLD response of the particular voxel. Thus the haemodynamic response appears to last for ~18 seconds in the motor areas. The response in this cluster has begun clearly before the onset of motor execution, observed as a 'folding over' of the upstroke of the positive wave, to the last 2-3 coefficients. Hence this cluster of voxels shows activation during the preparation phase of the task. In fig. 3 is shown the centre of another cluster. These voxels have a BOLD response that occurs soon after the onset of the execution period. The response during the preparation phase is very small, if at all present.

Conclusion
The voxels in the motor area clearly showed differences in activation in terms of temporal relation to the planning and execution phase of a motor task. FIR filtering followed by K-means clustering on the coefficients, provide a good tool not only for separate voxels with different temporal activation patterns in single-trial experiments but also to visualize the BOLD response.

Acknowledgment
Supported by NIH (MH57180 and RR08079).

References
1. Goutte, C. et al. Neuroimage, 1998, 7: S610;
2. Nielsen, F.A. et al. Neuroimage, 1997, 5: S473