Introduction
Multivariate pattern analysis (MVPA) of functional MRI (fMRI) data has grown steadily since its beginnings in 2001(Haxby, 2012). Following Raizada and Kriegeskorte (2010), we illustrate the growth of the literature by showing the citation rate for several key MVPA papers in Fig.1. Interest in MVPA spans disciplines. Advances have arisen from synergistic interactions with the machine learning community, which has developed new methods for addressing fMRI datasets and questions, as seen in the proliferation of relevant articles (e.g. Cuingnet et al., 2011, Mitchell et al., 2004, Van De Ville and Lee, 2012) and dedicated conference workshops (e.g. the International Conference on Pattern Recognition, NIPS, Cosyne, etc.). Interest in the cognitive neuroscience applications of MVPA is just as great (e.g. Heinzle et al., 2012, Tong and Pratte, 2012, Yang et al., 2012). The growing popularity of MVPA within neuroimaging has been driven by multiple factors, including: a) suggestions that it provides greater sensitivity and specificity than mass-univariate analyses with generally complementary results (Haynes and Rees, 2005, Jimura and Poldrack, 2012, Kamitani and Tong, 2005); b) the possibility of designing tests to address hypotheses which cannot be addressed with mass-univariate methods (e.g. Knops et al., 2009, Quadflieg et al., 2011, Stokes et al., 2009); and c) the intuitive appeal of a method which incorporates the signal from multiple voxels at once.
Searchlight analysis (also called information mapping) is an MVPA method introduced as a technique for identifying locally informative areas with greater power and flexibility than mass-univariate analyses (Kriegeskorte and Bandettini, 2007a, Kriegeskorte et al., 2006). Searchlight approaches are relatively unique, in that they were developed specifically for fMRI analysis, addressing both the common localization goal (many fMRI studies aim to identify small brain areas) and the spatial structure of the BOLD signal (adjacent voxels tend to have similar activation timecourses). Searchlight analysis produces maps by measuring the information in small spherical subsets (“searchlights”) centered on every voxel; the map value for each voxel thus derives from the information present in its searchlight, not the voxel individually. Note that the word “information” is not used here in its formal sense (as in the field of information theory), but rather following its conventional use in the MVPA application literature. Specifically, we use the word “information” to indicate that the activity in a group of voxels varies consistently with experimental condition: a highly informative voxel cluster can be used to identify experimental condition more accurately than a weakly informative one.
Appealing aspects of searchlight analysis include its whole-brain approach (i.e., a priori region specification is not needed), the ability to pool over subject-specific activation patterns, and its minimization of the extremes of the curse of dimensionality associated with whole-brain MVPA (the “curse” refers to computational difficulties which can occur when there are more voxels than examples, see (Clarke et al., 2008, Jain et al., 2000); it is minimized in searchlight analysis since relatively few voxels are typically included in each searchlight). Additionally, searchlight analysis produces a whole-brain results map that is superficially similar in appearance to the whole-brain significance maps produced by more familiar mass-univariate analyses (based on the general linear model); thus, searchlight analysis results are potentially easier to interpret. These appealing aspects, plus promising early results, have led to a rapid increase in the number of studies using searchlight analyses (note the rapid rise in citations for Kriegeskorte et al., 2006 in Fig.1, particularly in the last few years). Its acceptance as a standard approach is reflected in its inclusion in recent MVPA review and methodology articles (e.g. Bandettini, 2009, Mourao-Miranda et al., 2006, Raizada and Kriegeskorte, 2010, Tong and Pratte, 2012), as well as in the most prominent MVPA software packages (BrainVoyager QX 2.0, the Princeton MVPA Toolbox, PyMVPA).
Reflecting its potential and appeal, variations of the searchlight technique have been developed. In the spatial domain, it has been extended to circular subsets on cortical surfaces (Chen et al., 2011, Oosterhof et al., 2010, Oosterhof et al., 2011), rather than the original volumetric spheres. Efforts have also been made to extend the technique to incorporate the temporal domain (Fogelson et al., 2011, Rao et al., 2011). The first searchlight analyses used the Mahalanobis distance as the similarity measure for information mapping, but a widely adopted variation is to use machine learning algorithms, often support vector machines (SVMs), instead (Haynes et al., 2007, Kriegeskorte and Bandettini, 2007b). In these approaches, generalization accuracy of the classifier is used as a proxy for information content. Group analysis is usually performed by combining individual subject's maps with a binomial or t-test at each voxel (with the null hypothesis that the group classification accuracy is at chance level), creating maps of voxels with significant searchlights. Here we primarily consider classification-based searchlight analysis, but much of the discussion applies regardless of the precise implementation.
