|
|
|
1Neurology Department, University of Minnesota, 2Electrical Engineering, University of Minnesota, 3Radiology Department, University of Minnesota |
| Modeling & Analysis Abstract |
| Introduction: fMRI processing pipeline options
have a big impact on the final fMRI analysis results due to the low SNR of
fMRI data. Therefore, it is important to evaluate different processing options
and to optimize fMRI pipelines[1,2]. The standard evaluation method--simulations
and ROC curves has the advantage of accuracy due to known ground truth, but
it has serious limitations due to the unknown match between a simulation
and any real data set[1,2,3]. Methods: The 16 subjects' static force dataset described by LaConte et al. was utilized ([2]). Optimization was based on two criteria: prediction accuracy and activation pattern reproducibility, generated by NPAIRS through cross-validation[4]. Four main preprocessing options were considered generating 10 processing pipelines: (1) motion correction (black lines); (2) spatial smoothing at 0,1.5,6 pixels with 2D Gaussian filter, plotted as filled circles, triangles and squares, respectively; (3) temporal filtering at 0,1,3 cosine cycles with cosine basis function in GLM, plotted as ; fourth, denoising through PCA/CVA dimension reduction with subspace selection 10,25,50,75 components. Data analysis was performed by PCA/CVA model within NPAIRS framework. According to the experiment design and research interest, different CVA models were employed. 11 class CVA model was grouped by 11 brain states (including 5 force levels and 6 baselines) in a single trial/run. In addition, 2 class CVA (grouped by baselines and activations) was performed as a comparison with the 11 class CVA. NPAIRS was integrated into Fiswidgets environment[4]. This allows us to exchange resources with other fMRI packages and to better control the execution of fMRI processing pipelines on the Fiswidgets platform. Results: 10 processing pipelines were formed. The results of 11 class CVA analysis are summarized in Fig.1 and Fig.2 which give the average prediction accuracy and reproducibility across all 16 subjects for baselines (Fig. 1) and for activations (Fig.2). Although there is considerable variance in the analysis results among subjects, the mean prediction vs. reproducibility plots illustrate the trend in 11 class CVA case. First, detrending flattens the baselines (at 3 cosine cycle (Fig.1)), but it also seriously lowers the prediction accuracy for activations and causes the model badly predict the data (below randomly guessing probability). Second, spatial smoothing helps increase prediction accuracy and reproducibility only when the data is not highly detrended and the dimension reduction (denoising) is properly performed. Fig. 3 and Fig. 4 provide 2 class CVA results which were obtained by the same pipeline options. We observed that detrending increases the prediction accuracy for ! baselines in this case while there is a reverse trend at 1 cosine and 3 cosine for force levels. Again, close relationship between spatial smoothing, temporal filtering and denoising was observed here. Further, 2 class CVA model reveals that motion correction plays a significant role in terms of increasing prediction accuracy and reproducibility as shown in Fig 3 and 4. References 1.Shaw ME et al, Neuroimage 19:988-1001, 2003. 2.LaConte S et al, NeuroImage 18:10-27, 2003 3.Strother SC et al, Neuroimage. Apr; 15(4): 747-71 2002 4.Fissell K et al, Neuroinformatics 1:1 111-160, 2003 Acknowledgments: This work is supported by NIH grand MH57180. |
|
|
|
|
|
|
|
|
