The Human Connectome Project’s Neuroimaging Approach Matthew F. Glasser1, Stephen M. Smith2, Daniel S. Marcus1, Jesper Andersson2, Edward J. Auerbach3, Timothy E. J. Behrens2, Timothy S. Coalson1, Michael P. Harms4, Mark Jenkinson2, Steen Moeller3, Emma C. Robinson2, Stamatios N. Sotiropoulos2, Junqian Xu5, Essa Yacoub3, Kamil Ugurbil3, David C. Van Essen1   1Department  of  Neuroscience,  Washington  University  Medical  School,  Saint  Louis,  MO,  USA;   2FMRIB  Centre,  Nuffield  Department  of  Clinical  Neurosciences,  John  Radcliffe  Hospital,   University  of  Oxford,  Oxford,  UK;  3Center  for  Magnetic  Resonance  Research  (CMRR),  University   of  Minnesota,  Minneapolis,  MN,  USA;  4Department  of  Psychiatry,  Washington  University   Medical  School,  Saint  Louis,  MO;  5Translational  and  Molecular  Imaging  Institute,  Icahn  School   of  Medicine  at  Mount  Sinai,  New  York,  NY,  USA   Corresponding author: David C. Van Essen; address correspondence to [email protected] and [email protected]. Abstract Non-invasive human neuroimaging has yielded many exciting discoveries about the brain. Numerous methodological advances have also occurred, though inertia has slowed their adoption. This paper presents an integrated approach to data acquisition, analysis, and sharing that builds upon recent advances, particularly from the Human Connectome Project (HCP). The “HCP-style” paradigm has seven core tenets: (1) collect multimodal imaging data from many subjects; (2) acquire data at high spatial and temporal resolution; (3) preprocess data to minimize distortions, blurring, and temporal artifacts; (4) represent data using the natural geometry of cortical and subcortical structures; (5) accurately align corresponding brain areas across subjects and studies; (6) analyze data using neurobiologically accurate brain parcellations; and (7) share published data via user-friendly databases. We illustrate the HCP- style paradigm using existing HCP datasets and provide guidance for future research. Widespread adoption of this paradigm should accelerate progress in understanding the brain in health and disease. Introduction The Human Connectome Project (HCP) began in 2010 with ~$40M awarded by the National Institutes of Health (NIH) to two consortia to develop improved neuroimaging methods and to acquire a dataset of unprecedented size and quality for mapping the normal human macro-scale connectome1,2 that is, the long-distance connections between all of the brain’s areas. Better maps of the brain’s areas and their connections will deepen our understanding of healthy brain function and may improve our ability to understand and treat neurological and psychiatric disorders. The “WU-Minn-Ox” HCP consortium, centered at Washington University, University of Minnesota, and Oxford University, spent its first two years developing the state of the art methods discussed here, and then proceeded to acquire, analyze, and share high quality multimodal neuroimaging, behavioral, and genotype data from 1200 healthy young adult twins and non-twin siblings3,4. The HCP focused primarily on MRI-based imaging modalities to measure brain architecture, function, connectivity, and the regular, topographic, organization of some brain areas; however, magnetoencephalography (MEG) data was also acquired in some   1 subjects (see Box 2). As the HCP’s core activities draw to a close in 2016, its contributions include (i) a set of broadly useful MRI acquisition protocols, consisting of thoroughly tested and optimized pulse sequences and image reconstruction algorithms; (ii) an unprecedented collection of exceptionally high quality, freely shared neuroimaging data; (iii) numerous publicly available neuroimaging software and informatics tools; (iv) a growing number of intriguing discoveries emerging from analyses of HCP data; (v) the emergence of an “HCP-style” paradigm for neuroimaging data acquisition, analysis, and sharing; and (vi) a growing number of HCP-style projects that will study different age ranges and brain disorders. This article is both a position paper and a primer, promoting awareness and adoption of the integrated HCP-style paradigm along with guidance on its use. We provide an accessible conceptual overview of seven core tenets of the paradigm and summarize several recent discoveries made possible through its application. Additional technical information and discussion is provided in the cited publications and in 18 sections of the Supplementary Information (SI) that expand on specific topics raised in the main text. Investigators interested in practical aspects of implementing the HCP paradigm in their own lab can reference the MRI protocols, HCP software, and HCP Course materials listed below. The seven tenets of the HCP-style paradigm span the domains of acquisition, analysis, and