Searchlight analysis is a powerful and attractive tool for understanding neuroimaging data. However, it has particular characteristics and limitations that can lead to serious interpretation errors in practice, and so we recommend that straightforward confirmatory and sensitivity tests (analogous to post-hoc tests after an ANOVA), such as the ones described here, be considered a standard part of the searchlight analysis procedure. In the following sections we describe two assumptions that often implicitly underlie the interpretation of searchlight analysis results. Unfortunately, as we illustrate, these assumptions do not always hold, and so may lead to distorted results. We then describe how confirmatory follow-up tests can be used to guard against particularly harmful distortions, using two hypotheses common in cognitive studies as illustrations. This manuscript is accompanied by Supplemental Information containing examples (with code) and technical details. Assumption1 Information is detected consistently.
A fundamental aspect of fMRI is that information is not distributed uniformly across voxels but rather has a three-dimensional structure: some groups of voxels (e.g. those corresponding to a specific anatomical region) are more informative for a particular task than other groups of the same size. Additionally, neuroimaging data contains information at multiple spatial frequencies (Kriegeskorte et al., 2010, Op de Beeck, 2010). For example, consider a cued finger-tapping task. The finger area of the primary motor cortex will be highly informative at a very small spatial frequency while the premotor and somatosensory cortices may be equally informative, but at a larger spatial frequency. The difference can be imagined as the size of box required to enclose the minimum set of voxels capable of task classification: a larger box is necessary to enclose the pattern in premotor or somatosensory cortices than to enclose the pattern in the primary motor cortex.
The distribution of information is relevant for searchlight analysis because interpretation of any particular map depends on whether the information can be detected equally across spatial frequencies. In a simulation designed with equal power in all spatial frequency bands, Kriegeskorte et al. (2006) showed that detection did not require a close match between the size of the searchlight and the informative area: a 4mm radius consistently performed well. When this finding holds, it simplifies searchlight analysis interpretation: the peak areas of the map are the most informative voxels. However, if information is not present and detected equally at all spatial frequencies, then searchlight analysis results will depend fairly strongly upon the searchlight size; moreover, no single searchlight radius will be universally optimal or sufficient.
Additionally, although the Mahalanobis distance may be consistently sensitive to information across spatial frequency bands (Kriegeskorte et al., 2006), this property does not hold for all information measures used with searchlight analysis, especially the linear SVM. Training a linear SVM algorithm results in a set of weights; its decision function is a weighted linear combination of the voxels (Norman et al., 2006). Two properties of the linear SVM are particularly relevant when used in searchlight analysis: (1) It is sometimes able to correctly classify when the searchlight contains a small minority of highly informative voxels (intermixed with a majority of uninformative voxels), and conversely, (2) It is sometimes able to correctly classify when the searchlight contains a large number of weakly informative voxels.
Since, as described above, linear SVMs are relatively resistant to the curse of dimensionality (Jain et al., 2000), they can sometimes classify a dataset accurately even when only a tiny minority of the voxels are informative. The degree to which this occurs varies depending on dataset properties, but it happens often enough to be relevant in practice. For instance, Supplemental Example 4 shows that introducing just five informative voxels from an actual fMRI dataset into a group of two hundred random (uninformative) voxels is sufficient to shift the median accuracy of an SVM from chance to 0.6. For an extreme example, a dataset containing a single highly informative voxel and 200 random voxels is accurately classified in Supplemental Example 5. Searchlight analysis generally includes fewer than 200 voxels in each searchlight, increasing the likelihood that searchlights containing a single or only a few informative voxels will be detected (see the “Detection of rare informative voxels” section of the Supplemental Information for further discussion).
This behavior can cause distortions in a searchlight map. To illustrate, suppose that a cluster of five highly informative voxels (capable of significant classification whenever included in a searchlight) is surrounded by hundreds of truly uninformative voxels. Any searchlight overlapping the five-voxel cluster will be significant, even if the majority of its voxels are uninformative. As a result, some voxels in the results map will be categorized as significant, not because they themselves are informative, but because they are at the center of a searchlight that contains the informative voxels. Fig.2 (Supplemental Example 7) gives examples of this occurrence in an actual fMRI dataset (see Supplemental Example 6 as well): for instance, the voxel in the lower-left corner (at coordinates 1, 1) changes its mapped classification accuracy from “uninformative” to “informative” when the starred (actually informative) voxel is moved, despite there being no change the properties of the (lower-left) voxel itself.
A second issue is that the number of voxels marked as informative in a searchlight map will tend to grow as the searchlight radius increases, even when the size of the truly informative cluster stays fixed (Fig.3), so long as the curse of dimensionality does not dominate; classifiers will vary in how many uninformative voxels can be added to the fixed informative cluster before performance declines. This phenomenon, which has been termed the “needle-in-the-haystack-effect”, was demonstrated as a formal proof in Viswanathan et al. (2012). As an extreme example, Viswanathan et al. (2012) showed how all 147,000 voxels of a simulated volume would be classified as “informative” in a 3 voxel radius searchlight map when the volume contained just 430 evenly distributed informative voxels.
