Fine-grain atlases of functional modes for fMRI analysis
Kamalaker Dadi, Gaël Varoquaux, Antonia Machlouzarides-Shalit, Krzysztof J. Gorgolewski, Demian Wassermann, Bertrand Thirion, Arthur Mensch
FFine-grain atlases of functional modes for fMRI analysis
Kamalaker Dadi a , Ga¨el Varoquaux a , Antonia Machlouzarides-Shalit a , Krzysztof J. Gorgolewski c , DemianWassermann a , Bertrand Thirion a , Arthur Mensch a,b a Inria, CEA, Universit´e Paris-Saclay, Palaiseau, 91120, France b ENS, DMA, 45 rue d’Ulm, 75005 Paris c Department of Psychology, Stanford University, California, USA
Abstract
Population imaging markedly increased the size of functional-imaging datasets, shedding new light on the neural basis ofinter-individual differences. Analyzing these large data entails new scalability challenges, computational and statistical.For this reason, brain images are typically summarized in a few signals, for instance reducing voxel-level measures withbrain atlases or functional modes. A good choice of the corresponding brain networks is important, as most data analysesstart from these reduced signals. We contribute finely-resolved atlases of functional modes, comprising from 64 to 1024networks. These dictionaries of functional modes (DiFuMo) are trained on millions of fMRI functional brain volumesof total size 2.4TB, spanned over 27 studies and many research groups. We demonstrate the benefits of extractingreduced signals on our fine-grain atlases for many classic functional data analysis pipelines: stimuli decoding from 12,334brain responses, standard GLM analysis of fMRI across sessions and individuals, extraction of resting-state functional-connectomes biomarkers for 2,500 individuals, data compression and meta-analysis over more than 15,000 statisticalmaps. In each of these analysis scenarii, we compare the performance of our functional atlases with that of other popularreferences, and to a simple voxel-level analysis. Results highlight the importance of using high-dimensional “soft”functional atlases, to represent and analyse brain activity while capturing its functional gradients. Analyses on high-dimensional modes achieve similar statistical performance as at the voxel level, but with much reduced computationalcost and higher interpretability. In addition to making them available, we provide meaningful names for these modes,based on their anatomical location. It will facilitate reporting of results.
Keywords:
Brain imaging atlases; Functional networks; Functional parcellations; Multi-resolution;
1. Introduction
Population imaging has been bringing in terabytes ofhigh-resolution functional brain images, uncovering theneural basis of individual differences (Elliott et al., 2008).While these great volumes of data enable fitting richer sta-tistical models, they also entail massive data storage (Pol-drack et al., 2013; Gorgolewski et al., 2017) and challeng-ing high-dimensional data analysis. A popular approachto facilitate data handling is to work with image-derivedphenotypes (IDPs), i.e. low-dimensional signals that sum-marize the information in the images while keeping mean-ingful representations of the brain (Miller et al., 2016).While brain atlases originated in characterizing thebrain’s microstructure (Brodmann, 1909), today they arewidely used to study functional connectomes (Spornset al., 2005; Varoquaux and Craddock, 2013) and for datareduction in functional imaging (Thirion et al., 2006; Crad-dock et al., 2012). For these applications, the choice ofbrain regions conditions the signal captured in the dataanalysis. To define regions well suited to brain-imagingendeavors, there is great progress in building atlases fromthe neuroimaging data itself (Eickhoff et al., 2018). Yet,most functional atlases describe the brain as parcellations, locally-uniform functional units, and thus do not representwell functional gradients (Huntenburg et al., 2018).For functional imaging, brain structures delineated byan atlas should capture the main features of the functionalsignal, e.g. the functional networks (Smith et al., 2011). Ina nutshell, there are two approaches to define well-suitedstructures. These can strive to select homogenous neuralpopulations, typically via clustering approaches (Goutteet al., 1999; Bellec et al., 2010; Craddock et al., 2012;Thirion et al., 2014; Schaefer et al., 2017). They can alsobe defined via continuous modes that map intrinsic brainfunctional networks (Damoiseaux et al., 2006; Varoquauxet al., 2011; Harrison et al., 2015). These functional modeshave been shown to capture well functional connectivity,with techniques such as Independent Component Analysis(Kiviniemi et al., 2009; Pervaiz et al., 2019) or sparse dic-tionary learning (Mensch et al., 2016b; Dadi et al., 2019).High-resolution atlases can give a fine-grained divi-sion of the brain and capture more functionally-specificregions and rich descriptions of brain activity (Schaeferet al., 2017). Yet, there is to date no highly-resolved setof “soft” functional modes available, presumably becauseincreasing the dimensionality raises significant computa- a r X i v : . [ q - b i o . N C ] M a r ional and statistical challenges (Mensch et al., 2016a; Per-vaiz et al., 2019). In this paper, we address this need withhigh-order dictionaries of functional modes (DiFuMo) ex-tracted at a large scale both in terms of data size (3 millionvolumes of total data size 2.4TB) and resolution (up to1024 modes). For this, we leverage the wealth of openly-available functional images (Poldrack et al., 2013) and ef-ficient dictionary-learning algorithms to fit on large data.This is unlike ICA which is hard to use for a high numberof modes (Pervaiz et al., 2019). Contributions.
We provide Dictionaries of FunctionalModes “ DiFuMo ” that can serve as atlases to extractfunctional signals, e.g. provide IDPs, with different di-mensionalities (64, 128, 256, 512, and 1024). These modesare optimized to represent BOLD data well, over a widerange of experimental conditions. They are more finely-resolved than existing brain decompositions with contin-uous networks. By providing validated fine functional at-lases, our goal is to streamline fMRI analysis with reducedrepresentations, to facilitate large-cohort and inter-studieswork. Through thorough benchmarking over classic dataanalysis tasks, we show that these modes gives IDPs thatground better analysis of functional images. Finally, weprovide a meaningful label to each mode, summarizing itsanatomical location, to facilitate reporting of results.
2. Methods: data-driven fine-grain functionalmodes
We describe in this section the models and methods un-derlying our definition of brain structures to extract IDPs.
While analysis of brain images has been pioneered atthe voxel level (Friston et al., 1995), image-derived pheno-types (IDP) are increasingly used in the context of pop-ulation imaging. Trading voxel-level signals for IDPs hasseveral motivations. First and foremost, it greatly facili-tates the analysis on large cohorts: the data are smaller,easier to share, requiring less disk storage, computer mem-ory, and computing power to analyze. It can also comewith statistical benefits. For instance, in standard analysisof task responses, e.g. in mass-univariate brain mapping,the statistical power of hypothesis test at the voxel levelis limited by multiple comparisons (Friston et al., 1995),while working at the level of IDPs mitigates this prob-lem (Thirion et al., 2006). For predictive modeling, e.g. inmulti-variate decoding (Mour˜ao-Miranda et al., 2005), thehigh-dimensionality of the signals is a challenge to learningmodels that generalize well—a phenomenon known as thecurse of dimensionality in machine learning (Hastie et al.,2009). Finally, for functional connectomes, working atvoxel-level is computationally and statistically intractable https://parietal-inria.github.io/DiFuMo as it entails modeling billions of connections. The stan-dard approach is therefore to average signals on regions ornetworks (Varoquaux and Craddock, 2013).Functional neuroimaging is currently largely dependenton neuroanatomy for mapping function to structure (De-strieux et al., 2010; Devlin and Poldrack, 2007). Someanatomical structures support well a direct mapping tospecific functions (Brett et al., 2002; Rademacher et al.,1993), e.g. the primary visual areas. Yet other functionalunits are not simply defined from anatomical features, forinstance in high-level regions such as the default mode,which is defined from functional data (Leech et al., 2011;Greicius et al., 2003). Compared to anatomical atlases, defining regions fromthe functional signal can lead to a better explanation ofbehavioral outcomes (Dadi et al., 2019), as they capturethe functional structure of the brain. Clustering of fMRItimeseries has been heavily used to define brain parcella-tions (Goutte et al., 1999), or for data reduction in pre-dictive models (Michel et al., 2012). Reference functionalbrain parcellations have been defined with various cluster-ing algorithms on resting-state fMRI (Bellec et al., 2010;Yeo et al., 2011; Craddock et al., 2012). Another classof approaches seeks modes of brain activity, decomposingthe signal as a product of spatial maps and correspond-ing time-series (Figure 1). The most popular model inneuroimaging is independent component analysis (ICA,Hyv¨arinen and Oja, 2000), which optimizes spatial inde-pendence between extracted maps. It has been extensivelyused to define resting-state networks (Kiviniemi et al.,2003; Beckmann et al., 2005; Calhoun et al., 2001) andimplicitly outlines soft parcellations of the brain at highorder (Kiviniemi et al., 2009; Varoquaux et al., 2010b).ICA-defined networks are used to extract the official IDPsof UK BioBank, the largest brain-imaging cohort to date;these have been shown to relate to behavior (Miller et al.,2016).We rely on another decomposition model, dictionarylearning (Olshausen and Field, 1997), that enforces spar-sity and non-negativity instead of independence on the V o x e l s Time = k spatial maps Time x Figure 1:
Linear decomposition model of fMRI time-seriesfor estimating brain networks : The fMRI time series X are fac-torized into a product of two matrices, D wich contain spatial modesand A temporal loadings of each mode. p - number of features, n -number of volumes in fMRI image, k - number of dictionaries. Rest and task fMRI.
Most functional brain atlases havebeen extracted from rest fMRI (Bellec et al., 2010; Poweret al., 2011; Craddock et al., 2012; Yeo et al., 2011; Milleret al., 2016; Schaefer et al., 2017). Brain networks can alsobe extracted from task fMRI data (Calhoun et al., 2008;Lee et al., 2010), and segment a similar intrinsic large-scale structure (Smith et al., 2009). In our work, we buildfunctional modes from datasets with different experimen-tal conditions, including task and rest. Our goal is to beas general as possible and capture information from differ-ent protocols. Indeed, defining networks on task fMRI canhelp representing these brain images and predicting thecorresponding psychological conditions (Duff et al., 2012).
We consider BOLD time-series from fMRI volumes, re-sampled and registered to the MNI template. After tempo-ral concatenation, those form a large matrix X ∈ R p × n ,where p is the number of voxels of the images (around2 · ), and n is the number of brain images, of the or-der of 10 in the following. To extract DiFuMos, eachbrain volume is modeled as the linear combination of k spa-tial functional networks, assembled in a dictionary matrix D ∈ R p × k . We thus assume that X approximately fac-torizes as DA , where the matrix A ∈ R k × n holds in everycolumn the loadings α i necessary to reconstruct the brainimage x i from the networks D . The dictionary D is tobe learned from data. For this, we rely on Stochastic On-line Matrix Factorization (Mensch et al., 2018, SOMF),that is computationally tractable for matrices large in bothdirections, as with high-resolution large-scale fMRI data.SOMF solves the constrained (cid:96) reconstruction problemmin D ∈ R p × k , A ∈ R k × n D (cid:62) , ∀ j ∈ [ k ] , (cid:107) d j (cid:107) (cid:54) (cid:107) X − DA (cid:107) F + λ (cid:107) A (cid:107) F , where λ is a regularization parameter that controls thesparsity of the dictionary D , via the (cid:96) and positivity con-straints. Encouraging sparsity in spatial maps is key toobtaining well-localized maps that outline few brain re-gions. The parameter λ is chosen so that the union of allmaps approximately covers the whole brain. Available at: https://arthurmensch.github.io/modl/
Input fMRI data.
