Detecting change in BOLD signal between sessions for atlas-based anatomical ROIs

doi:10.1016/j.neuroimage.2008.01.001

NeuroImage

Volume 40, Issue 3, 15 April 2008, Pages 1157-1165

https://doi.org/10.1016/j.neuroimage.2008.01.001 Get rights and content

Abstract

An algorithm using pre-defined regions of interest (ROIs) to detect differences between sessions in Blood Oxygen-Level Dependent (BOLD) signal is proposed and results from a reproducibility study are reported here. It is important to know whether tests for change have the desired statistical properties, e.g., low variability between sessions and unbiased false-positive rates, under null conditions so one is confident in any conclusions based on the metric and test used. This study examined three cognitive tasks: Stroop, response inhibition, Sternberg digits, and a visually-cued finger tapping task, in 20 healthy subjects. Each subject had two fMRI sessions, one week apart. A series of ROI summaries was constructed by choosing different proportions of voxels from the ROI and calculating the mean of t values for the selected voxels. The choice of voxels was based on the magnitude of the t values, selecting the maximum value, then the top 1%, and so on until all voxels in the ROI were included. No ROIs were found to have significant differences between sessions based on paired comparison t tests. Generally, the observed false-positive rates were near the expected rates for all summaries and tasks, although the Sternberg and Response Inhibition tasks did have higher false-positive rates when α ≥ 0.1 for some ROI summaries. This study indicates results are reproducible and also have the desired statistical properties under null conditions.

Introduction

In drug development, we define a biomarker to be “a characteristic that is objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes, or pharmacological responses to a therapeutic intervention.” (Atkinson et al., 2001). In order to use fMRI as a biomarker, we need to be able to measure the amount of change due to a therapeutic intervention, not simply detect the occurrence of a change.

The limitations of voxel-based group analysis are well known (Brett et al., 2002). Even with smoothing and spatial normalization to a standard space, there is enough anatomical and functional variability between people that there can be significant differences between voxel-based group activation and activation in individuals (Vandenbroucke et al., 2004, Zahn et al., 2004, Eugene et al., 2003). These limitations and the complications of finding an appropriate way of quantifying the change make voxel-based group analysis difficult to use as a biomarker.

One alternative to voxel-based group analysis is to define regions of interest (ROI) in individuals and combine a summary of the ROIs to obtain a group result. Summarizing over the voxels in an ROI essentially smoothes over a restricted neighborhood of voxels thereby helping to reduce the effect of voxel-level spatial variability that is present in voxel-based group analysis. Another benefit of using ROIs is that translation to standard space is no longer required. Hence the manipulation of the “original” signal values during translation to standard space is removed from the analysis process.

There are some potential issues with summarizing over ROIs. Summarizing over areas larger than a voxel may decrease sensitivity as active voxels are grouped with inactive voxels in the summary. Also, active functional regions may not correspond to the defined anatomical ROIs. The functional regions may be a subregion of the anatomical ROI or it may be split between several ROIs leading to the signal being diluted during summarization of the ROI if the summary function is not chosen carefully.

Two summaries, location of the peak activation, and the number of voxels above a threshold (Rombouts et al., 1998, Constable et al., 1998), have been used to examine reproducibility of the fMRI signal, but are not good biomarker candidates because they do not easily quantify change due to a drug. The peak location is not expected to change with therapeutic intervention, although the magnitude might. The number of voxels above a threshold has several known limitations. The measure is dependent on an arbitrary threshold and results could be misleading if activation decreases to just below the threshold. The number of voxels above a threshold has also been shown to be highly variable compared to mean % signal change in an ROI (Cohen and Dubiosn, 1999).

The mean activation (e.g., mean t value or z score) over all or some voxels in a region has also been used (e.g., Dunckley et al., 2005, Iannetti et al., 2005). Typically, in a study only a single summary mean value is examined. In this paper we examine a family of summary measures which average over an increasing proportion of voxels in the ROI. This family of results presents a continuum of summaries from maximum signal in the ROI to the average of all signal values in the ROI. It also allows the active voxels to be filtered from the inactive voxels to some degree, hence improving the sensitivity of the summarized response.