Another property of linear SVMs relevant for their use in searchlight analysis is that they can pool weak biases across many voxels, with the result that it is possible for a group of voxels to be classified accurately while the individual voxels making up the group do not yield significant classification, either singly or as subsets. This information “pooling” is often a useful characteristic for fMRI data, which is sometimes structured as weak information present in a large number of voxels. However, it can be troublesome for searchlight analysis interpretation. For example, suppose that there is a large cluster of voxels, each with the same small bias (i.e. a uniformly weakly informative voxel cluster). Ten voxels from this cluster (a small searchlight) may not yield significant classification, but thirty voxels (a larger searchlight) could produce a weakly significant classification, and fifty voxels, a highly significant classification (Fig.4 and Supplementary Example 1). This can be thought of as a case of discontinuous detection of information: at the extreme, a voxel cluster can change from “uninformative” to “informative” upon the addition of a single voxel (Supplementary Examples 2 and 3).
Discontinuous detection makes it possible for groups of weakly informative voxels to be partially or entirely missed when mapping information. Continuing the example, with a searchlight encompassing fewer than 30 voxels, the cluster will be classified as uninformative because no single searchlight can include enough voxels to enable accurate classification (Fig.5a). Larger searchlights could detect the cluster, but only when the shape of the searchlight matches the shape of the cluster: a spherical searchlight could miss an elliptical cluster (Fig.5b). An additional complication comes from assigning each searchlight's accuracy to its center voxel: large, weakly informative clusters will appear smaller in the information map if the searchlight radius is less than the cluster diameter, since only searchlights fully overlapping the cluster will be significant (Fig.5c).
Prior reports in the literature have documented the failure of weakly informative areas to be detected in searchlight analysis, mirroring our experience that widespread, weakly informative areas are common in fMRI datasets (see also Gonzalez-Castillo et al., 2012). For example, Eger et al. (2009) found that searchlight analysis (linear SVM, 3-voxel radius) identified no ROI voxels as informative, despite significant classification when using the whole ROI. Likewise, Diedrichsen et al. (in press) report needing to expand their searchlight size to achieve adequate sensitivity in one experimental condition (increasing from 80 to 160 voxels, with regularized linear discriminant analysis as the classification algorithm). Assumption2 Spatial variation between subjects is small compared to the searchlight radius.
Most applications using searchlight analysis interpret results primarily based on group-level aggregation of single-subject information maps, even though strategies for constructing and interpreting these maps have not been fully explored. Methods for constructing group-level maps often parallel those used in mass-univariate analysis: a t-test (for average accuracy across individuals greater than chance) is conducted at every voxel independently, followed by multiple-comparisons correction (Kriegeskorte and Bandettini, 2007a). Alternatively, the individual maps are statistically thresholded and the group-level map is reported in terms of the proportion of subjects with a significant searchlight at each voxel (Pereira and Botvinick, 2011). Permutation-based tests have also been proposed (Kriegeskorte et al., 2006), with new techniques increasing their interpretability and computational tractability (Gaonkar and Davatzikos, 2012, Stelzer et al., 2013). Some authors perform the searchlight analysis in native space then normalize the individual maps to an atlas, while others normalize the images first and then perform the searchlight analysis in atlas space (both of which can introduce distortions). This proliferation of techniques reflects the importance placed on group information maps in cognitive neuroscience applications of MVPA, and also the lack of agreement regarding the best method for constructing them. All of these techniques rely on a common assumption, however: that spatial variation in the information maps between individuals is minimal compared to the searchlight radius. Group maps may be misleading if this does not hold.
Spatial variation between individuals is not a concern unique to searchlight analysis but a factor in all neuroimaging techniques. For example, smoothing is used during mass-univariate analysis to help reduce the impact of inter-individual variability. However, evaluating results when inter-individual variability is present is particularly complex in searchlight analysis because of distortions that can occur when constructing individual information maps, particularly distortions causing a mismatch between the actual informative voxels and their appearance in the searchlight map (such as those shown in Fig.3, Fig.5). Since all methods of constructing a group information map involve combining some version of the individual maps, distortions in the individual maps are carried to the group level, where their effects may be magnified.
For example, spatial variation in the location of an informative cluster between individuals may cause the cluster to be missed in the group-level map. In Fig.6a, weakly informative clusters overlap in the individual maps, but since the individual searchlight mapping detects only a minority of the informative voxels (as in Fig.5c), the individual information maps do not overlap at the group level (Fig.6b green area), and so the cluster is missing from the group information map.
At the opposite extreme, voxels that are uninformative in each individual when examined separately can be identified as being informative at the group level. To illustrate that this can occur, suppose half of the individuals have a cluster of highly informative voxels towards the left side of a ROI while the rest of the individuals have the same cluster of informative voxels, but shifted towards the right side (Fig.7a). The group-level information map will not identify the voxels corresponding to either cluster as informative but rather the voxels between the two clusters, because this is where the individual maps overlap (Fig.7b). While Fig.7 is a simple illustration contrived to show the problem, such an outcome can occur in many actual situations. Fig.8 (Supplemental Example 9) shows an occurrence in real fMRI data: The most informative voxel in the group information map (starred voxel at left) has the lowest average accuracy when the voxels are tested for classification in a univariate manner (i.e. as single voxels; Fig.8, right).
