ausblenden:
Schlagwörter:
-
Zusammenfassung:
Introduction:
We have recently proposed a new approach called LISA for statistical inference of fMRI data (Lohmann et al, 2018). LISA incorporates spatial context via a nonlinear filter so that no initial cluster-forming threshold is needed, and spatial precision is largely preserved. Multiple comparison correction is achieved by controlling the false discovery rate in the filtered maps. In our previous publication, we have described this method only for first-level (single-subject) designs using precoloring as a technique for incorporating temporal autocorrelation, and for simple second-level designs (onesample and twosample group studies). Here we introduce an extension of LISA so that most scenarios in fMRI-based research are covered. Specifically, LISA can now also use prewhitening for single-subject analyses, and it can handle arbitrary 2nd-level designs matrices. LISA can thus serve as a general tool for statistical inference in fMRI.
Methods:
GLM-based 2nd level analysis: The input into LISA for a general 2nd-level inference is a list of 1st-level contrast maps together with a 2nd-level design file and a contrast vector. The output is a map thresholded such that FDR < alpha for every voxel. For ease of use, the 2nd-level LISA software is compatible with format requirements of FSL (Jenkinson et al, 2012). In particular, the 2nd-level design file is a text file that may be generated manually or using a GUI such as FSL's Glm. LISA uses random permutations for statistical inference, (Winkler et al 2014, 2016). In some cases - for example repeated measures designs - permutations must be constrained to ensure exchangeability. These constraints are specified in an additional file which also follows format specifications compatible with FSL.
Single subject analysis using prewhitening: The input into LISA for a single subject analysis is a preprocessed 4D fMRI data set, a design matrix and a contrast vector. The output is a map thresholded such that FDR < alpha for every voxel. For the prewhitening approach, temporal autocorrelations are modelled as described in (Worsley et al, 2002). A null model is estimated based on random permutations of task labels. This approach requires that the experimental design is properly randomized and the inter-trial distances are large enough so that the trials are independent and task labels are exchangeable.
We applied LISA using prewhitening and SPM using the Gauss Random field model to single subject data acquired at a 9.4T Siemens Magnetom scanner. For SPM, we performed spatial smoothing using FWHM=2.4mm (twice the voxel size) with an initial cluster-forming threshold p=0.001. For LISA no smoothing was used. The subject (female, 29 yrs) performed an n-back task, where each trial consisted of the presentation of ten faces. Trials lasted for 20 secs and started with a brief presentation of the task type (2-back or 0-back). Each run consisted of 8 trials (4x 2-back and 4x 0-back), the order of the trials was randomized. Five runs of about 4.5 minutes each were recorded.
For an application of the new 2nd-level LISA in a reach-and-grasp experiment, see (Molla et al, submitted to OHBM 2019).
Results:
We applied statistical inference to the single subject data as described above, focusing on the contrast '2-back minus 0-back'. We found that LISA found more activation areas in anatomically plausible regions. In the SPM based result, spatial acuity was diminished due to spatial smoothing (Fig. 1).
Conclusions:
Due to problems regarding lack of statistical power and inflated false positive rates improved statistical inference tools are urgently needed. Furthermore, ultrahigh field scanners >= 7T are becoming more widespread and require algorithms that preserve spatial precision, so that spatial smoothing and cluster-forming thresholding can be avoided. Here we have introduced a generalized version of the LISA algorithm. We believe that it will help to address the needs of the human mapping community.
