|iTT Toolbox for SPM|
IntroductionThe purpose of thresholding in functional MRI (fMRI) analyses is to classify image voxels as activated, deactivated, or not involved. To accomplish this goal most statistical methods determine a level of significance (i.e. define a threshold). Voxel-based approaches [1, 2] that apply only one threshold always rely on a compromise between specificity and sensitivity. Some more advanced approaches attempt to provide both high specificity and sensitivity e.g. using information about clusters of activated voxels [3-5]. However, current thresholding strategies for the analysis of functional MRI (fMRI) datasets may suffer from specific limitations (e.g. with respect to the required smoothness) or lead to reduced performance for a low signal-to-noise ratio (SNR). Although a previously proposed two-threshold (TT) method offers a promising solution to these problems , the use of preset settings limits its performance. This work offers an optimised TT approach that estimates the required parameters in an iterative manner.
The TT method assumes that the actual distribution (histogram) of statistical parameters comprises both truly activated voxels and underlying noise. Because the noise is expected to be Gaussian distributed [7-10], a Gaussian function is fitted to the central part of the histogram to identify activated voxels above a strict upper threshold (p = 0.0001). Subsequently, a second lower threshold (p = 0.05) is applied to direct neighbours of the voxels selected in the first step. These probabilistic thresholds are derived from the fitted Gaussian curve. Altogether, the TT method therefore employs four parameters: the upper and lower cuts determining the central part of the histogram for Gaussian fitting (horizontal lines in Fig. 1) and the upper and lower thresholds determining statistically significant activated voxels (vertical lines in Fig. 1).
In general, the iTT method presents with remarkable sensitivity and good specificity that outperforms almost all conventional approaches in all cases. This also holds true for challenging conditions such as high spatial resolution (right column in Fig. 2), the absence of filtering (lower row in Fig. 2), high noise level, or a low number of task repetitions.
iTT Toolbox for SPM
The iterative Two Threshold (iTT) approach has been added to SPM in the form of a Toolbox thus enabling the user to choose “iTT” or “TT” (original TT – without iterative optimisation) for thresholding approach and select the lower threshold (default p = 0.05). The upper threshold is determined by the iterative optimization (iTT) or set to p = 0.0001 (TT), although this default value can be changed.
The iTT toolbox contains a separate graphic user interface for configuration, which may be selected from the Toolbox list.
When using this toolbox, please use the following reference:
1. Genovese, C.R., N.A. Lazar, and T. Nichols, Thresholding of statistical maps in functional neuroimaging using the false discovery rate. Neuroimage, 2002. 15(4): p. 870-8.
|< Prev||Next >|