CommentTen simple rules for reporting voxel-based morphometry studies
Introduction
Voxel-based morphometry (Ashburner and Friston, 2000, Mechelli et al., 2005) is becoming increasingly widely used as a tool to examine patterns of brain change in healthy aging (Good et al., 2001) or neurodegenerative disease (Baron et al., 2001) and neuroanatomical correlates of behavioural or cognitive deficits (Abell et al., 1999). VBM essentially involves voxel-wise statistical analysis of pre-processed structural MR images. Although much of the processing and analysis is automated in software packages such as SPM,1 many methodological decisions remain, including what template to use for normalisation, what level and type of correction to use and how best to display results. Different approaches, such as VBM using RAVENS maps (Davatzikos et al., 2001), introduce yet more options. It can therefore be difficult to replicate or draw conclusions from VBM studies if the processing steps are not clearly described. Similarly, if unusual methods or parameters are employed without sufficient justification, it can be challenging for readers to judge the potential impact on results or to compare the work with other studies. In light of these issues, this editorial presents a set of recommendations, in the form of ten “rules” accompanied by a checklist, which we hope will be helpful to authors when writing up VBM studies. The rules are intended to outline core principles that should be followed and information that should be included when reporting a VBM study in order to make it transparent, replicable and useful. Since the field is rapidly developing, such rules must not be overly restrictive; therefore in some instances, where a clear protocol cannot be stated, general advice is given in the hope of aiding the reader to follow good practice. As VBM data sets accumulate and alternative procedures and techniques proliferate, we feel that guidelines are crucial for clear scientific communication and further development of the field. Additional motivation for this work came from a related effort in the field of functional brain imaging (“Guidelines for reporting an fMRI study”, Poldrack et al., in press).2
Section snippets
Set out the rationale for your study and describe the data fully
What are the key experimental questions, and why was VBM preferred over other techniques in order to address these questions? Prior hypotheses should be stated; either experimental ones or a priori anatomical or spatial regions in which effects might be expected (Maguire et al., 2000). This is particularly important if search volumes are restricted when correcting for multiple statistical tests during data analysis (see Rule 5). The study design should be described in enough detail for readers
Explain how the brain segmentations are produced
The inputs to VBM's statistical analysis are derived from structural MR images using tissue segmentation, spatial normalisation and smoothing. Additional pre-processing is often performed before the main segmentation step, generally using automatic algorithms such as MR bias correction or skull stripping, or manual techniques such as semi-automatic brain segmentation or interactive reorientation. Multiple processes may be combined within unified algorithms, such as that of Ashburner and Friston
Describe the method of inter-subject spatial normalisation
In order to compare different subjects, it is essential to use some kind of registration algorithm to bring the images into at least approximate correspondence. Both the technique used and the reference space to which brains are aligned can impact on the results (Senjem et al., 2005), so clear reporting is crucial. As with the other pre-processing steps (see Rule 2), if a popular software package is used, deviations from the default options should be highlighted. If a non-standard approach is
Make your statistical design transparent
There are two issues here, model specification and contrast testing. When constructing a model, it is important to be clear about which variables are included, and why. In the case of factorial designs, it should be obvious to the reader exactly what the factors were, the levels of each factor, and which interactions between factors were modelled. With estimation methods more advanced than ordinary least squares, it may be necessary to report extra information; for example, SPM5 includes
Be clear about the significance of your findings
As with other mass-univariate image analysis techniques, a large number of statistical tests are performed in a VBM study. The method used to correct for multiple testing should be both clearly stated and carefully considered—ideally, a priori. VBM is often performed on limited numbers of subjects (for example, to investigate rare disorders), when there is a temptation to report uncorrected results due to low statistical power. If this is done, it should be made obvious and it is probably best
Present results unambiguously
The type and level of correction should be stated in all figure and table legends, and if the statistical parametric map (SPM) is displayed as orthogonal slices or sections then coordinates should be given. It is helpful to present tables that include statistic values and cluster sizes, as well as coordinates of local maxima. SPMs should be displayed on a template that represents some form of average anatomy, for example, the MNI T1 template often used for normalisation, or ideally, a
Clarify and justify any non-standard statistical analyses