We build the input data matrix X with BOLD time-series from 25 different task-based fMRIstudies and 2 resting state studies, adding up to 2 192functional MRI recording sessions. We gather data fromOpenNeuro (Gorgolewski et al., 2017) –Table A4 lists thecorresponding studies while Table A5 gives their data-acquisition parameters.We use fMRIprep (Esteban et al., 2019) for minimalpreprocessing: brain extraction giving as a reference tocorrect for head-motion (Jenkinson et al., 2002), and co-registration to anatomy (Greve and Fischl, 2009). All thefMRI images are transformed to MNI template space. Wethen use MRIQC (Esteban et al., 2017) for quality control.
Multi-dimensional DiFuMo atlases.
We estimate dictio-naries of dimensionality k ∈ { , , , , } . Thisis useful as the optimal dimensionality for extracting IDPsoften depends on the downstream data analysis task. Theobtained functional modes segment well-localized regions,as illustrated in Figure 2. The functional modes take continuous values (we re-fer to them as soft ) and can have some overlap –thoughin practice this overlap is small. As a consequence, sig-nal extraction calls for more than averaging on regions.The natural formulation is that the extracted signals (theIDPs) should best approximate the brain image x ∈ R p as a linear combination α ∈ R k of the set of modes in thedictionary D ∈ R p × k . This is solved by linear regression: α = argmin α ∈ R k (cid:107) x − Dα (cid:107) , i.e. α = D † x , (1)where D † = ( D T D ) − D T ∈ R k × p is the pseudo-inverseof D . For atlases composed of non-overlapping regions,such as classic brain parcellations—e.g. BASC (Bellecet al., 2010) or normalized cuts (Craddock et al., 2012)—linear regression simply amounts to averaging the imagesvalues in every cluster of D . For overlapping modes as theones of DiFuMo or the ICA maps used in UKBB (Milleret al., 2016), the linear regression formulation caters forthe overlap and softness of the regions. Relating IDPs to known brain structures facilitates in-terpretation and discussion of results. Though the Di-FuMo atlases are defined from functional signal, we chooseto reference their regions by their anatomical location, asit is a common and well-accepted terminology in neuro-science. For each resolution, we match the modes with re-gions in references of brain structure: the Harvard-Oxfordatlas (Desikan et al., 2006), Destrieux atlas (Destrieuxet al., 2010), the MIST atlas (Urchs et al., 2019), JohnsHopkins University (JHU) atlas (Hua et al., 2008), and theDierdrichsen cerebellum atlas (Diedrichsen et al., 2009).We name each mode from the anatomical structure that3 ub li c d a t a s e t s Reducedrepresentations S t a nd a r d f M R I d a t a a n a l y s i s Predictive modeling from functional connectomesEncoding brain activitygiven experimental designGLM
HOUSEFACEPUNISHREWARD
Decoding experimentaldesign given brain activityComponentsprojectionTime series7 resting-statestudiesContrast maps7 task fMRIstudies
2. Validation benchmarks
Map reconstruction and meta-analysis D i c t i o n a r y l ea r n i n go n O p e n f M R I r e c o r d s BASC444 ROIs UKBB ICA55 componentsCraddock et al.400 ROIs
128 components256 components 1024 components512 components64 components etc.
Figure 2:
Schema of DiFuMo atlases and their usage in typical fMRI analyses.
DiFuMo atlases are extracted from a massiveconcatenation of BOLD time-series across fMRI studies, using a sparsity inducing matrix factorization algorithm. We compute the DiFuMoatlases at different resolutions, up to 1024 components. We assess our atlases in 4 benchmarks that measure suitability to classic fMRIanalyses. Those are performed on reduced and non-reduced data, with different atlas sizes and a comparison between atlases. The easiestway to view and download DiFuMo atlases is via the online interactive visualizations: parietal-inria.github.io/DiFuMo. it most overlaps with. When the overlap was weak, atrained neuroanatomist (AMS) looked up the structure instandard classic anatomy references (Henri, 1999; Schmah-mann et al., 1999; Rademacher et al., 1992; Ono et al.,1990; Catani and de Schotten, 2012). Appendix F givesmore details on the naming of the brain areas.
3. Brain-image analysis on functional modes
We use the reduced representations (IDPs) introducedabove for various functional-imaging analytic tasks: stan-dard mass-univariate analysis of brain responses ( § § § § To gauge the usefulness of the extracted IDPs, we com-pare each analysis pipeline across several functional at- lases: DiFuMo and reference atlases are used to computefunctional IDPs. We use the same signal-extraction func-tion (1), but vary the spatial components D . As a baseline,we also perform the voxel-level analyses, though it entailsignificantly larger computational costs.We consider other functional atlases that are multi-resolutions, accessible to download, and volumetric (Ta-ble 1): ICA maps with k ∈ { , } components,extracted on large-scale rs-fMRI from UKBB (Milleret al., 2016); bootstrap analysis of stable clusters (BASC) built with hierarchical clustering on rs-fMRI, with var-ious number of clusters (Bellec et al., 2010); spatially-constrained clustering on rs-fMRI, with k ∈ { , } clusters (Craddock et al., 2012); k = 333 cortical ar-eas derived from rs-fMRI using a local gradient ap-proach (Gordon et al., 2014); k ∈ { , } functionalregions covering cortical and subcortical gray matterwith ICA and Ward clustering (Shirer et al. (2012),Altmann et al. (2015)); and brain parcellations derivedwith gradient-weighted Markov Random Field, with reso-lutions similar to ours (Schaefer et al., 2017, k up to 1000).4 ame Dimensionality a
90, 499 15 Yes ICA; Ward clustering Shirer et al. (2012); Altmann et al. (2015)Gordon 333 120 No Local-gradient approach Gordon et al. (2014)UKBB ICA 21, 55 4100 Yes Selected ICA components b Miller et al. (2016)Schaefer 100, 200, 300, 400, 500,600, 800, 1000 1489 No Gradient-weighted MarkovRandom Field (gwMRF) Schaefer et al. (2017)
DiFuMo c
64, 128, 256, 512, a https://findlab.stanford.edu/functional_ROIs.html b c https://parietal-inria.github.io/DiFuMo Table 1: Functional atlases that we benchmark; they define IDPs for brain-images. analysis
Standard analysis in task fMRI relates psychologicalmanipulations to brain activity separately for each voxelor region. It models the BOLD signal as a linear com-bination of experimental conditions—the General LinearModel (GLM, Friston et al., 1995). The BOLD signalforms a matrix Y ∈ R n × p , where p is the number of voxels.With data reduction, we use as input the reduced signal Y red = Y voxel ( D † ) (cid:62) ∈ R n × k (Equation 1). The GLMmodels Y or Y red as Y = Xβ + ε where X ∈ R n × q is thedesign matrix formed by q temporal regressors of interestor nuisance and ε is noise (Friston et al., 1998). In ourexperiments, we use the Nistats library .With reduced input Y red , we obtain one signal per re-gion, as β ∈ R q × k . The full β -maps can then be recon-structed by setting β rec = βD (cid:62) ∈ R q × p . We transformthe reconstructed β -maps into z-maps z ∈ R q × p using basecontrasts, before thresholding them with Benjamini andHochberg (1995) FDR correction for multiple comparisons.We then compare the z -maps obtained using voxels as in-put, and z -maps using reduced input and reconstructed β -maps, using the Dice (1945) similarity coefficient. We alsoperform an intra-subject analysis detailed in Appendix D. Data.
We consider the Rapid-Serial-Visual-Presentation(RSVP) language task of Individual Brain Charting (IBC)(see Pinho et al., 2018, for experimental protocol and pre-processing). We model six experimental conditions: com-plex meaningful sentences, simple meaningful sentences,jabberwocky, list of words, lists of pseudowords, conso-nant strings. β -maps are estimated for each subject usinga fixed-effect model over 3 out of the 6 subject’s sessions.We randomly select 3 sessions 10 times to estimate thevariance of the Dice index across sessions. As a baseline,we evaluate the mean and variance of the Dice index across z -maps when varying the sessions used in voxel-level GLM. predicts psychological conditions from task-related z -maps (Haynes and Rees, 2006). The validity ofa decoding model is evaluated on left-out data (following https://nistats.github.io/ Varoquaux et al., 2017), e.g. left-out subjects for inter-subject decoding (Poldrack et al., 2009). We use lineardecoding models: ridge regression for continuous targetand Support Vector Machine (SVC, Hastie et al., 2009)for classification. For each study, we separate sessions (forintra-subject decoding) or subjects (for inter-subject de-coding) into randomly-chosen train and test folds (20 foldswith 30% test size), and measure the test accuracy. Wecompare the performance of predictive models using thevoxel-level z -maps or using the data reduced with func-tional atlases. Data.
We use 6 open-access task fMRI studies. We per-form inter-subject decoding in the emotional and sensi-tivity to pain experiences from Chang et al. (2015), andin three studies from HCP900 (Van Essen et al., 2012): working memory , gambling (Delgado et al., 2000), and relational processing (Smith et al., 2007). We perform intra-subject decoding using the several sessions of left and right button press responses in IBC (ARCHI proto-col, Pinel et al., 2007). The unthresholded z -maps used inthe decoding pipeline are either obtained from Neurovault(Gorgolewski et al., 2015), or computed with the GLMfollowing § Resting-state fMRI can be used to predict phenotypictraits (Richiardi et al., 2010). For this, each subject is rep-resented by a functional connectivity matrix that capturesthe correlation between brain signals at various locations.Our functional-connectome prediction pipeline comprisesthree steps: we extract a reduced representation of theBOLD signal, projecting voxel-level data onto a functionalatlas as in § we compute a functional connectome from the reduced BOLD signals; we use it as input to alinear model. We compute a connectome from activationswith the Ledoit and Wolf (2004) covariance estimator asVaroquaux and Craddock (2013); Brier et al. (2015). Wethen derive single-subject features from covariance matri-ces using their tangent space parametrization (Varoquauxet al., 2010a; Barachant et al., 2013; Pervaiz et al., 2019).Those are used to fit an (cid:96) -penalized logistic regression forclassification and a ridge regression for continuous targets.5e assess predictive performance with 20 folds, randomsplits of subjects in train and test sets, with 25% test size. Data.
We use 7 openly-accessible datasets with diversephenotypic targets, as summarized in Table A3. We pre-dict diagnostic status for Alzheimer’s disease on
ADNI (Mueller et al., 2005), PTSD on
ADNIDOD ; AutismSpetrum Disorder on
ABIDE (Di Martino et al., 2014)and schizophrenia on
COBRE (Calhoun et al., 2012);drug consumption on
ACPI ; IQ measures on
HCP (Van Essen et al., 2013); and age (with a regression model)in normal aging with
CamCAN (Taylor et al., 2017).
The signals extracted on a brain atlas can be seen asa compression, or simplification, of the original signal. In-deed, a full image can be reconstructed from these signals.We quantify the signal loss incurred by this reduction. Forthis, we project a brain map x onto an atlas (solving Eq.(1)), and compute the best reconstruction of x from theloadings α , namely ˆ x = Dα ∈ R p . We compare originaland reconstructed images through the R coefficient, R ( x , ˆ x ) = 1 − (cid:107) x − ˆ x (cid:107) (cid:107) x − ¯ x (cid:107) , (2)where ¯ x ∈ R is the spatial mean of map x . The R coeffi-cient is averaged across all images. Higher R coefficientsmeans that the reduced signals (IDPs) explain more vari-ance of the original images, where R = 1 corresponds tono signal loss. The larger the number of signals used, theeasier it is to explain variance; it is therefore interesting tocompare this measure across atlases with similar numberof components. Data.
We use
NeuroVault (Gorgolewski et al., 2015), thelargest public database of statistical maps. To avoid circu-larity, we exclude maps derived from the studies used to ex-tract the DiFuMo atlases, along with maps that fail semi-automated quality inspection (filtering out thresholded ornon-statistical maps), resulting in maps.
