Topological FDR for neuroimaging.

Chumbley J, Worsley K, Flandin G, Friston K. · 2010

Read Topological FDR for neuroimaging. by Chumbley J, Worsley K, Flandin G, Friston K. (2010) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3221040.txt.

Introduction

Justin Chumbley, Keith Worsley, Guillaume Flandin, and Karl Friston described a new statistical safeguard for brain imaging in this 2010 technical note in NeuroImage. Brain scans are tested at thousands of locations, so researchers must control false discoveries without wiping out real signal. Their procedure works on the peaks of a statistical map rather than on individual points.

Drawing on random field theory, the method assigns each local maximum a significance value, then sets an adaptive threshold that keeps the false discovery rate in check. Simulations compared it with family-wise error control and with cluster-wise and voxel-wise alternatives. Peak-level control proved more sensitive than the stricter standard at little cost in false positives, and far more spatially accurate than voxel-wise control. An attention study illustrates the method.

Catalog record

CREATORChumbley J, Worsley K, Flandin G, Friston K.
YEAR2010
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC3221040

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Topological FDR for neuroimaging By J. Chumbley, K. Worsley, G. Flandin, K. Friston Published in: Neuroimage DOI: 10.1016/j.neuroimage.2009.10.090 ABSTRACT In this technical note, we describe and validate a topological false discovery rate (FDR) procedure for statistical parametric mapping. This procedure is designed to deal with signal that is continuous and has, in principle, unbounded spatial support. We therefore infer on topological features of the signal, such as the existence of local maxima or peaks above some threshold. Using results from random field theory, we assign a p-value to each maximum in an SPM and identify an adaptive threshold that controls false discovery rate, using the Benjamini and Hochberg (BH) procedure (1995). This provides a natural complement to conventional family wise error (FWE) control on local maxima. We use simulations to contrast these procedures; both in terms of their relative number of discoveries and their spatial accuracy (via the distribution of the Euclidian distance between true and discovered activations). We also assessed two other procedures: cluster-wise and voxel-wise FDR procedures. Our results suggest that (a) FDR control of maxima or peaks is more sensitive than FWE control of peaks with minimal cost in terms of false-positives, (b) voxel-wise FDR is substantially less accurate than topological FWE or FDR control. Finally, we present an illustrative application using an fMRI study of visual attention. INTRODUCTION Numerous authors have remarked that the fundamental effects of interest in fMRI are distributed and spatially continuous (e.g. Heller et al 2006, see also Pacifico et al 2004 and Benjamini and Heller 2007). A recent communication (Chumbley and Friston, 2009) proposed a solution based upon conditional p-values associated with the spatial extent of excursion sets above an arbitrary threshold (conditional on such a set existing). Here, we revisit the same issue and present an equivalent solution based on p-values associated with maxima or peaks in a Statistical Parametric Map (SPM). This allows us to assign FDR adjusted p-values (which we shall refer to as q-values) to both peaks and clusters within any SPM on image data. Most instances of image analysis deal with signals that are continuous functions of some support; for example, anatomical space, time and frequency in EEG analyses or various combinations of space, time and frequency: In Chumbley and Friston (2009) we discuss why signal is continuous in the majority of neuroimaging applications and have reviewed the various image reconstruction procedures and biophysical mechanisms that generate continuous signals. If signal is a continuous function of the search space, strictly speaking, every discovery is a true discovery, because signal is everywhere. These conditions present a challenge to the conventional application of FDR to separate tests or voxels. In Chumbley and Friston (2009) we described a solution based upon the topological features of the SPM. This involved controlling the false-discovery rate, not of voxels, but of connected components of the excursion set above some threshold (i.e., controlling the false-discovery rate of clusters). In this note, we pursue the same theme but control the false-discovery rate of maxima or peaks. Maxima or peaks are the most common topological features of activation profiles reported in the literature. Peaks are a useful characterisation of regionally specific responses because they have a well-defined location (usually reported in a standard anatomical space) and lend themselves to data-basing. Note that a simple1 suprathreshold region in an SPM could contain several peaks. Only at high thresholds, will the number of suprathreshold regions and peaks be equal. Usually, people display the SPM, at some threshold, as a maximum intensity projection or rendered image and then discuss peaks that survive a statistical criterion. In what follows, we will consider signal to be a continuous function of a statistical search space and an effect or activation to be some topological feature of this function. In particular, we will consider an activation to be a peak of the signal and a discovery to be the corresponding peak in the statistical process or SPM. We thus convert a continuous process (signal) to a topological feature (activation) that does or does not exist. Finally, we may choose to define “true positive” and “false positive” topological decisions (with a latitude that is not possible in the definition of decisions in discrete tests). For example, one might reasonably label a discovered peak in an SPM “true positive” if it lies within one FWHM of a true peak in the underlying signal, and “false positive” otherwise. However, this categorisation of discoveries as true or false depends on an arbitrary threshold (here one FWHM). More generally then, we consider the full distribution on the distance between observed and true maxima under a procedure. This measure of the spatial accuracy is of interest per se and can be readily transformed back into a qualitative false-positive rate (by evaluating its cumulative density at any specified distance). To assess accuracy, one can compare these distributions of spatial error, over procedures. This note comprises three sections. In the first, we present the theory behind the control for false-discovery rates of peaks. We then assess its performance in relation to conventional peak FWE in terms of two complementary measures: (a) the relative number of discoveries and (b) the spatial displacement of each discovered peak from the true underlying peak (defined as the minimum distance between true activations and discoveries). We report conditions under which the peak-BH procedure identifies more activations (as measured by a) at a negligible cost of identifying more false ones (as measured by b). This enables us to compare different procedures in quantitative terms, while side-stepping the different definitions of a discovery (e.g. of “activation”) under voxel-wise and topological approaches. In the final section, we apply the alternative procedures (FWE and FDR on peaks and FDR on clusters and voxels) to a standard fMRI dataset (the fMRI study of visual attention available at http://www.fil.ion.ucl.ac.uk/spm). THEORY In this section, we present the theory behind topological inference based on FDR for peaks. The usual approach to topological inference is as follows. First, derive the null distribution of the SPM, assuming no signal and Gaussian error fields. Then define a random set Ah,s of simple regions within the SPM, where each region exceeds h in height and s in spatial extent (hi > h,si > s)with i = 1, …, Mh,s. Finally, derive the distribution of Mh,s, under the assumptions described in Friston et al., 1996. This allows one to calibrate the thresholds (h,s) to control the expected number of false positives E(Mh,s), or until the probability of getting any false positives p(Mh,s > 0) is sufficiently small; this is the family-wise error (FWE) rate. Having established the decision criterion (h,s) any observed region satisfying (hi > h,si > s) indicates the qualitative existence of one or more maxima in the underlying signal field (e.g., “an activation”). An alternative approach, used in this work, views the random excursion set Ah,s as comprising two random subsets Ah,s⊇ = {Ah,s,f, Ah,s,t} of false- and true-positive domains. The integer count of simple regions in Ah,s is now Mh,s = Mh,s,f + Mh,s,t. We use a simple algorithm (Benjamini and Hochberg, 1995) to calibrate (h,s) so that Mh,s,f/Mh,s < α on average. This approach does not seek to restrict the absolute number of

Read the complete text