As a general principle, the less standard the analysis, the more thoroughly it should be explained. Here, we discuss three of the more common examples. Contrast masking may be used to disambiguate multiple possible causes of an effect or to define smaller search regions, in which case authors should clarify not only which contrasts were analysed, which were used for masking and at what threshold, but also the motivation for doing so and their interpretation. If a conjunction of analyses is
Guard against common pitfalls
Here we discuss a few potential problems with VBM analyses that might be easily overlooked. Firstly, note that while voxel-wise multiple testing is usually corrected for (see Rule 5), most software packages do nothing to correct for the user's investigation of multiple contrasts—the more conventional multiple-comparison problem (Hochberg and Tamhane, 1987). A simple example of this occurs if two opposite single-tailed t-contrasts are analysed: if findings in either contrast could be considered
Recognise the limitations of the technique
Like all image analysis methods, VBM has inherent limitations (Bookstein, 2001). The basic premise of inter-subject spatial normalisation is problematic: different subjects can have different gyral variants with no “true” correspondence between them and information from structural MRI (even manual sulcal labelling) does not necessarily predict underlying cytoarchitectonic borders (Amunts et al., 2007). Normalisation accuracy is also likely to vary between brain regions, for example, highly
Interpret your results cautiously and in context
When implemented rigorously and interpreted carefully, VBM can be a powerful technique. Authors should be forthright in discussing potential sources of bias or imprecision, whether they arise from the study's design or analysis, or from the nature of VBM itself. Particular care should be taken when interpreting results which appear fragile with respect to more arbitrary aspects of the method such as pre-processing options and nuisance variables. A conservative approach based on robust findings,
Acknowledgments
We are grateful to John Ashburner for his helpful comments and to Karl Friston for inviting the submission of this paper for review. We thank the authors of “Guidelines for reporting an fMRI study” for sharing an early draft of their work. Chris Frost provided helpful statistical advice. We are grateful to the reviewers, who made numerous detailed suggestions.
GRR is supported by the Engineering and Physical Sciences Research Council and GlaxoSmithKline through an Industrial CASE Studentship.
References (38)
- et al.
Cytoarchitecture of the cerebral cortex—more than localization
NeuroImage
(2007) A fast diffeomorphic image registration algorithm
NeuroImage
(2007)- et al.
Unified segmentation
NeuroImage
(2005) - et al.
Why voxel-based morphometry should be used
NeuroImage
(2001) - et al.
Voxel-based morphometry—the methods
NeuroImage
(2000) - et al.
Multimodal image coregistration and partitioning—a unified framework
NeuroImage
(1997) - et al.
In vivo mapping of gray matter loss with voxel-based morphometry in mild Alzheimer's disease
NeuroImage
(2001) “Voxel-based morphometry” should not be used with imperfectly registered images
NeuroImage
(2001)- et al.
Zen and the art of medical image registration: correspondence, homology, and quality
NeuroImage
(2003) - et al.
In praise of tedious anatomy
NeuroImage
(2007)
Voxel-based morphometry using the RAVENS maps: methods and validation using simulated longitudinal atrophy
NeuroImage
Distribution of grey matter atrophy in Huntington's disease patients: a combined ROI-based and voxel-based morphometric study
NeuroImage
Conjunction revisited
NeuroImage
Thresholding of statistical maps in functional neuroimaging using the false discovery rate
NeuroImage
A voxel-based morphometric study of ageing in 465 normal adult human brains
NeuroImage
The effect of filter size on VBM analyses of DT-MRI data
NeuroImage
Unbiased diffeomorphic atlas construction for computational anatomy
NeuroImage
Comparison of standard and optimized voxel-based morphometry for analysis of brain changes associated with temporal lobe epilepsy
NeuroImage
Acquisition-related morphological variability in structural MRI
Acad. Radiol.
Cited by (219)
Contributions of Polygenic Risk and Disease Status to Gray Matter Abnormalities in Major Depression
2024, Biological Psychiatry: Cognitive Neuroscience and NeuroimagingHyperacusis is associated with smaller gray matter volumes in the supplementary motor area
2023, NeuroImage: ClinicalThe cerebellum gains weight: A systematic review of alterations in cerebellar volume and cerebro-cerebellar functional alterations in individuals with eating disorders
2022, Neuroscience and Biobehavioral ReviewsMolecular biomarkers correlate with brain grey and white matter changes in patients with mitochondrial m.3243A > G mutation
2022, Molecular Genetics and MetabolismGray matter structural plasticity in patients with basal ganglia germ cell tumors: A voxel-based morphometry study
2022, Magnetic Resonance ImagingCitation Excerpt :Preprocessing began with the conversion of files from DICOM (Digital Imaging and Communications in Medicine) to NIfTI (Neuroimaging Informatics Technology Initiative) format using dcm2niigui. Next, a correlation was implemented using the Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) algorithm to increase image accuracy and generate native and DARTEL versions for GM and white matter and native versions for CSF [15]. During this step, DARTEL produces a GM template for further normalization of each GM subject (the data quality control report is shown in Fig. S1), which increase the accuracy of inter-subject alignment by modeling the shape of each brain according to millions of different parameters [16].