Cited by (153)
Machine learning of large-scale multimodal brain imaging data reveals neural correlates of hand preference
2022, NeuroImage
Lateralization is a fundamental characteristic of many behaviors and the organization of the brain, and atypical lateralization has been suggested to be linked to various brain-related disorders such as autism and schizophrenia. Right-handedness is one of the most prominent markers of human behavioural lateralization, yet its neurobiological basis remains to be determined. Here, we present a large-scale analysis of handedness, as measured by self-reported direction of hand preference, and its variability related to brain structural and functional organization in the UK Biobank (N=36,024). A multivariate machine learning approach with multi-modalities of brain imaging data was adopted, to reveal how well brain imaging features could predict individual's handedness (i.e., right-handedness vs. non-right-handedness) and further identify the top brain signatures that contributed to the prediction. Overall, the results showed a good prediction performance, with an area under the receiver operating characteristic curve (AUROC) score of up to 0.72, driven largely by resting-state functional measures. Virtual lesion analysis and large-scale decoding analysis suggested that the brain networks with the highest importance in the prediction showed functional relevance to hand movement and several higher-level cognitive functions including language, arithmetic, and social interaction. Genetic analyses of contributions of common DNA polymorphisms to the imaging-derived handedness prediction score showed a significant heritability (h2=7.55%, p <0.001) that was similar to and slightly higher than that for the behavioural measure itself (h2=6.74%, p <0.001). The genetic correlation between the two was high (rg=0.71), suggesting that the imaging-derived score could be used as a surrogate in genetic studies where the behavioural measure is not available. This large-scale study using multimodal brain imaging and multivariate machine learning has shed new light on the neural correlates of human handedness.
See AlsoEin Framework für die schnelle visuelle Bildsuche unter Verwendung von im Gehirn hervorgerufenen Reaktionen in einem einzigen VersuchDekodierung von Gehirnzuständen mittels Rückwärtskanteneliminierung und Graphkernen in fMRI-KonnektivitätsnetzwerkenDataset basics and concepts — PyMVPA 2.4.0 documentationLocal similarity of activity patterns during auditory and visual processing
2022, Neuroscience Letters
Citation Excerpt :
Where H(X) and H(Y) correspond to the entropy of the distributions of X and Y, respectively; H(X,Y) is the joint entropy. Using the MI approach, we used the same size seed pattern and the same coordinates as in the cross-correlation analysis, the same size of the cube as in the Sadoun et al.2020 [7,21], which is 10 voxels. We used the rotation radius of 5 voxels with respect to the center of the cube (corresponding to the size of the cube of 10 voxels).
Neuroimaging studies have shown that brain activity is variable and changes according to stimuli and the environmental context, reflecting brain coding or information representations at different processing levels. However, little is known about activity organization that reflects coding strategies. Here, we explored and compared two different coding approaches, spatial via cross-correlation and intensity-based coding using mutual information. Using two fMRI datasets and different seeds, we searched for the spatial and intensity-based similarities with the seeds in brain activity. Our results showed that, apart from the seed regions, significant regions detected by intensity-based similarity analysis differ completely from those found using cross-correlation. These findings may indicate that information shared through spatial coding differs from that transmitted via non-spatial coding processes. Our results suggest that brain coding is organized in several different ways to optimize information processing.
Pattern analysis of neuroimaging data reveals novel insights on threat learning and extinction in humans
2022, Neuroscience and Biobehavioral Reviews
Several decades of rodent neurobiology research have identified a network of brain regions that support Pavlovian threat conditioning and extinction, focused predominately on the amygdala, hippocampus, and medial prefrontal cortex (mPFC). Surprisingly, functional magnetic resonance imaging (fMRI) studies have shown inconsistent evidence for these regions while humans undergo threat conditioning and extinction. In this review, we suggest that translational neuroimaging efforts have been hindered by reliance on traditional univariate analysis of fMRI. Whereas univariate analyses average activity across voxels in a given region, multivariate pattern analyses (MVPA) leverage the information present in spatial patterns of activity. MVPA therefore provides a more sensitive analysis tool to translate rodent neurobiology to human neuroimaging. We review human fMRI studies using MVPA that successfully bridge rodent models of amygdala, hippocampus, and mPFC function during Pavlovian learning. We also highlight clinical applications of these information-sensitive multivariate analyses. In sum, we advocate that the field should consider adopting a variety of multivariate approaches to help bridge cutting-edge research on the neuroscience of threat and anxiety.
More complex than you might think: Neural representations of food reward value in obesity
2022, Appetite
(Video) FLSA + Common Wage and Hour Issues in BenefitsObesity reached pandemic proportions and weight-loss treatments are mostly ineffective.