Regions of interest that are investigated during a study are usually few in number and defined either functionally (for example, by making a mask from the active voxels determined from a voxel-based group analysis or separate scanning session) or anatomically. We chose to use an anatomical atlas to define the ROIs. This approach allowed us to predefine the ROIs and provide independence between ROI definition and functional response. We also chose to divide the entire brain into anatomical regions rather than focus on a few ROIs where activation or change in activation are expected. This allows us to systematically test regions where we expect activation and regions where we do not expect activation. A template based ROI map is easier to produce than a hand drawn map, and evidence suggests these automated mapping procedures may be as reliable as hand-drawn maps (Sun et al., 2007). We used the atlas developed by Tzourio-Mazoyer et al., 2002, for this paper, and have not investigated other available atlases (e.g., Hammers et al., 2002, Van Essen, 2005, Klein and Hirsch, 2005, Eickhoff et al., 2006).

In this paper we outline a methodology for whole brain analysis using ROIs which are summarized using a family of summary statistics. To assess the value of this approach as a biomarker for fMRI it is important to understand the degree of reproducibility between sessions in the metrics to be used and to have some confidence in the validity of the statistical tests that are likely to be used for second-level analysis. Reproducibility is characterized here by the standard deviation of the difference between sessions and potential temporal trends in response are examined through tests for mean differences between sessions. Intraclass correlation coefficients (ICCs) are not used to evaluate reproducibility here because of the variation in activation across all ROIs. For example, many ROIs are expected to have no association with a given task, so variation between subjects is likely to be low since the observed signals should just be random noise. Hence the ICC would be very low in these cases due to lack of variation between subjects rather than high variation between sessions, yet variability between sessions could be acceptably low as well. To assess the validity of the statistical tests it is necessary to investigate whether the tests have the desired statistical properties so as to have confidence in conclusions when the statistical tests reject the null hypothesis. To assess these statistical properties a study of the properties for the paired comparison t test under null conditions are examined. In particular the normality of the ROI summary values is evaluated and possible bias in the expected false-positive rates is examined.

Section snippets

Study design

Twenty healthy 18- to 65-year-old males and females with no history of neurological or mental disorders were recruited from the New Haven, CT area. Participants provided informed consent in accordance with the guidelines set forth by Western IRB.

The subjects completed two fMRI sessions approximately one week apart. During each fMRI session, 4 tasks were completed: fingertapping, response inhibition, Stroop, and Sternberg. The order of tasks was randomized within each session, except

Results

The set of gray matter voxels was considered to be the intersection of the SPM2 brain mask and the atlas gray matter mask. The atlas gray matter masks, however, were not always a subset of the SPM2 masks. The percent of voxels that were present in the atlas mask but not in the SPM2 mask ranged between 1.0 and 1.5% for the four tasks. In spite of our attempt to avoid having very small regions by a priori combining small regions with neighboring regions, we were not entirely successful and for

Discussion

The goal of this study was to begin to explore the viability of fMRI as a biomarker in drug trials. Two important steps in deciding whether fMRI has potential as a biomarker is assessing the quantitative reproducibility of a chosen signal over time across subjects and determining if the underlying assumptions for anticipated statistical tests used in future drug trials hold for replicate observations.

There were some important deviations in reproducibility during the creation of the ROIs,

References (26)

ConstableR.T. et al.
Quantifying and comparing region-of-interest activation patterns in functional brain MR imaging: methodology considerations
Magn. Reson. Imaging
(1998)
EickhoffS.B. et al.
Testing anatomically specified hypotheses in functional imaging using cytoarchitectonic maps
Neuroimage
(2006)
KleinA. et al.
Mindboggle: a scatterbrained approach to automate brain labeling
NeuroImage
(2005)
MostofskyS.H. et al.
fMRI evidence that the neural basis of response inhibition is task-dependent
Cogn. Brain Res.
(2003)
PetersonB.S. et al.
An event-related functional MRI study comparing interference effects in the Simon and Stroop tasks
Cogn. Brain Res.
(2002)
RomboutsS.A. et al.
Within-subject reproducibility of visual activation patterns with functional magnetic resonance imaging using multislice echo planar imaging
Magn. Reson. Imaging
(1998)
SunF.T. et al.
Automated template-based PET region of interest analyses in the aging brain
NeuroImage
(2007)
ThirionB. et al.
Analysis of a large fMRI cohort: Statistical and methodological issues for group analyses
NeuroImage
(2007)
Tzourio-MazoyerN. et al.
Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single subject brain
Neuroimage
(2002)
Van EssenD.C.
A population-average, landmark- and surface-based (PALS) atlas of human cerebral cortex
NeuroImage
(2005)

VandenbrouckeM.W. et al.

