Bootstrap generation and evaluation of an fMRI simulation database - PubMed (original) (raw)

Bootstrap generation and evaluation of an fMRI simulation database

Pierre Bellec et al. Magn Reson Imaging. 2009 Dec.

Abstract

Computer simulations have played a critical role in functional magnetic resonance imaging (fMRI) research, notably in the validation of new data analysis methods. Many approaches have been used to generate fMRI simulations, but there is currently no generic framework to assess how realistic each one of these approaches may be. In this article, a statistical technique called parametric bootstrap was used to generate a simulation database that mimicked the parameters found in a real database, which comprised 40 subjects and five tasks. The simulations were evaluated by comparing the distributions of a battery of statistical measures between the real and simulated databases. Two popular simulation models were evaluated for the first time by applying the bootstrap framework. The first model was an additive mixture of multiple components and the second one implemented a non-linear motion process. In both models, the simulated components included the following brain dynamics: a baseline, physiological noise, neural activation and random noise. These models were found to successfully reproduce the relative variance of the components and the temporal autocorrelation of the fMRI time series. By contrast, the level of spatial autocorrelation was found to be drastically low using the additive model. Interestingly, the motion process in the second model intrisically generated some slow time drifts and increased the level of spatial autocorrelations. These experiments demonstrated that the bootstrap framework is a powerful new tool that can pinpoint the respective strengths and limitations of simulation models.

PubMed Disclaimer

Figures

Fig. C.1

A diagram representation of the data-generating process (DGP) of an individual fMRI dataset, and the derivation of a measure of interest.

Fig. C.2

A diagram representation of the evaluation of a simulation database. The central scheme was highlighted by a dotted line, and represented the replication of the real database. The rest of the scheme summarised the process of generating statistical features of interest, deriving group-level summaries of these features and using a non-parametric i.i.d. bootstrap to assess significant departures between the real and simulation databases.

Fig. C.3

A diagram representation of the ADD data-generating process. The first four components of the model were deterministic linear mixtures (baseline, slow time drifts, physiological noise and activation). The noise was a sample of a Gaussian random noise independent in space and time and with voxel-specific variance. These five components were added to generate one fMRI simulation.

Fig. C.4

A diagram representation of the SM data-generating process. The three deterministic components of the model (baseline, physiological noise and activation) were first mixed to produce a motion-free space-time dataset with high spatial resolution. For each slice and each volume, the rigid-body motion was applied to the high-resolution volume corresponding to the slice time acquisition and the result was further averaged on a low-resolution voxel grid corresponding to the slice. Some random Gaussian noise independent in space and time and with voxel-specific variance was finally added to the low-resolution dynamics to generate one sample of fMRI simulation.

Fig. C.5

Axial slices of the MNI 152 anatomical template which were used to represent the maps, along with _z_-coordinates. The position of the axial slices were also indicated on one sagital slice.

Fig. C.6

Group-average maps of relative variance for real, ADD and SM data. The variance of the estimated components was derived at every voxel as a percentage of the total variance for the real dataset. The individual maps were spatially transformed in the MNI 152 space and averaged across all subjects and tasks, with the exception of the activation component, which is averaged only across subjects and is represented here for task HA.

Fig. C.7

Group-average maps of spatial and temporal autocorrelation for motion– corrected and residual data. The maps derived for each dataset were spatially interpolated in the MNI 152 space and averaged across all subjects and tasks.

Fig. C.8

Distribution of the relative energy in a principal component analysis (PCA). A spatial PCA was performed on real and simulated datasets and the variance contribution of each component was derived as a percentage of the total contribution of the first top 60 components. The average over all motion-corrected and residual datasets are represented in panel a and b respectively.

References

1. Ogawa S, Lee TM, Kay AR, Tank DW. Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc Natl Acad Sci U S A. 1990;87:9868–9872. -PMC -PubMed
1. Logothetis NK, Pauls J, Augath M, Trinath T, Oeltermann A. Neurophysiological investigation of the basis of the fMRI signal. Nature. 2001;412:150–157. -PubMed
1. Fox PT, Laird AR, Lancaster JL. Coordinate-based voxel-wise metaanalysis:dividends of spatial normalization Report of a virtual workshop. Hum Brain Mapp. 2005;25:1–5. -PMC -PubMed
1. Della-Maggiore V, Chau W, Peres-Neto PR, McIntosh AR. An empirical comparison of SPM preprocessing parameters to the analysis of fMRI data. Neuroimage. 2002;17:19–28. -PubMed
1. Logan BR, Rowe DB. An evaluation of thresholding techniques in fMRI analysis. Neuroimage. 2004;22:95–108. -PubMed

Bootstrap generation and evaluation of an fMRI simulation database - PubMed (original) (raw)