The level of brain activity in the reward circuitry is proposed to be proportionate to the reward value of food stimuli, and stronger in people with obesity. However, empirical evidence is inconsistent. This may be due to the double-sided nature of high caloric palatable foods: at once highly palatable and high in calories (unhealthy).
This study hypothesizes that, viewing high caloric palatable foods, a hedonic attentional focus compared to a health and a neutral attentional focus elicits more activity in reward-related brain regions, mostly in people with obesity. Moreover, caloric content and food palatability can be decoded from multivoxel patterns of activity most accurately in people with obesity and in the corresponding attentional focus.
During one fMRI-session, attentional focus (hedonic, health, neutral) was manipulated using a one-back task with individually tailored food stimuli in 32 healthy-weight people and 29 people with obesity. Univariate analyses (p<0.05, FWE-corrected) showed that brain activity was not different for palatable vs. unpalatable foods, nor for high vs. low caloric foods. Instead, this was higher in the hedonic compared to the health and neutral attentional focus. Multivariate analyses (MVPA) (p<0.05, FDR-corrected) showed that palatability and caloric content could be decoded above chance level, independently of either BMI or attentional focus.
Thus, brain activity to visual food stimuli is neither proportionate to the reward value (palatability and/or caloric content), nor significantly moderated by BMI. Instead, it depends on people's attentional focus, and may reflect motivational salience. Furthermore, food palatability and caloric content are represented as patterns of brain activity, independently of BMI and attentional focus. So, food reward value is reflected in patterns, not levels, of brain activity.
Frontopolar activity carries feature information of novel stimuli during unconscious reweighting of selective attention
2022, Cortex
Adapting to novelty is essential for an organism's survival in an uncertain world. Neuroimaging evidence consistently links the anterior prefrontal, specifically the frontopolar cortex (FPC; BA10), to exploratory reweighting of attentional weights thereby underscoring the role of the FPC in responding to environmental changes that are often complex and may occur very rapidly. Here we report new evidence showing that the FPC serves a role in attentional reallocation even in the absence of conscious awareness. Both mass-univariate and multivariate pattern analyses of fMRI data revealed that the right FPC and other attention-related areas not only are sensitive to unaware changes in the relevant stimulus dimension, but also that unconsciously processed information of the novel stimulus was globally represented across these regions. Our results indicate that unconsciously processed information can reach a global level of representation outside the occipitotemporal cortex, and that the FPC is crucial for the reweighting of selection biases in the absence of visual awareness.
Fast construction of interpretable whole-brain decoders
2022, Cell Reports Methods
Researchers often seek to decode mental states from brain activity measured with functional MRI. Rigorous decoding requires the use of formal neural prediction models, which are likely to be the most accurate if they use the whole brain. However, the computational burden and lack of interpretability of off-the-shelf statistical methods can make whole-brain decoding challenging. Here, we propose a method to build whole-brain neural decoders that are both interpretable and computationally efficient. We extend the partial least squares algorithm to build a regularized model with variable selection that offers a unique “fit once, tune later” approach: users need to fit the model only once and can choose the best tuning parameters post hoc. We show in real data that our method scales well with increasing data size and yields interpretable predictors. The algorithm is publicly available in multiple languages in the hope that interpretable whole-brain predictors can be implemented more widely in neuroimaging research.
Recommended articles (6)
Research article
Resolving Ambiguities of MVPA Using Explicit Models of Representation
Trends in Cognitive Sciences, Volume 19, Issue 10, 2015, pp. 551-554
(Video) Marianne gives a talk on chronic pain and socioeconomic statusWe advocate a shift in emphasis within cognitive neuroscience from multivariate pattern analysis (MVPA) to the design and testing of explicit models of neural representation. With such models, it becomes possible to identify the specific representations encoded in patterns of brain activity and to map them across the brain.
Research article
GLMdenoise improves multivariate pattern analysis of fMRI data
NeuroImage, Volume 183, 2018, pp. 606-616
GLMdenoise is a denoising technique for task-based fMRI. In GLMdenoise, estimates of spatially correlated noise (which may be physiological, instrumental, motion-related, or neural in origin) are derived from the data and incorporated as nuisance regressors in a general linear model (GLM) analysis. We previously showed that GLMdenoise outperforms a variety of other denoising techniques in terms of cross-validation accuracy of GLM estimates (Kay etal., 2013a). However, the practical impact of denoising for experimental studies remains unclear. Here we examine whether and to what extent GLMdenoise improves sensitivity in the context of multivariate pattern analysis of fMRI data. On a large number of participants (31 participants across 4 experiments; 3 T, gradient-echo, spatial resolution 2–3.75 mm, temporal resolution 1.3–2 s, number of conditions 32–75), we perform representational similarity analysis (Kriegeskorte etal., 2008a) as well as pattern classification (Haxby etal., 2001). We find that GLMdenoise substantially improves replicability of representational dissimilarity matrices (RDMs) across independent splits of each participant's dataset (average RDM replicability increases from r = 0.46 to r = 0.61). Additionally, we find that GLMdenoise substantially improves pairwise classification accuracy (average classification accuracy increases from 79% correct to 84% correct). We show that GLMdenoise often improves and never degrades performance for individual participants and that GLMdenoise also improves across-participant consistency. We conclude that GLMdenoise is a useful tool that can be routinely used to maximize the amount of information extracted from fMRI activity patterns.
