Item

Abstract

Introduction:
The inflation of false positive detections due to insufficient control of the effects of massive multiple independent testing is a crucial challenge for the statistical validity of fMRI analyses. One approach uses cluster extent, calculating threshold cluster sizes based on an acceptable False Discovery Rate (FDR) or Family Wise Error Rate (FWER), e.g. derived from Gaussian Random Field Theory (RFT) (Chumbley et al, 2009). However, the dependency of the RFT on the shape of the autocorrelation function and on a constant spatial smoothing across the brain may result in invalid cluster-wise inference (Eklund et al, 2016). The Local Indicator of Spatial Association algorithm (LISA) has been recently developed (Lohmann et al, 2018) with the aim of overcoming such shortcomings. LISA does not require spatial smoothing during data preprocessing and performs a non-linear spatial filtering only after a z-map has been calculated. It tests for statistical significance at a voxel level using a non-parametric, permutation-based test. In this study, we assessed the performance of LISA in a group level analysis of data from a visuomotor experiment that incorporated kinematic covariates of no interest.
Methods:
The dataset consisted of measurements from 27 healthy subjects (3T Siemens TRIO, slice thickness = 3mm; 36 slices interleaved acquisition; in-plain resolution 3mm × 3mm; TR = 2.47s; TE = 33ms). The participants reached and grasped either an object commonly used in everyday life or a geometrical object with no distinctive feature which was matched for dimension (Sheygal, 2015). Two MR-compatible cameras recorded hand movements during each run. Seven durations of kinematic components of the reach-to-grasp movement were identified offline and later used as covariates in the group analysis (Sheygal, 2015). LISA expects the following inputs: 1) A contrast image for each participant and each condition; 2) a text file with the design matrix including additional covariates; 3) the definition of the blocks within which permutations are allowed; 4) a contrast vector. We calculated a paired test between the two conditions across subjects using LISA corrected for multiple comparisons (FDR < 0.05).
Results:
The group analysis using LISA detected a bilateral activation in the anterior intraparietal sulcus (aIPS), ventral premotor cortex (vPM) and anterior cingulate cortex (aCC), as well as a unilateral activation at the left lateral occipital cortex (LOC). These findings were highly plausible, given previous observations in similar experiments. The group analysis using SPM12 and cluster extent corrections detected fewer clusters, not detecting for example aIPS. Detected clusters had a smaller volume at all locations (Fig. 1).
Conclusions:
Our results showed that LISA detects precisely activity associated with the task in anatomically plausible regions also when additional covariates have been included. The high specificity does not come with a loss of sensitivity. When compared to the results produced with SPM, it becomes clear that no cluster of activation is neglected by the algorithm. On the opposite, new well-defined - anatomically plausible - clusters of activation are detected.