Interindividual differences of medial temporal lobe activation during encoding in an elderly population studied by fMRI

NeuroImage

(2004)

ZahnR. et al.

Recovery of semantic word processing in global aphasia: a functional MRI study

Brain Res. Cogn. Brain Res.

(2004)

BenjaminiY. et al.

Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing

R. Stat. Soc., Ser. A Stat. Soc. Ser. B.

(1995)

Cited by (15)

Reduced brain activation during spoken language processing in children with developmental language disorder and children with 22q11.2 deletion syndrome
2021, Neuropsychologia
Citation Excerpt :
By providing information about both the strongest activating and the strongest de-activating voxels, we aimed to offer a representation of the full range of beta values for each group, task and ROI. In earlier studies, the selection of a subset of voxels, based on their level of activity, as a basis for a single measure of task-related signal changes within an anatomically defined ROI, has been found to be a useable and valid approach (Buck et al., 2008; Buma et al., 2016; Mitsis et al., 2008; Tong et al., 2016). Using each of these two voxel selections, we subsequently calculated the mean (de-)activation per ROI.
Language difficulties of children with Developmental Language Disorder (DLD) have been associated with multiple underlying factors and are still poorly understood. One way of investigating the mechanisms of DLD language problems is to compare language-related brain activation patterns of children with DLD to those of a population with similar language difficulties and a uniform etiology. Children with 22q11.2 deletion syndrome (22q11DS) constitute such a population. Here, we conducted an fMRI study, in which children (6-10yo) with DLD and 22q11DS listened to speech alternated with reversed speech. We compared language laterality and language-related brain activation levels with those of typically developing (TD) children who performed the same task. The data revealed no significant differences between groups in language lateralization, but task-related activation levels were lower in children with language impairment than in TD children in several nodes of the language network. We conclude that language impairment in children with DLD and in children with 22q11DS may involve (partially) overlapping cortical areas.
The role of fMRI in drug development
2018, Drug Discovery Today
Citation Excerpt :
With such prescribed methodology in hand, an fMRI experiment can be reduced to a binary outcome more suitable to inform drug development decisions. In drug development, primary endpoints and hypotheses are typically based on prespecified ROIs [68,69] and power calculations are performed accordingly to avoid underpowering. Significant logistical requirements must be met when deploying fMRI to large numbers of heterogeneous, nonacademic imaging facilities, as is required for typical late-stage clinical trials.
Functional magnetic resonance imaging (fMRI) has been known for over a decade to have the potential to greatly enhance the process of developing novel therapeutic drugs for prevalent health conditions. However, the use of fMRI in drug development continues to be relatively limited because of a variety of technical, biological, and strategic barriers that continue to limit progress. Here, we briefly review the roles that fMRI can have in the drug development process and the requirements it must meet to be useful in this setting. We then provide an update on our current understanding of the strengths and limitations of fMRI as a tool for drug developers and recommend activities to enhance its utility.
Dynamic causal modelling of EEG and fMRI to characterize network architectures in a simple motor task
2016, NeuroImage
Citation Excerpt :
The order of sessions was counterbalanced, so a systematic order effect on task performance or brain activation is unlikely. However, we cannot exclude that this strategy might have introduced some variance because BOLD and spectral EEG responses can vary from day to day (Buck et al., 2008; Fernandez et al., 1993; Menz et al., 2006; Napflin et al., 2007). We opted to do so because we assume that the better data quality of our 63-channel active electrode scalp recordings compared to EEG data measured in our MR scanner settings outweigh the potentially better comparative basis offered by a synchronous recording since it is generally assumed that repeated measures of EEG and BOLD signals, i.e., averaging across multiple trials, during motor tasks provide a valid basis for generalizable results.
