|
|
|
1Department of Neurology,University of Minnesota, USA, 2Department of Electrical Engineering,University of Minnesota, USA, 3Informatics and Mathematical Modeling, Technical University of Denmark,DK, 4Department of Health Information,University of Minnesota, USA, 5Department of Radiology,University of Minnesota, USA |
| Modeling & Analysis Abstract |
| INTRODUCTION. Several frequency domain activation-detection
methods based on multi-taper spectral estimation[1] were evaluated using
fMRI simulation data and a normalized partial ROC index within the NPAIRS
framework. The results were also compared with univariate Bayesian
and GLM techniques, as well as multivariate Canonical Variates Analysis (CVA)
in the time domain. METHODS. Data Simulation:The simulation data were generated using a brain-mask volume(64*64*32) from one subject for a 1.5T fMRI experiment in which every volunteer was asked to perform two runs of a static force task alternating six rest and five force periods/run (44s/period, TR=4s). Four artificial Gaussian blob (FWHM=1,1.5,2,4 pixels) activations, each restricted to a 7*7 square, were added to different locations in a single slice. To form the simulated time sequence, the blobs were then multiplied by 1) Block: the on-off reference function for two parametric static force runs; 2) a sinusoidal wave of 0.011 Hz—the fundamental block frequency; both were convolved with a Poisson shaped (λ=7.3) HRF. After adding white noise to the sequence and normalizing the SNR at the blobs’ centers to be 1, two sets of simulation data (Block & Sinusoidal) were obtained. Finally the spectra of the time series corresponding to each voxel were estimated using multi tapers[1] (MTM, the number of Slepian sequences K=3, time–half band width product NW=2, the Rayleigh frequency N=128). Analysis: In frequency domain: 1) Harmonic F test; 2) regression with reference F test; 3) CVA (each frequency component corresponding to a group, total group number=N/2); 4) Complex Singular Value Decomposition (SVD) was performed over the MTM spectral data [2]. In time Domain: 1) GLM; 2) Bayesian Detection [3]; 3) 2 Class CVA; 4) 11 class CVA. All of the above 8 methods were implemented in the NPAIRS framework [4] with consensus, reproducing Z-score images as the final results for each data set. The normalized area under a partial ROC curve (pAUC) with FPRs ranging between 0 and 0.1 was calculated as the performance measurement. RESULTS/CONCLUSIONS.The figures illustrate the pAUC trends as a function of Gaussian blob threshold— only the voxels in the Gaussian Blob with values no less than the threshold were treated as truly activated —for the 8 detection methods applied to the Sinusoidal and Block simulated data sets. Our conclusions require further justification using real fMRI data: 1) GLM performs the worst of all methods tested; 2) Time domain CVA, whether 2 class or 11 class, performs the best; 3) the Bayesian method dramatically outperforms all other univariate methods because it takes maximum advantage of prior information; 4) Among frequency domain methods, complex SVD performs the best; 5) Regression with reference F test uses the 0.011Hz priori frequency and therefore performs better than the harmonic F test. REFERENCES.[1] Mitra PP, et al, Biophysical Journal 76: 691-708, 1999; [2] Sujit KP, master thesis, UMN2757487, 2003; [3] Hansen LK, et al, AIM 25: 35-44, 2002;[4] Strother SC, et al, NeuroImage 15, 747-771, 2002. ACKNOWLEDGEMENTS.This work is supported in part by NIH grant MH57180. |
|
|