sharing of MRI-based data: (1) acquiring large amounts of high quality, multi-modal data on as many subjects as feasible to measure architectural, functional, connectional, and topographical information4 (2) acquiring this MRI data with high spatial, temporal, and angular resolution for structural, functional, and diffusion MRI using cutting-edge accelerated acquisition protocols5; (3) minimizing blurring introduced by preprocessing and removing distortions, noise, and temporal artifacts as selectively and completely as possible6,7; (4) representing cortical data (surface vertices) and subcortical data (volume voxels) in a common geometrical framework (“CIFTI grayordinates”) that is optimal for each6; (5) accurately aligning corresponding brain areas across subjects and studies using “areal features” related to connectivity and architecture8,9; (6) using structurally and functionally relevant brain parcellations (preferably based on multiple modalities) to provide a strong neuroanatomical framework for condensing complex neuroimaging data and enhancing statistical sensitivity and power without blurring across areal boundaries8; and (7) routine sharing of extensively analyzed results such as statistical maps10 (plus raw and preprocessed data when feasible11) together with the code used for the analysis, so that other neuroscientists can make precise comparisons across studies, along with replicating and extending findings. The HCP-style paradigm builds on recent advances in data acquisition, analysis, and sharing from many neuroimaging labs worldwide, both within and outside the HCP consortia. It benefitted greatly from the NIH’s emphasis on methodological development prior to high- throughput data acquisition. The HCP-style paradigm differs from common practices in the human neuroimaging community that emerged during the early years of PET and fMRI (the ‘traditional paradigm’), which often involve (i) voxel-based (‘volumetric’) analysis of data acquired at relatively coarse spatial and temporal resolutions on small groups of subjects; (ii) extensive unconstrained spatial smoothing that attempts to compensate for imperfect cross- subject alignment, improve signal-to-noise ratio (SNR), and satisfy oversimplified statistical assumptions12; (iii) not using parcellations at all or using parcellations derived from Brodmann’s 2D schematic drawing; and (iv) reporting summary results using 3D volumetric coordinates while keeping the extensively analyzed data private. Relative to the new paradigm, the traditional paradigm has limitations in spatial fidelity, neuroanatomical fundamentals, and the robustness of cross-study comparisons, along lines also noted by Turner13,14.   2 Though most neuroimaging studies still use the traditional paradigm, important exceptions dating as far back as the 1990’s include numerous “retinotopy” studies that use functional MRI (fMRI) data to map the arrangement of visual areas on cortical surface meshes of individual subjects15,16. Thanks in part to this emphasis on surface-based analyses, the visual system is one of the better understood functional systems in the human brain. We anticipate accelerated progress in understanding other brain regions as the new paradigm is applied to them more routinely. A solid conceptual understanding of MRI acquisition, preprocessing, and analysis (tenets 1-5) is important for applying the HCP-style paradigm to one’s own studies, and for understanding the methodological choices made by the HCP and its successors (see below). Many of these choices were driven by fundamental properties of human brain anatomy, MR physics, and human physiology. Tenets 1 and 2: Multi-modal Data Acquisition Informed by Brain Anatomy and Physiology. Comprehensive non-invasive brain mapping of many subjects using multi-modal MRI. The HCP acquired ~4 hours of structural, functional, and diffusion images (along with auxiliary ‘field map’ images to help with image preprocessing). The structural images were T1-weighted (T1w) and T2-weighted (T2w) images (see Supplementary Topic #1 for a discussion of the relationship between these terms and MRI contrast). These images allow non-invasive measurement of architectural properties like myelin content within the cerebral cortex and cortical thickness that match well with corresponding invasive measures17-19, see also Supplementary Topic #1. The structural images also generate geometrical models of brain anatomy (cortical surface meshes and subcortical segmentations17,20-27) that are critical for all other analyses. The functional images included task-based fMRI, where the subjects carried out behavioral tasks while in the scanner28, and resting-state scans, where subjects fixated on a crosshair. These functional images enable measurement of functional activation within brain areas, and ‘functional