Dynamic causal modelling (DCM) has extended the understanding of brain network dynamics in a variety of functional systems. In the motor system, DCM studies based on functional magnetic resonance imaging (fMRI) or on magneto-/electroencephalography (M/EEG) have demonstrated movement-related causal information flow from secondary to primary motor areas and have provided evidence for nonlinear cross-frequency interactions among motor areas. The present study sought to investigate to what extent fMRI- and EEG-based DCM might provide complementary and synergistic insights into neuronal network dynamics. Both modalities share principal similarities in the formulation of the DCM. Thus, we hypothesized that DCM based on induced EEG responses (DCM–IR) and on fMRI would reveal congruent task-dependent network dynamics.
Brain electrical (63-channel surface EEG) and Blood Oxygenation Level Dependent (BOLD) signals were recorded in separate sessions from 14 healthy participants performing simple isometric right and left hand grips. DCM–IR and DCM–fMRI were used to estimate coupling parameters modulated by right and left hand grips within a core motor network of six regions comprising bilateral primary motor cortex (M1), ventral premotor cortex (PMv) and supplementary motor area (SMA). We found that DCM–fMRI and DCM–IR similarly revealed significant grip-related increases in facilitatory coupling between SMA and M1 contralateral to the active hand. A grip-dependent interhemispheric reciprocal inhibition between M1 bilaterally was only revealed by DCM–fMRI but not by DCM–IR. Frequency-resolved coupling analysis showed that the information flow from contralateral SMA to M1 was predominantly a linear alpha-to-alpha (9–13 Hz) interaction. We also detected some cross-frequency coupling from SMA to contralateral M1, i.e., between lower beta (14–21 Hz) at the SMA and higher beta (22–30 Hz) at M1 during right hand grip and between alpha (9–13 Hz) at SMA and lower beta (14–21 Hz) at M1 during left hand grip.
In conclusion, the strategy of informing EEG source-space configurations with fMRI-derived coordinates, cross-validating basic connectivity maps and analysing frequency coding allows for deeper insight into the motor network architecture of the human brain. The present results provide evidence for the robustness of non-invasively measured causal information flow from secondary motor areas such as SMA towards M1 and further contribute to the validation of the methodological approach of multimodal DCM to explore human network dynamics.
Cerebral mechanisms for different second language writing systems
2013, Neuropsychologia
Citation Excerpt :
Subsequently, to examine pairwise differences between the groups for each word condition and differences between the word conditions within each group, the % BOLD signal changes in the significant clusters (as determined by FSL Featquery) were further investigated via unpaired and paired t-tests, respectively. Note that the 90th percentile measure of % BOLD signal change was utilized, rather than the mean or maximum value, because it is considered a more representative measure of a typically active voxel within the ROI (Buck, Singhal, Arora, Schlitt, & Constable, 2008). Table 1 gives a summary of the demographic and cognitive measures (performed outside the scanner) for each group.
In this fMRI study, we examined the cerebral processing associated with second language (L2) reading in different writing systems in late L2 learners. To examine the impacts of cross-linguistic differences between the first language (L1) and L2 on learning to read in L2, we employed a bidirectional approach and compared brain activation during single word processing in two groups of late L2 readers: (1) L2 readers of English whose L1 was Japanese (Japanese-L1/English-L2) and (2) L2 readers of Japanese (of syllabic Kana only) whose L1 was English (English-L1/Japanese-L2). During English reading, the L2 readers of English (Japanese-L1/English-L2) exhibited stronger activation in the left superior parietal lobule/supramarginal gyrus, relative to the L1 readers of English (English-L1/Japanese-L2). This is a region considered to be involved in phonological processing. The increased activation in the Japanese-L1/English-L2 group likely reflects the increased cognitive load associated with L2 English reading, possibly because L1 readers of Kana, which has an extremely regular orthography, may need to adjust to the greater phonological demands of the irregular L2 English orthography. In contrast, during Kana reading, the L2 readers of Japanese Kana (English-L1/Japanese-L2) exhibited stronger activation in the lingual gyrus in both the left and right hemispheres compared to the L1 readers of Kana (Japaese-L1/English-L2). This additional activation is likely to reflect the lower level of visual familiarity to the L2 symbols in the English-L1/Japanese-L2 group; Kana symbols are uniquely used only in Japan, whereas Roman alphabetic symbols are seen nearly everywhere. These findings, bolstered by significant relationships between the activation of the identified regions and cognitive competence, suggest that the cerebral mechanisms for L2 reading in late learners depends both on which language is their L1 and which language is to be learnt as their L2. Educational implications of these results are discussed.