Meta-analysis of contrasts maps.
Ideally, the extractedIDPs should allow to compute meta-analytical summariesof brain activity maps. In this setting, a single map, corre-sponding to a certain cognitive concept, is computed frommany z -maps across different studies, associated to con-ditions that involve this cognitive concept. We comparethe summaries obtained at voxel-level, i.e. averaging themaps { x } , with the ones obtained using reconstructed im-ages, i.e. averaging the maps { ˜ x } used in Eq. (2). We usemaps from our curated subset of NeuroVault annotatedwith terms motor, language and face recognition .
4. Results: comparing atlases for analyses
We report benchmarking results on the analytic taskslisted in the previous section.
32 64 128 256 512 1024Dimension0.00.20.40.60.8 M e a n D i c e i n d e x r e l a t i v e t o v o x e l - l e v e l m a p s UKBB ICADiFuMo BASCCraddock FINDGordon SchaeferVoxel-level
Across-fold consistencyat voxel-level
General Linear Model on task fMRI x=14 L Rz=0 -4.5-2.202.24.5
GLM at the voxel level, p ~ 200,000 voxels GLM on DiFuMo reduction, p = 1024 voxels
Group-level z -map: complex sentence Figure 3:
Overlap between GLM maps obtained with func-tional atlases and voxel-level analysis . Top:
The overlap ismeasured with the Dice similarity coefficient. The black line givesa baseline the mean overlap between voxel-level contrast maps overseveral random selections of sessions per subject. The figure givesDice similarity scores between the GLM maps computed with sig-nals extracted on functional atlases and at the voxel-level, after re-construction of full z -maps and voxel-level thresholding with FDRcontrol. The best similarity is achieved for highest dimensional-ity, though 1024-dimensional DiFuMo atlas largely dominates 1000-dimensional Schaefer parcellation. Each point is the mean and theerror bar denotes the standard deviation over contrast maps. Bot-tom:
The activity maps encoded on 1024-dimensional space capturethe same information and much smoother to voxel-level.
Figure 3 reports the results of standard analysis of taskfMRI (GLM), comparing analysis at the voxel-level withanalyses on signals extracted from functional atlases. Bestcorrespondence is obtained at highest dimensionality, asthe regions are finer. Notably, analysis with DiFuMo ofdimensionality 1024 is markedly closer to voxel-level analy-sis than using the largest alternative, the 1000-dimensionalSchaefer parcellation. In addition, the Dice index relativeto the voxel-level gold standard is comparable to the Diceindex between runs of voxel-level GLM estimated acrossfolds. We note that using soft functional modes from only55 ICA components shows excellent results, comparableto those obtained using the 1000 components Schaeferatlas. This stresses the benefit of continuous functionalmodes for the analysis of task responses. Overall, stan-dard task-fMRI analysis on signals derived from 512 or1024-dimensional DiFuMo gives results close to the voxel-6 -10%-8%-6%-4%-2%0%+2%+4% A cc u r a c y g a i n r e l a t i v e t o m e d i a n a t l a s p e r f o r m a n c e UKBB ICADiFuMo BASCCraddock FINDGordon SchaeferVoxel-level
Decoding mental processes from statistical maps
Figure 4:
Impact of the choice of atlas on decoding perfor-mance.
Each point gives the relative prediction score, over 6 differ-ent task-fMRI experiments. The thick lines give the median relativescore per atlas. The baseline (black) is the relative score. High-order resolutions increase prediction accuracy. Using high-order Di-FuMo ( k = 1024) and Schaefer parcellations ( k = 1000) gives thebest performance and, on average, outperforms voxel-level predic-tion. Appendix B.2 reports absolute prediction scores for each taskseparately. Voxel-level DiFuMo=1024 Schaefer=1000
L Rz=-10Voxel-level
L Rz=-10DiFuMo=1024
L Rz=-10Schaefer=1000
Figure 5:
Decoding maps of the working memory task, faceversus rest , showed for Voxel-level analysis, DiFuMo, and Schaefer.The maps are highly interpretable with high-dimensional soft modes(DiFuMo 1024) compared to voxel-level analysis. Brain areas impor-tant in the visual working memory task –fusiform gyrus and lateraloccipital cortex– are clearly visible. Figure A4 gives a full view ofdecoding weights across atlases and resolutions. level gold standard (Figure 3 shows that the maps arealso qualitatively similar). Figure A6 shows similar trendswhile comparing intra-subject explained-variance maps,both qualitatively and quantitatively. Dimension reduc-tion have the additional benefit of alleviating the burdenof correcting for multiple comparisons.
Figure 4 shows the impact on decoding performance ofreducing signals with various functional atlases. It reportsthe performance relative to the median across methods foreach of the 6 tasks. These results clearly show the impor-tance of high-dimensional functional modes for decoding.Indeed, the higher the atlas resolution, the better the pre-dictions. Using DiFuMo k = 1024 or Schaefer k = 1000
32 64 128 256 512 1024Dimension -10%-5%0%+5%+10% A cc u r a c y g a i n r e l a t i v e t o m e d i a n a t l a s p e r f o r m a n c e UKBB ICADiFuMo BASCCraddock FINDGordon Schaefer
Predicting traits from functional connectomes
Figure 6:
Impact of the choice of atlas for predictions basedon functional connectomes.
Each data point gives the predictionaccuracy relative to the median for one of the 7 phenotypic predictiontargets, i.e. each point a dataset. The thick line shows the medianover the datasets. While the results are noisy, the optimal dimension-ality seems to lie around 300 nodes, and the best-performing atlas isDiFuMo k = 256, followed by Craddock k = 400 and BASC k = 444.Figure A5 report absolute results for each prediction problem. gives the best performance. In addition, as these func-tional atlases segment sufficiently-fine regions, predictionfrom the corresponding signals tends to outperform voxel-level prediction. Indeed, applying multivariate models toa larger number of signals with a limited amount of datais more prone to overfitting—data reduction acts here asa welcome regularization. Qualitatively, brain maps con-taining decoding weights can be reconstructed. With high-dimensional atlases, they are interpretable and capture in-formation similar to voxel-level analysis (Figure 5). Figure 6 shows the impact of the choice of functionalatlas when predicting phenotypes from functional connec-tomes. It reports the relative prediction accuracy for 7different prediction problems (each composed of a datasetand a target phenotype); the lines give the median acrossthe prediction problems. Here, we do not report a voxel-level baseline, as it requires to compute covariance matri-ces of dimensions around 100 , × ,
000 and is there-fore computationally and statistically intractable. Unlikewith the previous results, high-resolution atlases do notprovide the best performance, likely because the complex-ity of the statistical models increases with the square of thenumber of nodes. The best prediction overall is achievedusing DiFuMo k = 256, followed by Craddock k = 400and BASC k = 444 atlases. Different outcomes have dif-ferent optimal dimensionality, consistently across atlases(Figure A5): k ∼
150 for age prediction; k ∼
300 forAutism Spectrum Disorder, PTSD, or IQ prediction; and k ∼
50 for Alzheimer’s Disease and drug use prediction.7 igure 7:
Image reconstruction qual-ity. Left:
Quantitative comparison on15542 statistical images. The R loss be-tween the true and recovered images af-ter compression with brain atlases of mul-tiple resolutions. In general, higher-orderatlases capture more signal. Right:
Meta-analysis summaries for the motor task.High R score (left) correspond to bettercapturing fine structures of images, as vis-ible on the qualitative images. DiFuMoatlases better capture the gradients andsmooth aspects of the original images thanhard parcellations, as BASC.
32 64 128 256 512 1024Dimension0.10.20.30.40.50.6 C o m p r e ss i o n R s c o r e UKBB ICADiFuMoBASCCraddockFINDGordonSchaefer
Measuring data fidelity across manystatistical images y=-26 x=-56 z=16
Non-reduced imageReduced with DiFuMo -87 -44 0 44 87
Reduced with BASC
Figure 7 (left) displays the R scores summarizing theloss of information when data are reduced on an atlas andreconstructed back to full images. Unsurprising, reduc-ing the images with lower-order dimensions (atlases withfewer regions) yields a high loss of information across allmethods. DiFuMo k = 1024 captures 70% of the originalvoxel-level signal. Qualitatively, the benefits of functionalmodes can be seen by comparing the meta-analytic mapsrelated to motor tasks (Figure 7 right)—Figure A7 showsadditional meta-analysis on other topics. The DiFuMohave clear visual benefits over brain discrete parcellations,such as BASC, as they better capture gradients.
5. Discussion
This paper introduces brain-wide soft functionalmodes, named DiFuMos and made of a few hundreds to athousand of brain sub-divisions. They are derived fromBOLD time-series across many studies to capture wellfunctional images with a small number of signals. In thecontext of population imaging, these signals are known asimage-derived phenotypes (IDP, Miller et al., 2016) andare crucial to easily scale statistical analysis, building a sci-ence of inter-individual differences by relating brain signalsto behavioral traits (Dubois and Adolphs, 2016). Reduc-ing the dimensionality of the signals not only come with a1000 × gain in storage, but also with 100 × computationalspeed-up for the analysis (Table A1). Even small-scalestudies may need functional nodes, e.g. for computingfunctional connectomes (Zalesky et al., 2010; Varoquauxand Craddock, 2013). There already exist many functionalbrain atlases; yet DiFuMos have the unique advantage ofbeing both soft and highly resolved. These features areimportant to capture gradients of functional information. Grounding better image-derived phenotypes.
Signals ex-tracted from a functional atlas should enable good sta-tistical analysis of brain function. We considered quan-titative measures for typical neuroimaging analytic sce-narii and compared the fitness of extracting signal on Di-FuMo with using existing functional brain atlases. The biggest gains in analysis come from increasing the dimen-sionality of brain sub-divisions, aside for functional con-nectome studies where an optimal is found around 200nodes. Choosing the number of nodes then becomes atradeoff between complexity of the representation and an-alytic performance. Importantly, the gains in analytic per-formance continue way beyond the dimensionality typi-cally used for IDPs (e.g. 55 components from Miller et al.,2016). These results extend prior literature emphasizingthe importance of high-dimensional parcellations for fMRI(Abou Elseoud et al., 2011; Thirion et al., 2014; Arslanet al., 2017; Sala-Llonch et al., 2019). To foster good anal-ysis, the second most important aspect of a parcellationappears that it be soft, i.e. continuously-valued. For agiven dimensionality, soft modes tend to outperform hardparcellations, whether they are derived with ICA or dic-tionary learning.
Modes well-adapted to the EPI signal.
The functionalmodes are optimized to fit well a large number of EPIimages: 2,192 sessions across 27 studies. As a result, theyform a division of the brain well adapted to the signal.For instance, they define regions larger in the white mat-ter and in the CSF than in the grey matter (Figure A1).A large dataset is needed to capture such implicit regular-ities of the signal with high-dimensional spatial decompo-sitions. Indeed, running the same model on less data ex-tracts modes with less spatial regularity (Figure A2). Thecombination of high dimensionality and large dataset leadsto significant computational demands. The extraction ofDiFuMos was possible thanks to fast algorithms for hugematrix factorization (Mensch et al., 2018), and gatheringdata representative of a wide variety of scanning protocolsvia openfMRI (Poldrack et al., 2013).We did not limit the DiFuMo modes to gray matter, asmeasures outside gray matter can be useful in subsequentanalysis, for instance to remove the global signal (Murphyand Fox, 2017). In addition, distributed modes extractedfrom full-brain EPI can separate out noise –such as move-ment artifacts– and help rejecting it in a later analysis(Perlbarg et al., 2007; Griffanti et al., 2014; Pruim et al.,2015). Some DiFuMo modes indeed segment ventricles or8 igure 8:
Modes around the puta-men , for DiFuMo dimensionality 64,256, and 512. As dimensionality in-creases: sub-divisions are more refined,modes are split into right and left hemi-sphere and anterio-posterior direction.Each color represents a single mode.Figure A8 details more this breakdown.