connectivity’ based on correlations of the fMRI signal between brain areas in the absence of a task or controlled external stimuli. They can also map the internal topographic organization within brain areas, such as the ordered representation of visual space within visual areas using task driven15,16,29 or resting-state-based methods (visuotopic organization)8. The similarity of these detailed maps across species30 and in comparison with invasive studies in non-human primates,31,32 confirms that fMRI can accurately capture properties of brain organization. Diffusion MRI uses directionally dependent diffusion of water within white matter to encode average axonal fiber orientations, which can be connected by streamlining tractography algorithms to estimate structural brain connectivity, yielding results that are modestly correlated with invasively measured structural connectivity33,34. Diffusion MRI also allows estimates of microscopic structural properties of brain tissue, such as the integrity and orderliness of axons (e.g.35-37. Thus, the HCP MRI data enables brain architecture, function, connectivity, and topography to be estimated non-invasively. The large number of subjects (~1100) imaged by the HCP enables statistically precise estimation of these measures in the healthy young adult population and investigation of the nature and extent of their individual variability, while the focus on twins (and their non-twin siblings) enables assessment of their heritability. How brain anatomy drives MRI spatial resolution. The cerebral cortex, the seat of human cognition, is a folded sheet, 1.6 - 4mm thick (mean ~2.6 mm), with a surface area of ~1,000 cm2/hemisphere18,38. High quality structural images (both T1w and T2w) are vital for generating accurate geometrical models upon which all other analyses are based (surface meshes for the cerebral cortex and grey matter ROIs for the subcortical structures6. For the main HCP, we   3 acquired these images at an isotropic spatial resolution of 0.7mm (voxel cubes 0.7mm on each side), enabling accurate individual subject maps of cortical myelin content and thickness together with precise models of the inner (white matter) and outer (pial) surfaces. In general, we recommend acquiring T1w and T2w images at 0.8mm isotropic or better (half the minimum thickness of the cortex), in contrast to the single 1mm T1w image traditionally acquired. Such high-resolution images benefit from tightly fitting multi-channel MR receive head coils, such as the Siemens 32-channel head coil used by the HCP. For fMRI, it is important to acquire images of sufficient spatial resolution to distinguish between cortical gray matter, cerebrospinal fluid (CSF), and white matter, and between opposite banks of cortical folds. The HCP acquired high-resolution fMRI data at 2 mm isotropic (with high resolution fMRI being defined as having voxels smaller than the mean cortical thickness of 2.6 mm, see Supplementary Fig. 1) on its customized 3-tesla (3T) scanner (and at 1.6 mm isotropic at 7T). High-resolution fMRI enables more precise localization of functional signals 6 in contrast to resolutions as low as 4 mm that have been traditionally used in fMRI. Diffusion MRI (dMRI) aims to address a fundamentally resolution-starved problem: mapping connections through brain white matter whose individual axons are mostly <1 micron diameter 39 with MRI voxels that exceed 1 mm isotropic, contain hundreds of thousands of axons 40, and typically include multiple populations of fibers crossing in various directions. Both of the HCP consortia used customized MRI scanners and customized MRI pulse sequences to acquire in vivo human diffusion data of unprecedented resolution5,41-46. The 3T WU-Minn HCP scanner uses very strong magnetic field gradients (100 mT/m), approximately double that which commercial scanners could achieve when the project began in 2010. Along with new and optimized dMRI pulse sequences, this enabled 1.25mm isotropic resolution in human dMRI in vivo (see Supplementary Topic #2 and Supplementary Figs. 2 and 3 for more details) and 1.05mm isotropic at 7T45,47, in contrast to traditional dMRI with 2-2.5 mm resolution. The recently introduced Siemens Prisma MRI scanner (80 mT/m) developed in part because of the HCP enables the community to acquire similar dMRI data. Faster, longer imaging for cleaner fMRI data and more robust white matter fiber orientation estimates. Most MRI-based functional imaging relies on the Blood Oxygen Level Dependent (BOLD) contrast mechanism in which increased neural activity indirectly alters the regional deoxygenated hemoglobin content over a time course (‘hemodynamic response function’, or HRF) of many seconds48. Accordingly, one might think that it is sufficient to acquire fMRI images at intervals modestly faster than the HRF (2-3 s Repetition