Graph-theory based parcellation of functional subunits in the brain from resting-state fMRI data
2010, NeuroImage
Resting-state fMRI provides a method to examine the functional network of the brain under spontaneous fluctuations. A number of studies have proposed using resting-state BOLD data to parcellate the brain into functional subunits. In this work, we present two state-of-the-art graph-based partitioning approaches, and investigate their application to the problem of brain network segmentation using resting-state fMRI. The two approaches, the normalized cut (Ncut) and the modularity detection algorithm, are also compared to the Gaussian mixture model (GMM) approach. We show that the Ncut approach performs consistently better than the modularity detection approach, and it also outperforms the GMM approach for in vivo fMRI data. Resting-state fMRI data were acquired from 43 healthy subjects, and the Ncut algorithm was used to parcellate several different cortical regions of interest. The group-wise delineation of the functional subunits based on resting-state fMRI was highly consistent with the parcellation results from two task-based fMRI studies (one with 18 subjects and the other with 20 subjects). The findings suggest that whole-brain parcellation of the cortex using resting-state fMRI is feasible, and that the Ncut algorithm provides the appropriate technique for this task.
Within-limb somatotopy in primary motor cortex - revealed using fMRI
2010, Cortex
Citation Excerpt :
Furthermore, although POA calculation was based on anatomical atlas-based M1 and conservative Bonferroni correction (range: .00006–.000003 for individual subjects), we found high variability for POA, also noted by Lotze et al. (2000) and Fesl et al. (2008). Fesl et al. (2008) and Buck et al. (2008) have suggested that as long as the thresholding technique used is ‘corrected’ and the threshold level is stringent, the variability of POA (t-weighted COM) is less relevant. This happens because at higher stringency, the activation shrinks and less significant voxels disappear, reducing noise in the data.
Controversy exists regarding within-limb somatotopy, i.e., organization of representations of within-limb segments (e.g., elbow, fingers), in the primary motor cortex (M1). While some believe that within-limb representations overlap, others postulate somatotopic distinction. The purpose of our study was to aid the resolution of this controversy by exploring a) overlap between finger and elbow representations using high-resolution functional magnetic resonance imaging (fMRI), b) somatotopic differences between the two representations with and without the overlap, c) different location markers – center of mass (COM) and peak of activation (POA) and d) intensity differences between the two representations with and without the overlap.
Twenty-four subjects underwent fMRI during finger and elbow tracking for definition of activation-based finger and elbow representations and their overlapping parts. Differential analysis was used to generate parts of finger and elbow representations that did not overlap. Location measures, COM and POA and fMRI signal intensity for all parts of finger and elbow representations were recorded.
Finger and elbow representations overlap but still possess distinct somatotopic centers, i.e., finger is lateral to overlap, which is lateral to the elbow. When overlap is excluded, locations of finger and elbow representations become more distinct. COM shows distinction along x, y and z-axes, but POA only shows distinction along the x-axis. Overlap has the highest intensity, while non-overlapping finger and elbow representations have the lowest intensity during finger and elbow tracking, respectively.
Somatotopic gradients exist in M1 despite the overlap. This finding supports ‘functional somatotopy’, i.e., within-limb representations overlap for multi-joint coordination, yet possess discrete centers for individuated control. Such a flexible somatotopy might allow representational reorganization. COM could be marker for ‘location’ and POA within in a high-intensity zone could emphasize ‘reorganization’ following learning/disease. Also, fMRI intensity could be a marker of change of overlap or somatotopic distinction.

View all citing articles on Scopus

View full text

Detecting change in BOLD signal between sessions for atlas-based anatomical ROIs

Abstract

Introduction

Section snippets

Study design

Results

Discussion

Magn. Reson. Imaging

Neuroimage

NeuroImage

Cogn. Brain Res.

Cogn. Brain Res.

Magn. Reson. Imaging

NeuroImage

NeuroImage

Neuroimage

NeuroImage

NeuroImage

Brain Res. Cogn. Brain Res.

Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing

R. Stat. Soc., Ser. A Stat. Soc. Ser. B.