L Rz=3x=-27 dimension=64
L Rz=3x=-27 dimension=256
L Rz=3x=-27 dimension=512 interfaces. Depending on the application, practitioners canchoose to restrict signal extraction to a grey-matter mask.
The functional modes are sharp and anatomically relevant.
To extract structures defined by brain anatomy or mi-crostructure, atlasing efforts have used anatomical or mul-timodal imaging (Mori et al., 2005; Desikan et al., 2006;Eickhoff et al., 2007; Glasser et al., 2016). The DiFuMoatlases capture a different signal: brain activity. Yet,thanks to the sparsity and non-negativity constraint, theyare made of localized modes which often have a naturalanatomical interpretation. Consequently, we have labeledthe modes with a unique name based on the most relevantanatomical structure, following Urchs et al. (2019) whoalso give anatomical labels to functional regions. Indeed,using a common vocabulary of brain structures is impor-tant for communication across the neuroimaging commu-nity. As visible on Figure 8, the modes are well anchoredon anatomical structures such as the putamen. They arehowever not constrained to contain only one connectedregion. Smaller dimension DiFuMos indeed capture dis-tributed networks, often comprising bilateral regions. Asthe dimensionality increases, the networks progressivelyseparate in smaller networks which eventually form singleregions. For instance, the left and right putamen appear inthe same mode at dimension 64, and are first sub-dividedalong the anterio-posterior direction, and later the left andright putamen are separated (Figure 8). Dimension choiceis data driven: it should best explain the functional signal.
6. Conclusion
We provide multidimensional atlases of functionalmodes that can be used to extract functional signals:parietal-inria.github.io/DiFuMo. They give excellent per-formance for a wide variety of analytic tasks: GLM-based analysis, mental-process decoding or functional-connectivity analysis. Their availability reduces compu-tational burdens: practitioners can readily perform anal-yses on a reduced signal, without a costly ROI-definitionstep. In addition, working on common functional modesacross studies facilitates comparison and interpretationsof results. To help communication, we have labeled ev-ery functional mode to reflect the neuroanatomical struc-tures that it contains. To date, these are the only high-dimensional soft functional modes available. As they havebeen extracted from a variety of data (more than 2,000 ses-sions across 27 studies, 2.4TB in size) and improve many analytic tasks, the rich descriptions of neural activity thatthey capture is well suited for a broad set of fMRI studies.
7. Acknowledgments
This project has received funding from the Euro-pean Union’s Horizon 2020 Research and Innovation Pro-gramme under Grant Agreement No. 785907 (HBP SGA2)and No 826421 (VirtualBrainCloud). The work of ArthurMensch has been supported by the European ResearchCouncil (ERC Grant Noria). This work acknowledges thesupport of ANR NeuroRef and ERC-StG NeuroLang.We also thank Pierre Bellec and Vincent Frouinfor their helpful discussions on the experimental work,the neuroimaging community for giving access to fMRIdatasets, and open-source contributors on the packages webuild upon (including nilearn , fMRIprep , and MRIQC eferences
Abou Elseoud, A., Littow, H., Remes, J., Starck, T., Nikkinen,J., Nissil¨a, J., Tervonen, O., Timonen, M., Kiviniemi, V., 2011.Group-ica model order highlights patterns of functional brain con-nectivity. Frontiers in Systems Neuroscience 5, 37.Abraham, A., Milham, M.P., Di Martino, A., Craddock, R.C., Sama-ras, D., Thirion, B., Varoquaux, G., 2017. Deriving reproduciblebiomarkers from multi-site resting-state data: An autism-basedexample. NeuroImage 147, 736–745.Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller,A., Kossaifi, J., Gramfort, A., Thirion, B., Varoquaux, G., 2014.Machine learning for neuroimaging with scikit-learn. Frontiers inneuroinformatics 8.Altmann, A., Ng, B., Landau, S.M., Jagust, W.J., Greicius, M.D.,2015. Regional brain hypometabolism is unrelated to regionalamyloid plaque burden. Brain 138, 3734–3746.Alvarez, R., Poldrack, R., 2011. Cross-language repetition priming.Stanford Digital Repository .Aron, A.R., Behrens, T.E., Smith, S., Frank, M.J., Poldrack, R.A.,2007. Triangulating a cognitive control network using diffusion-weighted magnetic resonance imaging (mri) and functional mri.Journal of Neuroscience 27, 3743–3752.Aron, A.R., Gluck, M.A., Poldrack, R.A., 2006. Long-term test-retest reliability of functional mri in a classification learning task.NeuroImage 29, 1000 – 1006.Arslan, S., Ktena, S.I., Makropoulos, A., Robinson, E.C., Rueckert,D., Parisot, S., 2017. Human brain mapping: A systematic com-parison of parcellation methods for the human cerebral cortex.NeuroImage .Barachant, A., Bonnet, S., Congedo, M., Jutten, C., 2013. Classifi-cation of covariance matrices using a riemannian-based kernel forbci applications. Neurocomputing 112, 172 – 178.Beckmann, C., DeLuca, M., Devlin, J., Smith, S., 2005. Investiga-tions into resting-state connectivity using independent componentanalysis. Philos Trans R Soc Lond B 360, 1001.Behzadi, Y., Restom, K., Liau, J., Liu, T., 2007. A component basednoise correction method (compcor) for BOLD and perfusion basedfMRI. Neuroimage 37, 90.Bellec, P., Rosa-Neto, P., Lyttelton, O., Benali, H., Evans, A., 2010.Multi-level bootstrap analysis of stable clusters in resting-statefMRI. NeuroImage 51, 1126.Benjamini, Y., Hochberg, Y., 1995. Controlling the false discoveryrate: A practical and powerful approach to multiple testing. J RSTAT SOC B (Methodological) 57, 289.Brett, M., Johnsrude, I.S., Owen, A.M., 2002. The problem of func-tional localization in the human brain. Nat Rev Neurosci 3, 243.Brier, M.R., Mitra, A., McCarthy, J.E., Ances, B.M., Snyder, A.Z.,2015. Partial covariance based functional connectivity computa-tion using ledoit–wolf covariance regularization. NeuroImage 121,29.Brodmann, K., 1909. Vergleichende Lokalisationslehre der Grosshirn-rinde in ihren Prinzipien dargestellt auf Grund des Zellenbaues.Barth.Calhoun, V., Sui, J., Kiehl, K., Turner, J., Allen, E., Pearlson, G.,2012. Exploring the psychosis functional connectome: Aberrantintrinsic networks in schizophrenia and bipolar disorder. Frontiersin Psychiatry .Calhoun, V.D., Adali, T., Pearlson, G.D., Pekar, J.J., 2001. Amethod for making group inferences from fMRI data using in-dependent component analysis. Hum Brain Mapp 14, 140.Calhoun, V.D., Kiehl, K.A., Pearlson, G.D., 2008. Modulation oftemporally coherent brain networks estimated using ICA at restand during cognitive tasks. Hum Brain Map 29, 828. Catani, M., de Schotten, M.T., 2012. Atlas of Human Brain Con-nections. Oxford University Press.Cera, N., Tartaro, A., Sensi, S.L., 2014. Modafinil alters intrinsicfunctional connectivity of the right posterior insula: A pharmaco-logical resting state fmri study. PLOS ONE 9, 1–12.Chang, L.J., Gianaros, P.J., Manuck, S.B., Krishnan, A., Wager,T.D., 2015. A sensitive and specific neural signature for picture-induced negative affect. PLOS Biology 13, 1–28.Craddock, R.C., James, G.A., Holtzheimer, P.E., Hu, X.P., Mayberg,H.S., 2012. A whole brain fMRI atlas generated via spatiallyconstrained spectral clustering. Hum brain map 33, 1914.Dadi, K., Rahim, M., Abraham, A., Chyzhyk, D., Milham, M.,Thirion, B., Varoquaux, G., 2019. Benchmarking functionalconnectome-based predictive models for resting-state fMRI. Neu-roImage 192, 115 – 134.Damoiseaux, J.S., Rombouts, S.A.R.B., Barkhof, F., Scheltens, P.,Stam, C.J., Smith, S.M., Beckmann, C.F., 2006. Consistentresting-state networks across healthy subjects. Proc Natl AcadSci 103, 13848.Delgado, M.R., Nystrom, L.E., Fissell, C., Noll, D.C., Fiez, J.A.,2000. Tracking the hemodynamic responses to reward and punish-ment in the striatum. Journal of Neurophysiology 84, 3072–3077.Desikan, R., S., S´egonne, F., Fischl, B., Quinn, B., T., Dickerson,B., C., Blacker, D., Buckner, R., L., Dale, A., M., Maguire, R.,P., Hyman, B., T., Albert, M., S., Killiany, R., J., 2006. An auto-mated labeling system for subdividing the human cerebral cortexon MRI scans into gyral based regions of interest. Neuroimage 31,968.Destrieux, C., Fischl, B., Dale, A., Halgren, E., 2010. Automatic par-cellation of human cortical gyri and sulci using standard anatom-ical nomenclature. NeuroImage 53, 1 – 15.Devlin, J.T., Poldrack, R.A., 2007. In praise of tedious anatomy.NeuroImage 37, 1033 – 1041.Di Martino, A., Yan, C.G., Li, Q., Denio, E., Castellanos, F.X.,Alaerts, K., Anderson, J.S., Assaf, M., Bookheimer, S.Y.,Dapretto, M., et al., 2014. The autism brain imaging data ex-change: towards a large-scale evaluation of the intrinsic brain ar-chitecture in autism. Molecular psychiatry 19, 659–667.Dice, L.R., 1945. Measures of the amount of ecologic associationbetween species. Ecology 26, 297–302.Diedrichsen, J., Balsters, J.H., Flavell, J., Cussans, E., Ramnani,N., 2009. A probabilistic mr atlas of the human cerebellum. Neu-roImage 46, 39 – 46.Dubois, J., Adolphs, R., 2016. Building a science of individual dif-ferences from fmri. Trends in cognitive sciences 20, 425–443.Duff, E.P., Trachtenberg, A.J., Mackay, C.E., Howard, M.A., Wilson,F., Smith, S.M., Woolrich, M.W., 2012. Task-driven ica featuregeneration for accurate and interpretable prediction using fmri.NeuroImage 60, 189 – 203.Duncan, K.J., Pattamadilok, C., Knierim, I., Devlin, J.T., 2009.Consistency and variability in functional localisers. Neuroimage46, 1018.Eickhoff, S.B., Paus, T., Caspers, S., Grosbras, M.H., Evans, A.C.,Zilles, K., Amunts, K., 2007. Assignment of functional activationsto probabilistic cytoarchitectonic areas revisited. Neuroimage 36,511.Eickhoff, S.B., Yeo, B.T.T., Genon, S., 2018. Imaging-based parcel-lations of the human brain. Nat Rev Neurosci 19, 672.Elliott, P., Peakman, T.C., et al., 2008. The UK biobank samplehandling and storage protocol for the collection, processing andarchiving of human blood and urine. Int J Epidemiology 37, 234.Esteban, O., Birman, D., Schaer, M., Koyejo, O.O., Poldrack, R.A.,Gorgolewski, K.J., 2017. MRIQC: Advancing the automatic pre-diction of image quality in MRI from unseen sites. PLOS ONE12, 1. steban, O., Markiewicz, C.J., Blair, R.W., Moodie, C.A., Isik, A.I.,Erramuzpe, A., Kent, J.D., Goncalves, M., DuPre, E., Snyder,M., et al., 2019. fMRIPrep: a robust preprocessing pipeline forfunctional MRI. Nature methods 16, 111.Foerde, K., Knowlton, B.J., Poldrack, R.A., 2006. Modulation ofcompeting memory systems by distraction. Proc Natl Acad Sci103, 11778.Friston, K., Fletcher, P., Josephs, O., Holmes, A., Rugg, M., Turner,R., 1998. Event-related fmri: Characterizing differential re-sponses. NeuroImage 7, 30 – 40.Friston, K.J., Holmes, A.P., Worsley, K.J., Poline, J.B., Frith, C.,Frackowiak, R., 1995. Statistical parametric maps in functionalimaging: A general linear approach. Hum Brain Mapp , 189.Gabitov, E., Manor, D., Karni, A., 2015. Patterns of modulation inthe activity and connectivity of motor cortex during the repeatedgeneration of movement sequences. J Cog Neurosci 27, 736.Glasser, M.F., Coalson, T.S., Robinson, E.C., Hacker, C.D., Har-well, J., Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C.F.,Jenkinson, M., Smith, S.M., Essen, D.C.V., 2016. A multi-modalparcellation of human cerebral cortex. Nature 536, 171–178.Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fis-chl, B., Andersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni,J.R., Essen, D.C.V., Jenkinson, M., 2013. The minimal prepro-cessing pipelines for the human connectome project. NeuroImage80, 105 – 124.Gordon, E.M., Laumann, T.O., Adeyemo, B., Huckins, J.F., Kelley,W.M., Petersen, S.E., 2014. Generation and Evaluation of a Cor-tical Area Parcellation from Resting-State Correlations. CerebralCortex 26, 288–303.Gorgolewski, K., Esteban, O., Gunnar, S., Brain, W., Poldrack, R.,2017. Openneuro – a free online platform for sharing and analysisof neuroimaging data., in: 23rd Annual Meeting of the Organiza-tion for Human Brain Mapping, p. 1677.Gorgolewski, K.J., Storkey, A., Bastin, M.E., Whittle, I.R., Ward-law, J.M., Pernet, C.R., 2013. A test-retest fmri dataset for motor,language and spatial attention functions. GigaScience 2, 2047–217X–2–6.Gorgolewski, K.J., Varoquaux, G., Rivera, G., Schwarz, Y., Ghosh,S.S., Maumet, C., Sochat, V.V., Nichols, T.E., Poldrack, R.A.,Poline, J.B., Yarkoni, T., Margulies, D.S., 2015. Neurovault.org:a web-based repository for collecting and sharing unthresholdedstatistical maps of the human brain. Frontiers in Neuroinformatics9, 8.Goutte, C., Toft, P., Rostrup, E., Nielsen, F.A., Hansen, L.K., 1999.On clustering fMRI time series. NeuroImage 9, 298–310.Greicius, M., Krasnow, B., Reiss, A., Menon, V., 2003. Functionalconnectivity in the resting brain: a network analysis of the defaultmode hypothesis. Proc Natl Acad Sci 100, 253.Greve, D.N., Fischl, B., 2009. Accurate and robust brain imagealignment using boundary-based registration. NeuroImage 48, 63.Griffanti, L., Salimi-Khorshidi, G., Beckmann, C.F., Auerbach, E.J.,Douaud, G., Sexton, C.E., Zsoldos, E., Ebmeier, K.P., Filippini,N., Mackay, C.E., et al., 2014. ICA-based artefact removal andaccelerated fMRI acquisition for improved resting state networkimaging. Neuroimage 95, 232–247.Hanson, S.J., Matsuka, T., Haxby, J.V., 2004. Combinatorial codesin ventral temporal lobe for object recognition: Haxby (2001)revisited: is there a “face” area? NeuroImage 23, 156 – 166.Harrison, S.J., Woolrich, M.W., Robinson, E.C., Glasser, M.F.,Beckmann, C.F., Jenkinson, M., Smith, S.M., 2015. Large-scaleprobabilistic functional modes from resting state fmri. NeuroIm-age 109, 217 – 231.Hastie, T., Tibshirani, R., Friedman, J., 2009. The elements of sta-tistical learning. Springer. Haxby, J.V., Gobbini, I.M., Furey, M.L., et al., 2001. Distributedand overlapping representations of faces and objects in ventraltemporal cortex. Science 293, 2425.Haynes, J.D., Rees, G., 2006. Decoding mental states from brainactivity in humans. Nat. Rev. Neurosci. 7, 523.Henri, M.D., 1999. The Human Brain: Surface, Three-dimensionalSectional Anatomy with MRI, and Blood Supply. Springer.Hua, K., Zhang, J., Wakana, S., Jiang, H., Li, X., Reich, D.S., Cal-abresi, P.A., Pekar, J.J., van Zijl, P.C., Mori, S., 2008. Tractprobability maps in stereotaxic spaces: Analyses of white matteranatomy and tract-specific quantification. NeuroImage 39, 336.Huntenburg, J.M., Bazin, P.L., Margulies, D.S., 2018. Large-scalegradients in human cortical organization. Trends in cognitive sci-ences 22, 21.Hyv¨arinen, A., Oja, E., 2000. Independent component analysis: al-gorithms and applications. Neural Networks 13, 411.Iannilli, E., Gasparotti, R., Hummel, T., Zoni, S., Benedetti, C.,Fedrighi, C., Tang, C.Y., Van Thriel, C., Lucchini, R.G., 2016.Effects of manganese exposure on olfactory functions in teenagers:A pilot study. PLOS ONE 11, 1–9.Jenkinson, M., Bannister, P., Brady, M., Smith, S., 2002. Improvedoptimization for the robust and accurate linear registration andmotion correction of brain images. NeuroImage 17, 825 – 841.Jimura, K., Cazalis, F., Stover, E.R.S., Poldrack, R.A., 2014. Theneural basis of task switching changes with skill acquisition. Fron-tiers in Human Neuroscience 8, 339.Kelly, A.C., Uddin, L.Q., Biswal, B.B., Castellanos, F.X., Milham,M.P., 2008. Competition between functional brain networks me-diates behavioral variability. NeuroImage 39, 527 – 537.Kim, J., Wang, J., Wedell, D.H., Shinkareva, S.V., 2016. Identi-fying core affect in individuals from fmri responses to dynamicnaturalistic audiovisual stimuli. PLOS ONE 11, 1–21.Kiviniemi, V., Kantola, J., Jauhiainen, J., Hyv¨arinen, A., Tervonen,O., 2003. Independent component analysis of nondeterministicfmri signal sources. Neuroimage 19, 253.Kiviniemi, V., Starck, T., Remes, J., Long, X., Nikkinen, J., Haapea,M., Veijola, J., et al., 2009. Functional segmentation of the braincortex using high model order group PICA. Hum Brain Map 30,3865.Ledoit, O., Wolf, M., 2004. A well-conditioned estimator for large-dimensional covariance matrices. J. Multivar. Anal. 88, 365.Lee, K., Tak, S., Ye, J.C., 2010. A data-driven sparse GLM forfMRI analysis using sparse dictionary learning with MDL crite-rion. IEEE Trans Med Imag 30, 1076.Leech, R., Kamourieh, S., Beckmann, C.F., Sharp, D.J., 2011. Frac-tionating the default mode network: Distinct contributions of theventral and dorsal posterior cingulate cortex to cognitive control.J Neurosci 31, 3217.Lepping, R.J., Atchley, R.A., Chrysikou, E., Martin, L.E., Clair,A.A., Ingram, R.E., Simmons, W.K., Savage, C.R., 2016a. Neuralprocessing of emotional musical and nonmusical stimuli in depres-sion. PLOS ONE 11, 1–23.Lepping, R.J., Atchley, R.A., Savage, C.R., 2016b. Developmentof a validated emotionally provocative musical stimulus set forresearch. Psychology of Music 44, 1012–1028.Mennes, M., Kelly, C., Colcombe, S., Castellanos, F.X., Milham,M.P., 2013. The extrinsic and intrinsic functional architectures ofthe human brain are not equivalent. Cerebral Cortex 23, 223–229.Mensch, A., Mairal, J., Bzdok, D., Thirion, B., Varoquaux, G.,2017. Learning Neural Representations of Human Cognitionacross Many fMRI Studies, in: Neural Information ProcessingSystems, p. 5885.Mensch, A., Mairal, J., Thirion, B., Varoquaux, G., 2016a. Dictio-nary Learning for Massive Matrix Factorization, in: InternationalConference on Machine Learning, pp. 1737–1746. ensch, A., Mairal, J., Thirion, B., Varoquaux, G., 2018. StochasticSubsampling for Factorizing Huge Matrices. IEEE Trans Sig Proc66, 113.Mensch, A., Varoquaux, G., Thirion, B., 2016b. Compressed OnlineDictionary Learning for Fast Resting-State fMRI Decomposition,in: Proc. ISBI, p. 1282.Michel, V., Gramfort, A., Varoquaux, G., Eger, E., Keribin, C.,Thirion, B., 2012. A supervised clustering approach for fMRI-based inference of brain states. Pattern Recognition 45, 2041.Miller, K.L., Alfaro-Almagro, F., et al., 2016. Multimodal popula-tion brain imaging in the UK biobank prospective epidemiologicalstudy. Nature Neuroscience .Moran, J.M., Jolly, E., Mitchell, J.P., 2012. Social-cognitive deficitsin normal aging. J Neurosci 32, 5553.Mori, S., Wakana, S., Van Zijl, P.C., Nagae-Poetscher, L., 2005. MRIatlas of human white matter. Elsevier.Mour˜ao-Miranda, J., Bokde, A.L., Born, C., Hampel, H., Stetter, M.,2005. Classifying brain states and determining the discriminatingactivation patterns: Support vector machine on functional MRIdata. NeuroImage 28, 980.Mueller, S., Weiner, M., Thal, L., Petersen, R., Jack, C., Jagust, W.,Trojanowski, J.Q., Toga, A.W., Beckett, L., 2005. The alzheimer’sdisease neuroimaging initiative. Neuroimaging Clinics of NorthAmerica 15, 869.Murphy, K., Fox, M.D., 2017. Towards a consensus regarding globalsignal regression for resting state functional connectivity MRI.Neuroimage 154, 169–173.Olshausen, B., Field, D., 1997. Sparse coding with an overcompletebasis set: A strategy employed by V1? Vision research 37, 3311.Ono, M., Kubik, S., Abernathey, C.D., 1990. Atlas of the cerebralsulci. G. Thieme Verlag.O’Toole, A.J., Jiang, F., Abdi, H., Haxby, J.V., 2005. Partially dis-tributed representations of objects and faces in ventral temporalcortex. J Cog Neurosci 17, 580.Pedregosa, F., Varoquaux, G., Gramfort, A., et al., 2011. Scikit-learn: Machine learning in Python. Journal of Machine LearningResearch 12, 2825.Perlbarg, V., Bellec, P., Anton, J.L., Pelegrini-Issac, M., Doyon,J., Benali, H., 2007. CORSICA: correction of structured noisein fMRI by automatic identification of ICA components. MagnReson Imaging 25, 35.Pervaiz, U., Vidaurre, D., Woolrich, M.W., Smith, S.M., 2019. Op-timising network modelling methods for fmri. bioRxiv .Pinel, P., Thirion, B., Meriaux, S., Jobert, A., Serres, J., Le Bihan,D., Poline, J., Dehaene, S., 2007. Fast reproducible identifica-tion and large-scale databasing of individual functional cognitivenetworks. BMC neuroscience 8, 