Research article
Full correlation matrix analysis (FCMA): An unbiased method for task-related functional connectivity
Journal of Neuroscience Methods, Volume 251, 2015, pp. 108-119
The analysis of brain imaging data often requires simplifying assumptions because exhaustive analyses are computationally intractable. Standard univariate and multivariate analyses of brain activity ignore interactions between regions and analyses of interactions (functional connectivity) reduce the computational challenge by using seed regions of interest or brain parcellations.
To meet this challenge, we developed full correlation matrix analysis (FCMA), which leverages and optimizes algorithms from parallel computing and machine learning to efficiently analyze the pairwise correlations of all voxels in the brain during different cognitive tasks, with the goal of identifying task-related interactions in an unbiased manner.
When applied to a localizer dataset on a small compute cluster, FCMA accelerated a naive, serial approach by four orders of magnitude, reducing running time from two years to one hour. In addition to this performance gain, FCMA emphasized different brain areas than existing methods. In particular, beyond replicating known category selectivity in visual cortex, FCMA also revealed a region of medial prefrontal cortex whose selectivity derived from differential patterns of functional connectivity across categories.
(Video) Lightstreams (PHT) | Crypto First-Look FundamentalsFor benchmarking, we started with a naive approach and progressively built up to the complete FCMA procedure by adding optimized classifier algorithms, multi-threaded parallelism, and multi-node parallelism. To evaluate what can be learned with FCMA, we compared it against multivariate pattern analysis of activity and seed-based analysis of functional connectivity.
FCMA demonstrates how advances in computer science can alleviate computational bottlenecks in neuroscience. We have released a software toolbox to help others evaluate FCMA.
Research article
The effect of spatial resolution on decoding accuracy in fMRI multivariate pattern analysis
NeuroImage, Volume 132, 2016, pp. 32-42
Multivariate pattern analysis (MVPA) in fMRI has been used to extract information from distributed cortical activation patterns, which may go undetected in conventional univariate analysis. However, little is known about the physical and physiological underpinnings of MVPA in fMRI as well as about the effect of spatial smoothing on its performance. Several studies have addressed these issues, but their investigation was limited to the visual cortex at 3T with conflicting results. Here, we used ultra-high field (7T) fMRI to investigate the effect of spatial resolution and smoothing on decoding of speech content (vowels) and speaker identity from auditory cortical responses. To that end, we acquired high-resolution (1.1mm isotropic) fMRI data and additionally reconstructed them at 2.2 and 3.3mm in-plane spatial resolutions from the original k-space data. Furthermore, the data at each resolution were spatially smoothed with different 3D Gaussian kernel sizes (i.e. no smoothing or 1.1, 2.2, 3.3, 4.4, or 8.8mm kernels). For all spatial resolutions and smoothing kernels, we demonstrate the feasibility of decoding speech content (vowel) and speaker identity at 7T using support vector machine (SVM) MVPA. In addition, we found that high spatial frequencies are informative for vowel decoding and that the relative contribution of high and low spatial frequencies is different across the two decoding tasks. Moderate smoothing (up to 2.2mm) improved the accuracies for both decoding of vowels and speakers, possibly due to reduction of noise (e.g. residual motion artifacts or instrument noise) while still preserving information at high spatial frequency. In summary, our results show that – even with the same stimuli and within the same brain areas – the optimal spatial resolution for MVPA in fMRI depends on the specific decoding task of interest.
Research article
A decade of decoding reward-related fMRI signals and where we go from here
NeuroImage, Volume 180, Part A, 2018, pp. 324-333
Information about potential rewards in the environment is essential for guiding adaptive behavior, and understanding neural reward processes may provide insights into neuropsychiatric dysfunctions. Over the past 10 years, multivoxel pattern analysis (MVPA) techniques have been used to study brain areas encoding information about expected and experienced outcomes. These studies have identified reward signals throughout the brain, including the striatum, medial prefrontal cortex, orbitofrontal cortex, dorsolateral prefrontal cortex, and parietal cortex. This review article discusses some of the assumptions and models that are used to interpret results from these studies, and how they relate to findings from animal electrophysiology. The article reviews and summarizes some of the key findings from MVPA studies on reward. In particular, it first focuses on studies that, in addition to mapping out the brain areas that process rewards, have provided novel insights into the coding mechanisms of value and reward. Then, it discusses examples of how multivariate imaging approaches are being used more recently to decode features of expected rewards that go beyond value, such as the identity of an expected outcome or the action required to obtain it. The study of such complex and multifaceted reward representations highlights the key advantage of using representational methods, which are uniquely able to reveal these signals and may narrow the gap between animal and human research. Applied in a clinical context, MVPA may advance our understanding of neuropsychiatric disorders and the development of novel treatment strategies.