91.Pinho, A.L., Amadon, A., Ruest, T., Fabre, M., Dohmatob, E.,Denghien, I., Ginisty, C., Becuwe-Desmidt, S., Roger, S., Laurier,L., Joly-Testault, V., M´ediouni-Cloarec, G., Doubl´e, C., Martins,B., Pinel, P., Eger, E., Varoquaux, G., Pallier, C., Dehaene, S.,Hertz-Pannier, L., Thirion, B., 2018. Individual Brain Charting,a high-resolution fMRI dataset for cognitive mapping. ScientificData 5, 180105.Poldrack, R.A., Barch, D.M., Mitchell, J.P., et al., 2013. Towardopen sharing of task-based fMRI data: the OpenfMRI project.Frontiers in neuroinformatics 7.Poldrack, R.A., Clark, J., Par´e-Blagoev, E.J., Shohamy, D.,Creso Moyano, J., Myers, C., Gluck, M.A., 2001. Interactive mem-ory systems in the human brain. Nature 414, 546–550.Poldrack, R.A., Halchenko, Y.O., Hanson, S.J., 2009. Decoding thelarge-scale structure of brain function by classifying mental statesacross individuals. Psychological Science 20, 1364. Power, J., Cohen, A., Nelson, S., Wig, G., Barnes, K., Church, J.,Vogel, A., Laumann, T., Miezin, F., Schlaggar, B., Petersen, S.,2011. Functional network organization of the human brain. Neu-ron 72, 665–678.Pruim, R.H., Mennes, M., Buitelaar, J.K., Beckmann, C.F., 2015.Evaluation of ICA-AROMA and alternative strategies for motionartifact removal in resting state fMRI. Neuroimage 112, 278–287.Rademacher, J., Caviness, V. S., J., Steinmetz, H., Galaburda, A.M.,1993. Topographical Variation of the Human Primary Cortices:Implications for Neuroimaging, Brain Mapping, and Neurobiol-ogy. Cerebral Cortex 3, 313–329.Rademacher, J., Galaburda, A.M., Kennedy, D.N., Filipek, P.A.,Caviness, V.S., 1992. Human cerebral cortex: Localization, par-cellation, and morphometry with magnetic resonance imaging. JCog Neurosci 4, 352.Repovs, G., Barch, D., 2012. Working memory related brain networkconnectivity in individuals with schizophrenia and their siblings.Frontiers in Human Neuroscience 6, 137.Richiardi, J., Eryilmaz, H., Schwartz, S., Vuilleumier, P., VanDe Ville, D., 2010. Decoding brain states from fMRI connectivitygraphs. NeuroImage .Rizk-Jackson, A., Aron, A., Poldrack, R., 2011. Classification learn-ing and stop-signal (1 year test-retest). Stanford Digital Reposi-tory .Romaniuk, L., Pope, M., Nicol, K., Steele, D., Hall, J., 2016. Neu-ral correlates of fears of abandonment and rejection in borderlinepersonality disorder. Wellcome Open Research 1.Roy, A., Bernier, R.A., Wang, J., Benson, M., French, Jr., J.J.,Good, D.C., Hillary, F.G., 2017. The evolution of cost-efficiencyin neural networks during recovery from traumatic brain injury.PLOS ONE 12, 1–26.Sala-Llonch, R., Smith, S.M., Woolrich, M., Duff, E.P., 2019. Spatialparcellations, spectral filtering, and connectivity measures in fmri:Optimizing for discrimination. Hum Brain Map 40, 407.Schaefer, A., Kong, R., Gordon, E.M., Laumann, T.O., Zuo, X.N.,Holmes, A.J., Eickhoff, S.B., Yeo, B.T.T., 2017. Local-Global Par-cellation of the Human Cerebral Cortex from Intrinsic FunctionalConnectivity MRI. Cerebral Cortex 28, 3095–3114.Schmahmann, J.D., Doyon, J., McDonald, D., Holmes, C., Lavoie,K., Hurwitz, A.S., Kabani, N., Toga, A., Evans, A., Petrides, M.,1999. Three-dimensional mri atlas of the human cerebellum inproportional stereotaxic space. NeuroImage 10, 233 – 260.Schonberg, T., Fox, C., Mumford, J., Congdon, E., Trepel, C., Pol-drack, R., 2012. Decreasing ventromedial prefrontal cortex ac-tivity during sequential risk-taking: An fmri investigation of theballoon analog risk task. Frontiers in Neuroscience 6, 80.Shirer, W., Ryali, S., Rykhlevskaia, E., Menon, V., Greicius, M.,2012. Decoding subject-driven cognitive states with whole-brainconnectivity patterns. Cerebral Cortex 22, 158.Smith, R., Keramatian, K., Christoff, K., 2007. Localizing the ros-trolateral prefrontal cortex at the individual level. NeuroImage36, 1387 – 1396.Smith, S., Fox, P., Miller, K., Glahn, D., Fox, P., Mackay, C., et al.,2009. Correspondence of the brain’s functional architecture duringactivation and rest. Proc Natl Acad Sci 106, 13040.Smith, S., Miller, K., Salimi-Khorshidi, G., Webster, M., Beckmann,C., Nichols, T., Ramsey, J., Woolrich, M., 2011. Network mod-elling methods for fMRI. Neuroimage 54, 875.Sporns, O., Tononi, G., Kotter, R., 2005. The human connectome:a structural description of the human brain. PLoS Comput Biol1, e42.Stephan-Otto, C., Siddi, S., Senior, C., Mu˜noz-Samons, D., Ochoa,S., S´anchez-Laforga, A.M., Br´ebion, G., 2017. Visual imageryand false memory for pictures: A functional magnetic resonanceimaging study in healthy participants. PLOS ONE 12, 1–17. aylor, J.R., Williams, N., Cusack, R., Auer, T., Shafto, M.A.,Dixon, M., Tyler, L.K., Cam-CAN, Henson, R.N., 2017. Thecambridge centre for ageing and neuroscience (cam-CAN) datarepository: Structural and functional MRI, MEG, and cognitivedata from a cross-sectional adult lifespan sample. NeuroImage144, 262.Thirion, B., Flandin, G., Pinel, P., Roche, A., Ciuciu, P., Poline,J., 2006. Dealing with the shortcomings of spatial normalization:Multi-subject parcellation of fMRI datasets. Hum brain map 27,678.Thirion, B., Varoquaux, G., Dohmatob, E., Poline, J., 2014. WhichfMRI clustering gives good brain parcellations? Frontiers in Neu-roscience 8, 167.Uncapher, M.R., Hutchinson, J.B., Wagner, A.D., 2011. Dissocia-ble effects of top-down and bottom-up attention during episodicencoding. Journal of Neuroscience 31, 12613–12628.Urchs, S., Armoza, J., Moreau, C., Benhajali, Y., St-Aubin, J., Or-ban, P., Bellec, P., 2019. MIST: A multi-resolution parcellationof functional brain networks. MNI Open Research 1.Van Essen, D., Ugurbil, K., Auerbach, E., Barch, D., Behrens, T.,Bucholz, R., Chang, A., Chen, L., Corbetta, M., Curtiss, S., DellaPenna, S., Feinberg, D., Glasser, M., Harel, N., Heath, A., Larson-Prior, L., Marcus, D., Michalareas, G., Moeller, S., Oostenveld,R., Petersen, S., Prior, F., Schlaggar, B., Smith, S., Snyder, A.,Xu, J., Yacoub, E., 2012. The human connectome project: A dataacquisition perspective. NeuroImage 62, 2222–2231.Van Essen, D.C., Smith, et al., 2013. The wu-minn human connec-tome project: an overview. Neuroimage 80, 62–79.Varoquaux, G., Baronnet, F., Kleinschmidt, A., Fillard, P., Thirion,B., 2010a. Detection of brain functional-connectivity differencein post-stroke patients using group-level covariance modeling, in:MICCAI.Varoquaux, G., Craddock, R.C., 2013. Learning and comparing func-tional connectomes across subjects. NeuroImage 80, 405.Varoquaux, G., Gramfort, A., Pedregosa, F., Michel, V., Thirion,B., 2011. Multi-subject dictionary learning to segment an atlas ofbrain spontaneous activity, in: Inf Proc Med Imag, p. 562.Varoquaux, G., Raamana, P.R., Engemann, D.A., Hoyos-Idrobo, A.,Schwartz, Y., Thirion, B., 2017. Assessing and tuning brain de-coders: cross-validation, caveats, and guidelines. NeuroImage 145,166.Varoquaux, G., Sadaghiani, S., Pinel, P., Kleinschmidt, A., Poline,J.B., Thirion, B., 2010b. A group model for stable multi-subjectICA on fMRI datasets. NeuroImage 51, 288.Verstynen, T.D., 2014. The organization and dynamics of corti-costriatal pathways link the medial orbitofrontal cortex to futurebehavioral responses. J Neurophysio 112, 2457.Xue, G., Aron, A.R., Poldrack, R.A., 2008. Common neural sub-strates for inhibition of spoken and manual responses. CerebralCortex 18, 1923–1932.Xue, G., Poldrack, R.A., 2007. The neural substrates of visual per-ceptual learning of words: Implications for the visual word formarea hypothesis. J Cog Neurosci 19, 1643.Yeo, B., Krienen, F., Sepulcre, J., Sabuncu, M., et al., 2011. Theorganization of the human cerebral cortex estimated by intrinsicfunctional connectivity. J Neurophysio 106, 1125.Zalesky, A., Fornito, A., Harding, I.H., Cocchi, L., Y¨ucel, M., Pan-telis, C., Bullmore, E.T., 2010. Whole-brain anatomical networks:Does the choice of nodes matter? NeuroImage 50, 970 – 983. L Ry=-35 x=40 L Rz=20 DiFuMo atlases capture well the EPI signal
Figure A1:
Region volume ( cm ) of modes on the brain with1024 dictionary of DiFuMo. The volume of the modes tends tobe larger corresponding to white matter when compared with thecortical gray matter. This justifies the adaptation of DiFuMo atlasto the fMRI signal. x=2 x=2 / th of full training size x=2 x=2 Full training size: 2192 volumes
Figure A2: 1024 components trained on two different sizes of theinput set of fMRI images. The components trained on the full datahave more spatial regularity, while the components trained on 100volumes have more overlap in some regions of the brain. The addi-tional spatial regularity shows the benefit of large-scale training sizein learning a data-driven based functional atlas.
Appendix A. Performance of DiFuMos
As discussed in §
5, we report how DiFuMOs compo-nents are well adapted to the fMRI EPI signal in Fig-ure A1. Figure A2 qualitatively compare components ob-tained training on the whole data corpus and training ona fraction of it. Better component regularity is obtainedwith more data. Finally, Table A1 reports the computa-tional speed-ups obtained using DiFuMos IDPs instead ofvoxel in the decoding experiment. Similar speed-ups areobserved in the other validation pipelines.
Appendix B. Details on stimulus decoding
We provide additional details for the decoding pipeline,to complete the description in § a s k s a m p l e s R e p r e s e n t a t i o n T i m e ( s e c ) S p e e d u p
Emotion 4924 Voxel-level 77.7 46 × Reduced 1.7Pain 84 Voxel-level 1.5 250 × Reduced 0.006Working 3140 Voxel-level 874.7 240 × memory Reduced 3.7Gambling 1574 Voxel-level 298.7 270 × Reduced 1.12Relational 1572 Voxel-level 263.1 405 × Reduced 0.65
Table A1: The comparison in computational times while predict-ing mental state on two set of brain features space: voxel-level ≈ ,
000 and reduced voxels to DiFuMo 1024. We report the aver-aged time over 20 cross-validation folds for several task-fMRI condi-tions. Clearly, there are benefits trading for reduced representationsin terms of computation time. On high-resolution brain images likeHCP, these are decreased by a factor 200.