Research article
Representational similarity encoding for fMRI: Pattern-based synthesis to predict brain activity using stimulus-model-similarities
NeuroImage, Volume 128, 2016, pp. 44-53
Patterns of neural activity are systematically elicited as the brain experiences categorical stimuli and a major challenge is to understand what these patterns represent. Two influential approaches, hitherto treated as separate analyses, have targeted this problem by using model-representations of stimuli to interpret the corresponding neural activity patterns. Stimulus-model-based-encoding synthesizes neural activity patterns by first training weights to map between stimulus-model features and voxels. This allows novel model-stimuli to be mapped into voxel space, and hence the strength of the model to be assessed by comparing predicted against observed neural activity. Representational Similarity Analysis (RSA) assesses models by testing how well the grand structure of pattern-similarities measured between all pairs of model-stimuli aligns with the same structure computed from neural activity patterns. RSA does not require model fitting, but also does not allow synthesis of neural activity patterns, thereby limiting its applicability. We introduce a new approach, representational similarity-encoding, that builds on the strengths of RSA and robustly enables stimulus-model-based neural encoding without model fitting. The approach therefore sidesteps problems associated with overfitting that notoriously confront any approach requiring parameter estimation (and is consequently low cost computationally), and importantly enables encoding analyses to be incorporated within the wider Representational Similarity Analysis framework. We illustrate this new approach by using it to synthesize and decode fMRI patterns representing the meanings of words, and discuss its potential biological relevance to encoding in semantic memory. Our new similarity-based encoding approach unites the two previously disparate methods of encoding models and RSA, capturing the strengths of both, and enabling similarity-based synthesis of predicted fMRI patterns.
(Video) Think Science: Artificial Intelligence In The 21st Century
Copyright © 2013 Published by Elsevier Inc.
FAQs
What is a SearchLight analysis? ›
SearchLight analysis was introduced in [Kriegeskorte et al. [3]], and consists of scanning the brain with a searchlight. Briefly, a ball of given radius is scanned across the brain volume and the prediction accuracy of a classifier trained on the corresponding voxels is measured.
What does multi-voxel pattern analysis rely on? ›Multi-voxel pattern analysis (MVPA) involves searching for highly reproducible spatial patterns of activity that differentiate across experimental conditions.
What is multivariate pattern analysis? ›MVPA refers to a set of methods that analyze neural responses as patterns of activity, thus affording investigation of the varying brain states that a cortical field or system can produce, thus increasing the amount of information that can be decoded from brain activity, in contrast to simpler univariate measures that ...
What is the purpose of searchlight? ›searchlight, high-intensity electric light with a reflector shaped to concentrate the beam, used to illuminate or search for distant objects or as a beacon.
Why is it called searchlight? ›Admission is free.
Founded in 1898, Searchlight takes its name from the Searchlight Mine.
Voxel size and detail in MR images is determined by the values selected for the three protocol factors: FOV, matrix size, and slice thickness. The dimension of a voxel in the plan of the image is determined by the ratio of the field of view (FOV) and the size of the matrix.
What are the limitations of fMRI? ›Limitations of fMRI
It was stated that fMRI does not measure neuron activity directly but rather does so indirectly by measuring blood flow changes to determine active areas.
Multi-voxel Chemical Shift Imaging (CSI) techniques offer two potential advantages over SVS: 1) a larger total coverage area (since the size of the entire multivoxel slab is greater), and 2) higher spatial resolution (since the individual voxels are smaller).
What are the 3 categories of multivariate analysis? ›Multiple linear regression. Multiple logistic regression. Multivariate analysis of variance (MANOVA) Factor analysis.
Which are the two most common multivariate analysis methods? ›Principal component analysis (PCA) and Factor analysis are two of the common techniques used to perform such a dimension reduction.
What is the most common multivariate analysis? ›
Multiple Regression Analysis
Multiple regression is the most commonly utilized multivariate technique. It examines the relationship between a single metric dependent variable and two or more metric independent variables.
Concave mirrors are used in search lights and torches so that we have a more focused light beam which will not diverge out and hence help in searching.
Are searchlights still used? ›Non-military use
Today, searchlights are used in advertising, fairs, festivals and other public events. Their use was once common for movie premieres; the waving searchlight beams can still be seen as a design element in the logos of 20th Century Studios and the Fox television network.
(b) A concave mirror is used as a search-light reflector to obtain a parallel beam of light.
What is the other term for searchlight? ›synonyms for searchlight
On this page you'll find 24 synonyms, antonyms, and words related to searchlight, such as: flashlight, light, beacon, gaslight, torch, and gas lamp.