Task-fMRI Prediction task
NV503: Emotion Rating:1 , , , , , , Table A2:
Dataset, prediction tasks and dataset size for eachof the 6 decoding tasks we consider in § z -maps from HCPand IBC were computed using the GLM, while NeuroVault directlyprovided the β -maps for Emotion and Pain. NV: NeuroVault. Appendix B.1. Input data and pre-processing pipelines
The decoding pipeline classifies input unthresholdedstatistical maps. Table A2 summarizes the task-basedstudies used to obtain these statistical maps.
Pre-encoded maps downloaded from Neurovault.org.
Wedownload maps related to emotion and pain (Chang et al.,2015) using
Neurovault , querying the collections and . We use the “Rating” & “PainLevel” labels as predic-tive targets. We predict emotion using ridge regression,and pain-level over 3 classes using Linear SVC. The super-vised learning pipeline, that includes cross-validation andlinear models is implemented with Python based scikit-learn (Pedregosa et al., 2011). We use nilearn (Abrahamet al., 2014) to download maps from Neurovault.org in-terface (Gorgolewski et al., 2015). The data acquisitionparameters, preprocessing details and estimation of statis-tical maps are described in Chang et al. (2015).
Statistical maps encoded using the GLM.
We compute z -maps from HCP900 (Van Essen et al., 2012) and IBC(Pinho et al., 2018) studies, that comprise high-qualiytask-fMRI experiments. -20%0%20%40% R s c o r e emotion A cc u r a c y pain A cc u r a c y face vs place A cc u r a c y punish vs reward A cc u r a c y relational vs match
32 64 128 256 512 1024 V o x e l s Dimension A cc u r a c y left vs right button press (intra-subject decoding) UKBB ICADiFuMoBASC CraddockFINDGordon SchaeferNon reduced
Predicting mental state: Task-fMRI
Figure A3:
Decoding prediction scores for each brain atlasand target : Each marker denotes the mean performance of usinga certain brain atlas; error bars are the standard deviation of theprediction scores for this atlas. Decoding from high-order dictionar-ies, and especially from DiFuMos, perform similarly or better thandecoding from voxels.
HCP.
We download fMRI data from the HCP900 re-lease; those are already preprocessed using HCP pipelines(Glasser et al., 2013). We use MNINonLinear-based reg-istered data as input for the GLM, that outputs one z -map per condition per subject. We consider three task-based studies, namely: for Working Memory , we con-14ider z -maps based on condition: “0-back faces”, “2-backfaces”, “0-back places”, “2-back places”. Similarly, for Gambling (Delgado et al., 2000), we consider z -maps forthe conditions “loss” and “reward”; finally, on Relationalprocessing , we consider z -maps for the conditions “rela-tional processing” and “matching”. For each study, weuse Linear SVC on encoded z -maps to predict psycholog-ical conditions. The predictive model therefore perform a2-class or 4-class classification. The experimental protocoland data acquisition parameters are detailed in Van Essenet al. (2012). IBC.
We consider the
Archi Standard (Pinel et al.,2007) motor task, where subjects are asked to press “left”and “right” button press based on audio and visual in-structions. We perform within-subject classification be-tween left and right button press, using z -maps corre-sponding to each repetition of the instruction. For eachof the 13 available subjects, a linear model is trained onthe z -maps from all but one session and prediction is per-formed on the left-out session. Each subject provides 80encoded z -maps across 4 sessions. We use data prepro-cessed following the pipelines of Pinho et al. (2018). GLM specification.
For both datasets, the input z -maps are estimated from the raw fMRI data by fitting aGLM. We use Nistats , a Python package for the statisticalanalysis of fMRI data. The temporal regressors of themodel are specified according to the timing of stimuluspresentations convolved with hemodynamic models ( spm+ derivative ). We use polynomial model to capture thelow-frequency drifts in the data. Appendix B.2. Detailed results
To complete the summarizing Figure 4, we report theraw prediction scores separately for each decoding tasks inFigure A3. Prediction accuracy increases with the size offunctional atlases. Using 1024 atlases allows to match orpass the performance of voxel-based prediction. In termsof interpretation, the weights are much smoother and blobsare clearly visible in the weight classification maps ob-tained using DiFuMo. This is illustrated on Figure A4 forface-versus-place decoding in the working-memory HCPstudy.
Appendix C. Details on biomarker prediction
We consider multiple datasets to account for the diver-sity of prediction targets in biomarker prediction problem.We report datasets, prediction groups and prediction tar-gets in Table A3.
Appendix C.1. Input data and prediction settings
The connectivity features built from functional atlasespredict various clinical outcomes (neuro-degenerative andneuro-psychiatric disorders, drug abuse impact) and psy-chological traits. https://nistats.github.io/ -0.012-0.00600.0060.012 -0.21-0.1100.110.21-0.21-0.100.10.21-0.25-0.1200.120.25-0.32-0.1600.160.32 -2.4-1.201.22.4 -0.28-0.1400.140.28-0.28-0.1400.140.28 Voxel-level 200,000 voxelsBASC
64 197 444
Craddock
200 400
FIND
90 499
Gordon
UKBB ICA
21 55
Schaefer
100 500 1000
DiFuMo
Decoding working memory: 0BK face versus rest
Figure A4:
Decoding classification weight maps for the HCPworking memory task (0BK face) , obtained with voxel-leveldecoding and decoding over various functional atlases. Using DiFu-Mos yield highly interpretable weight maps; it clearly delineates thefusiform gyrus and lateral occipital cortex.
Group classification.
We use the Alzheimer’s Disease Neu-roimaging Initiative (ADNI) and (ADNIDOD) (Muelleret al., 2005) to predict neuro-degenerative diseases. Wediscriminate between Alzheimer’s Disease (AD) from Mild est-fMRI Prediction groups Samples HCP900 High IQ vs Low IQ 443 subjects213/230ABIDE Autism vs control 866 subjects402/464ACPI Marijuana use vs control 126 subjects62/64ADNI Alzheimers vs MCI 136 subjects40/96ADNIDOD PTSD vs control 167 subjects89/78COBRE Schizophrenia vs control 142 subjects65/77CamCAN Age 626 subjects24 − Table A3:
Resting-state fMRI datasets used in the pipelinedescribed on § In CamCAN, age is predicted usingridge regression. The groups from other datasets are predicted usinglogistic regression. IQ - Fluid intelligence, PTSD - Post TraumaticStress Disorder, MCI - Mild Cognitive Impairment.
Cognitive Impairment (MCI) group on ADNI. We discrim-inate between post-traumatic stress disorder (PTSD) andhealthy individuals on ADNIDOD. We use data from theCenter for Biomedical Research Excellence (COBRE Cal-houn et al. (2012) to predict schizophrenia diagnosis of in-dividuals. We classify autism and healthy individuals onAutism Brain Imaging Data Exchange database (ABIDE,Di Martino et al. (2014), Finally, we consider data fromAddiction Connectome Preprocessed Initiative (ACPI),where we discriminate Marijuana consumers versus con-trol subjects. Psychological traits.
We first stratify individuals fromHCP900 release (Van Essen et al., 2013) into groups of highand low IQ, and perform binary classification on these.The details about the stratification into these groups aredescribed in Dadi et al. (2019).
Age regression.
We use Cambridge Center for Ageing andNeuroscience (CamCAN) dataset (Taylor et al., 2017) tostudy brain ageing. This study comprises wide range ofage groups spanning from 24 – 86. We use ridge regressionto predict age on this dataset.
Appendix C.2. Data acquisition parameters and pre-processing pipelines
The data acquisition details for ADNI, ADNIDOD,COBRE, ABIDE, ACPI and HCP are described in Dadiet al. (2019); those for CamCAN in Taylor et al. (2017).We pre-process individuals from CamCAN, ADNI, AD-NIDOD and COBRE. All rs-fMRI acquistions are pre-processed with standard steps, described in Dadi et al. http://fcon_1000.projects.nitrc.org/indi/ACPI/html/ R s c o r e Age A cc u r a c y Autism vs Control A cc u r a c y Marijuana use vs Control A cc u r a c y Alzheimers Disease vs MCI A cc u r a c y PTSD vs Control A cc u r a c y Schizophrenia vs Control
32 64 128 256 512 1024
Dimension A cc u r a c y High IQ vs Low IQ
UKBB ICADiFuMoBASC CraddockFIND GordonSchaefer
Predicting phenotypes from functional connectomes
Figure A5:
Connectome prediction scores for each brain at-lases and target : Each marker denotes the mean performance ofusing a certain brain atlas; error bars are the standard deviation ofthe prediction scores for this atlas. BASC and DiFuMo-based atlasesgive good prediction scores up to k = 256 ROIs. Confound removal and temporal signal pre-processing.
The strategy we use for cleaning temporal signals is thesame as in Dadi et al. (2019). We brieftly outline thesesteps here. We regress out 10 CompCor (Behzadi et al.,2007) components on the whole brain and six motion re-lated signals which are provided in the ADNI, ADNIDOD,COBRE, CamCAN datasets. We do not perform any ad-ditionnal preprocessing steps on ABIDE, ACPI and HCP.For all datasets, the signal of each region is normalized,detrended and bandpass-filtered between 0.01 and 0.1Hz.All these steps are done with nilearn (Abraham et al.,2014).
Appendix C.3. Detailed results
Figure 6 summarizes the impact of the brain atlasesand ROIs in predicting diverse targets on rs-fMRI images.Figure A5 shows the absolute prediction scores for eachtarget separately. DiFuMo-based predictions are on parwith those using UKBB ICA components, Craddock et al.(2012) and BASC atlases.
Appendix D. Intra-subject encoding In § z -maps computed atthe voxel-level and on reduced data using the DICE sim-ilarity coefficient. We also performed an intra-subject,across sessions, standard analysis. We consider the Rapid-Serial-Visual-Presentation (RSVP) language task of Indi-vidual Brain Charting (IBC) (see Pinho et al. (2018) fordetails on experimental protocol and data pre-processing). Encoding model.
In this setting, we fit a GLM on the sev-eral acquisition sessions of each subject considered sepa-rately. That is, we compute a single β -map per sessionand condition, forming a set of maps β ∈ R q × p . β iseither computed directly at the voxel-level or using func-tional atlases, in which case we set β = β red D (cid:62) , with β ∈ R q × k .We then use a leave-one-session-out cross-validationscheme to compare the observed, single-session, time se-ries Y ∈ R n × p to the reconstructed time-series ˆ Y = X ¯ β ,where ¯ β are the average β -maps across the 5 training ses-sions. We obtain R -maps, where each voxel holds theproportion of variance explained by the model r i = 1 − (cid:107) y i − ˆ y i (cid:107) (cid:107) y i − ¯ y i (cid:107) , where y i is the univariate time-series in R n associated tovoxel i and ¯ y i is its temporal mean. We finally average ('BASC', 64) s =0.48=0.51 x=-55 L Rz=2 ('BASC', 122) s =0.57=0.56 x=-55 L Rz=2 Encoding activations across single-subject sessions ('BASC', 197) s =0.62=0.59 x=-55 L Rz=2 ('BASC', 325) s =0.69=0.64 x=-55 L Rz=2 ('BASC', 444) s =0.73=0.66 x=-55 L Rz=2 ('Craddock', 200) s =0.57=0.57 x=-55 L Rz=2 ('Craddock', 400) s =0.60=0.59 x=-55 L Rz=2 ('ICA', 21) s =0.43=0.34 x=-55 L Rz=2 ('ICA', 55) s =0.50=0.39 x=-55 L Rz=2 ('DiFuMo', 64) s =0.40=0.49 x=-55 L Rz=2 ('DiFuMo', 128) s =0.61=0.62 x=-55 L Rz=2 ('DiFuMo', 256) s =0.73=0.68 x=-55 L Rz=2 ('DiFuMo', 512) s =0.86=0.74 x=-55 L Rz=2 ('DiFuMo', 1024) s =0.94=0.79 x=-55 L Rz=2 R score from voxelwise R s c o r e f r o m m a p Non-reduced s =1.00=1.00 x=-55 L Rz=2 -0.08-0.0400.040.08 Figure A6:
Intra-subject univariate prediction of brain re-sponse in the language task protocol of the IBC dataset.
We compare R -maps obtained using voxel based and functional-atlas based encoding models. Encoding models based on high-orderatlases better explain the variance of an unseen session. The compar-ison is made for a single subject; results are similar across subjets. R scores across leave-one-session-out folds, and thresh-old non-positive values. The resulting R -maps providesinformation on how much encoded β -maps are able to pre-dict univariate voxel activation on new sessions. A valueclose to 1 means that the voxel activation is well predictedby the encoding model, while a 0 value means that thevoxel activation cannot be predicted. We compare the R -maps across the various data-reduction methods forestimating β .17 alidation. To measure the difference between R maps R computed from voxels and R maps ˜ R computed fromDiFuMos, we report correlation coefficients ρ between R and ˜ R , and the slope s predicting the activations ˜ R fromthe activations R . This slope indicates a form of signalloss due to using functional atlases. We expect it to besmaller than 1, in part because projection on functionalatlases have a noise reduction effect. Results.
Figure A6, using higher order DiFumo atlasesleads to a loss of explained variance R of only 6% com-pared to working directly with voxels, which may imputedto a denoising effect. Qualitatively, the R maps are muchcomparable. DiFuMo ( k = 1024) is therefore suitable forintra-subject encoding tasks; they make these much lesscostly. Using lower-order atlases yield stronger signal loss. Appendix E. Extra meta-analysis maps
Figure A7 shows the meta-analysis summary imagesfor two additional cognitive topics: language and face . Wecompare non-reduced images with reduced images usingDiFuMo ( k = 1024) and BASC ( k = 444). The images re-duced with DiFuMos are easier to interpret than the onesreduced with BASC for both topics. Quantitatively, werecall that Figure 7 shows the better performance of Di-FuMos for image compression. Appendix F. DiFuMos naming details
A measure of overlap with a reference anatomical atlasallows to match each DiFuMo component with a specificanatomical region, e.g. “postcentral gyrus”. Where thereare more than one component for each anatomical region,the functional atlas region are further characterized by ananatomical spatial descriptions, e.g. “postcentral gyrus in-ferior”. Finally, we append the localisation of the region inthe left or right hemisphere, e.g. “postcentral gyrus infe-rior RH”. Some of the nodes from DiFuMo atlases overlapsa fraction of several regions in the anatomical atlas—thoseare named by a trained neuroanatomist. 18 =-38 x=-55 z=2 -97-4804897 a. Meta-analysis of ”language” y=-75 x=28 z=-11 -14-70714 b. Meta-analysis of ”face”
Non-reduced images Image reduced with DiFuMo Image reduced with BASC
Figure A7:
Meta-analysis on cognitive topics –language (a.) and face (b.) – from statistical images : We compare imagesreconstructed with DiFuMo ( k = 1024) and BASC ( k = 444) with voxel-level averages (right). The topic-related activations are bettervisualized using DiFuMo (middle) than using BASC (left). DiFuMo results are closer to voxel-level averages, as the signal loss is minimalwhen projecting on this atlas. dimension=64Putamen dimension=128Putamen, anterior dimension=256Putamen dimension=256Putamen, anterior Modes around putamen: a. Smaller atlases contains bilateral networks dimension=512 postero-superior left hemisphere superior postero-inferior anterior right hemisphere b. Increasing the atlas dimension splits those networks into different components
Figure A8:
Interpretation of higher-dimensional modes of DiFuMo : The putamen segmentation is refined as dimension of DiFuMosincreases. A single mode contain the left and right putamen in lower dimension (a) , when higher order atlases holds separate components forthem. Larger atlases model the detailed organization within the sub-structures, which may be crucial in discriminative tasks. MRI study Version Cognitive task
Sub j ec t s S e ss i o n s R un s Conditions(Schonberg et al., 2012) ds000001 R2.0.4 Balloon Analog 16 3 balloon analog riskRisk-taking(Aron et al., 2006) ds000002 R2.0.5 Classification learning 17 2 deterministic classificationmixed event related probeprobabilistic classification(Xue and Poldrack, 2007) ds000003 R2.0.2 Rhyme judgment 13 rhyme judgment(Jimura et al., 2014) ds000006 R2.0.1 ds000006 14 2 6 living nonliving decision-with plain or mirror reversed text(Xue et al., 2008) ds000007 R2.0.1 Stop-signal task with 20 2 stop manualspoken & manual responses stop vocalstop word(Aron et al., 2007) ds000008 R2.0.0 Stop-signal task with 14 3 conditional stop signalunconditional and conditional stop signalstopping(Foerde et al., 2006) ds000011 R2.0.1 Classification learning 14 2 Classification probe withoutand tone counting feedbackDual task weather predictionSingle task weather predictionTone counting(Rizk-Jackson et al., 2011) ds000017 R2.0.1 Classification learning 8 2 3 probabilistic classificationand stop-signal (1 year test-retest) selective stop signal task(Alvarez and Poldrack, 2011) ds000051 R2.0.2 Cross-language 13 8 abstract concrete judgmentrepetition priming(Poldrack et al., 2001) ds000052 R2.0.0 Classification learning 14 2 weather predictionand reversal reversal weather prediction(Mennes et al., 2013) ds000101 R2.0.0 Simon task 21 2 simon(Kelly et al., 2008) ds000102 R2.0.0 Flanker task 26 2 flanker(event-related)(Haxby et al., 2001) ds000105 R2.0.2 Visual object recognition 6 12 object viewing(O’Toole et al., 2005)(Hanson et al., 2004)(Duncan et al., 2009) ds000107 R2.0.2 Word and 49 2 1-back taskobject processing(Moran et al., 2012) ds000109 R2.0.2 False belief task 36 2 theory of mind(Uncapher et al., 2011) ds000110 R2.0.1 Incidental encoding task 18 10 Incidental encoding task(Posner Cueing Paradigm)(Gorgolewski et al., 2013) ds000114 R2.0.1 A test-retest fMRI dataset 10 2 covert verb generationfor motor, language and finger footlipsspatial attention functions line bisectionovert verb generationovert word generation(Repovs and Barch, 2012) ds000115 R2.0.0 Working memory in healthy 1 letter 0-back taskand schizophrenic individuals letter 1-back taskletter 2-back task(Cera et al., 2014) ds000133 R1.0.0 Modafinil alters intrinsic 26 2 3 restfunctional connectivity of theright posterior insula: apharmacologicalresting state fMRI study(Verstynen, 2014) ds000164 R1.0.1 Stroop task 28 stroop(Gabitov et al., 2015) ds000170 R1.0.1 Learning and memory: motor 15 3 Trained Hand Trained Sequenceskill consolidation and Trained Hand Untrained Sequenceintermanual transfer Untrained Hand Trained Sequence(Lepping et al., 2016a) ds000171 R1.0.0 Neural Processing of Emotional 39 5 music(Lepping et al., 2016b) Musical and Nonmusical non musicStimuli in Depression(Iannilli et al., 2016) ds000200 R1.0.0 Pre-adolescents Exposure 1 olfactoryto Manganese(Stephan-Otto et al., 2017) ds000203 R1.0.2 Visual imagery and 26 2 visual imagery-false memory for pictures false memory(Kim et al., 2016) ds000205 R1.0.0 Affective Videos 11 2 functional localizerview(Romaniuk et al., 2016) ds000214 R1.0.0 EUPD Cyberball 40 Cyberball(Roy et al., 2017) ds000220 R1.0.0 Cost Analysis TBI 26 2 restTable A4: Large-scale fMRI datasets downloaded from OpenNeuro to build our multi-scale functional atlases. Data are pre-processed using fMRIprep . The data acquisition parameters of each study are listed on Table A5. The corpus is 2.4TB in total. MRI study MR scanner Slice FoV Voxel size Matrix TR TE Flip angle Number oforientation ( mm ) ( mm ) size ( msec ) ( msec ) ( ◦ ) volumes(Schonberg et al., 2012) 3T Siemens AG axial - 4 × × ×
64 2000 30 90 300Allegra (Erlangen,Germany)(Aron et al., 2006) 3T Siemens - - 4 × × ×
64 2000 30 90 180Allegra(Xue and Poldrack, 2007) 3T Siemens - 200 4 × × ×
64 2000 30 90 160Allegra (Iselin, NJ)(Jimura et al., 2014) 3T Siemens - 200 4 × × ×
64 2000 30 90 205Allegra (Erlangen,Germany)(Xue et al., 2008) 3T Siemens - 200 4 × × ×
64 2000 30 90 182Allegra(Aron et al., 2007) 3T Siemens - 200 4 × × ×
64 2000 30 90 176Allegra(Foerde et al., 2006) 3T Siemens - 200 4 × × ×
64 2000 30 - 208Allegra(Poldrack et al., 2001) 3T Siemens axial 200 5 × × ×
64 3000 30 - 225Allegra(Mennes et al., 2013) 3T Siemens - 192 3 × × ×
64 2000 30 80 101Allegra(Kelly et al., 2008) 3T Siemens - 192 3 × × ×
64 2000 30 80 146Allegra(Haxby et al., 2001) 3T GE sagittal 240 3 . × . × . × × . ×
64 3000 50 - 165(Moran et al., 2012) 3T Siemens axial - 3 × × . × . × .
44 64 ×
64 2000 30 75 186(Gorgolewski et al., 2013) 1.5T GE Signa - 256 4 × × ×
64 2500 50 90 varied(Repovs and Barch, 2012) 3T Tim Trio - 256 4 × × ×
64 2500 27 90 137(Cera et al., 2014) 3T Philips transaxial 256 4 × × ×
64 1671 35 75 145(Verstynen, 2014) 3T Siemens - - 3 . × . × . × . × . ×
64 3000 35 90 45(Lepping et al., 2016a) 3T Siemens axial 220 2 . × . × ×
64 3000 25 90 105Skyra (Erlangen,Germany)(Iannilli et al., 2016) 1.5T Siemens axial - 3 . × . × . ×
64 2500 50 - 120Aera (Erlangen,Germany)(Stephan-Otto et al., 2017) 1.5T GE Signa axial 240 4 × × ×
64 2000 40 90 267(Kim et al., 2016) 3T Siemens axial - 3 × × ×
64 2200 35 90 365Trio (Erlangen)(Romaniuk et al., 2016) 3T Siemens axial 220 3 . × . × ×
64 1560 26 66 341Magnetom Verio(Roy et al., 2017) 3T Philips axial 240 4 × × ×
80 2000 30 90 144AchievaTable A5: Data acquisition parameters for each fMRI study that we use for training DiFuMo atlases. Data are downloaded from OpenNeuro.80 2000 30 90 144AchievaTable A5: Data acquisition parameters for each fMRI study that we use for training DiFuMo atlases. Data are downloaded from OpenNeuro.