On January 17, 2020, it was announced that the "Fox" name would be dropped from several of the Fox assets that were acquired by Disney, shortening the company's name to "Searchlight Pictures", in order to` avoid brand confusion with Fox Corporation.
What is the difference between spotlight and searchlight? ›Re: Searchlight vs Spotlight
IMHO a "spotlight" is a narrow focus beam only and a "searchlight" is a hard mounted moveable light.
A voxel is a cube within a 3D model with a position in a 3D grid and a single color value, whereas a pixel is a square inside a 2D image with a location in a 2D grid and a single color value. The fact that pixels are preserved in picture formats like PNG and JPG is another important factor.
What is the difference between pixel size and voxel size? ›A pixel represents the smallest sampled 2D element in an image. It has dimensions given along two axes in mm, dictating in-plane spatial resolution. Pixel sizes range in clinical MRI from mm (e.g., 1 × 1 mm2) to sub-mm. A voxel is the volume element, defined in three dimensions.
What is a good voxel size? ›A typical voxel size for a structural MRI is 1 mm x 1mm x 1mm; for fMRI, 3 mm x 3 mm x 3 mm. However, these can vary substantially from study to study.
What is the main weakness of fMRI and why? ›
Strengths and weaknesses of fMRI
Because BOLD contrast derives from the sluggish hemodynamic response to metabolic changes, a significant weakness is its low temporal resolution.
Drawbacks of MRI scans include their much higher cost, and patient discomfort with the procedure. The MRI scanner subjects the patient to such powerful electromagnets that the scan room must be shielded.
What are potential limitations of brain imaging techniques? ›Repetition of a high quality test may not produce any findings but increases the financial and psychological strain on the patient. Finally, a major limitation of neuroimaging lies in its inability to detect reliably all neuro-ophthalmic disorders of intracranial and intra- orbital etiology.
What is the smallest voxel size? ›The small field of view units may use a small voxel size of 0.076 mm, which enables visualization of very small changes to structures. Other voxel sizes available for CBCT units are variable, such as 0.2 mm, 0.3 mm, and 0.4 mm.
What is the minimum voxel size? ›This is a single value which results in the generation of equal volume, cubic, isotropic voxels. The minimum allowable setting is 0.10 cm and the resolution for incremental increases from this value is 0.01 cm.
Is voxel 2D or 3D? ›Voxel is an image of a three-dimensional space region limited by given sizes, which has its own nodal point coordinates in an accepted coordinate system, its own form, its own state parameter that indicates its belonging to some modeled object, and has properties of modeled region.
What is multivariate voxel-based lesion symptom mapping? ›This voxel‐based lesion‐symptom mapping (VLSM) technique produces a statistical map showing the strength of the relationship between damage at any given voxel and performance on a behavioral measure of interest across a group of individuals with brain lesions (Bates et al., 2003).
What is Mvpa in neuroscience? ›Multi-Voxel Pattern Analysis (MVPA)
What is voxel analysis? ›Voxel-based analysis involves registration of diffusion maps into a standard space (i.e., normalization) to achieve accurate registration between subjects across voxels and subsequently across anatomical structures.
What is a voxel and what does it measure in the brain? ›Each pixel corresponds to a three-dimensional square or rectangular chunk of brain tissue called a volume element (or voxel). Each pixel of the image is typically assigned either a level of visual grayness ranging from black to white (Fig. 1) or an arbitrary color that represents a numerical value.
Does MRI use voxels? ›
An MRI image is composed of a number of voxels; the voxel size is the spatial resolution of the image. A voxel is a 3D unit of the image with a single value, just as for digital photographs a pixel is a 2D unit of the image with a single value.
What is the purpose of a voxel? ›A voxel layer represents multidimensional spatial and temporal information in a 3D volumetric visualization. For example, you can visualize atmospheric or oceanic data, a geological underground model, or space-time cubes as voxel layers.
How many voxels are in an MRI scan? ›Typically, fMRI machines divide the brain into three-dimensional pixels called voxels, each about five cubic millimeters in size. The complete activity of the brain at any instant can be recorded using a three-dimensional grid of 60 x 60 x 30 voxels.
How do you detect MVPA? ›HOW DO I KNOW IF MY ACTIVITIES ARE OF MODERATE TO VIGOROUS INTENSITY? To gauge your level of intensity, you can simply measure your heart rate during your activities. Heart rate is measured in beats per minute (bpm) and can vary greatly from person to person depending on factors like age and fitness level.
What are some examples of Mvpa? ›- Hiking.
- Jogging at 6 mph.
- Shoveling.
- Carrying heavy loads.
- Bicycling fast (14-16 mph)
- Basketball game.
- Soccer game.
- Tennis singles.
Exercises should involve motor skills (balance, agility, coordination and gait), proprioceptive exercise training, and multifaceted activities (yoga) to improve physical function and prevent falls in older adults.