Time, TR, has been typical for whole brain imaging). However, fMRI images are corrupted by many sources of time- dependent (temporal) artifacts, including subject movement, physiology, and scanner instability. Many artifacts occur more rapidly than the typical fMRI volume TR and thus are aliased into the signal. High temporal resolution fMRI data (i.e., TR < 1 s) facilitates more selective and effective artifact removal7, and improves statistical efficiency, particularly for multi-variate statistics49. Recent data also suggest that there may be faster components to the BOLD response50,51, which would be better captured with faster scans. The HCP acquired fMRI data at TR = 0.72 s for 3T and 1 s for 7T using MRI pulse sequences that acquire many brain slices at a time (‘multi-band’ - also known as ‘simultaneous multi slice’ - sequences, see Supplementary Topic #35,43,52-54 and Supplementary Fig. 4). For diffusion imaging, using these sequences to reduce the TR allows for more images to be acquired with diffusion weightings at different strengths and along a larger number of unique spatial directions in the same amount of time. The resulting higher angular resolution and effective SNR enables multiple fiber populations to be more robustly identified in each imaging voxel, helping to resolve crossing fiber pathways and improve tractography5,44.   4 The HCP used 1 hour of scan time for each of task fMRI, resting state fMRI, and diffusion imaging. This enabled seven fMRI tasks to be acquired covering multiple cognitive domains and involving most of the brain28 http://protocols.humanconnectome.org/HCP/3T/task-fMRI- protocol-details.html, along with four 15-min resting state fMRI runs (helping to provide more stable connectivity estimates55), and 270 different diffusion measures (90 unique diffusion directions for each of three different diffusion weighting strengths (i.e. “shells”), enabling better fiber orientation and tissue microstructure modeling at high resolution56,57. See Supplementary Topic #4 for particulars on various “HCP-Short” protocols optimized for subjects who are less tolerant of long scans than healthy young adults. Tenets 3 and 4: Optimizing the Precision of Data Analysis Aligning all MRI modalities within the subject’s physical space. A key early task of preprocessing is to align all MRI modalities so that every image volume is in the subject’s undistorted physical space. Head movements occur in the scanner, both between acquisition of each image and during acquisition. Because head movement during T1w and T2w images causes uncorrectable blurring, the HCP acquired at least two pairs of T1w and T2w images, and used all scans that were rated good or excellent in overall quality (11; see Supplementary Topic #5 for further discussion of this issue and potential solutions). Spatial effects of movement during fMRI or dMRI acquisitions can be corrected after the fact using image registration to realign the images; however, movement also induces temporal variations of image intensities that must be corrected through other means described below. Movement between images of different modalities is corrected with advanced registration algorithms that use knowledge of brain tissue boundaries and the expected image intensity gradients across those boundaries to align images58. Ideally, image registration within each subject would entail simple translations or rotations, but MRI images have many forms of spatial distortion that must be unwarped before an image faithfully represents the individual brain’s real physical dimensions and thus can be precisely aligned. These distortions all derive from imperfections in the scanner’s encoding of physical space using perturbations (linear gradients) of the main magnetic field. All MRI images are slightly distorted by small nonlinearities in the space-encoding gradients that are correctible using information measured and specified by the manufacturer (gradient distortion, see Supplementary Topic #6). In addition, placement of the subject’s head in the scanner induces idiosyncratic variations in the magnetic field (field inhomogeneities), particularly near the air- filled spaces in the head (sinuses and ear canals), which can be measured using the auxiliary ‘field map’ scan mentioned above. Correcting these distortions with a field map is critical for accurately aligning data across modalities within each subject, though traditionally this scan is often omitted, leading to poor alignment in some brain regions. Scanner vendors provide a standard field map scan (based on gradient echoes), but the HCP developed an alternative ‘spin echo EPI field map’, that is faster, at least as accurate, and also enables correction of fMRI image intensity biases (see Supplementary Topic #7). These field map scans should exactly match the geometry and distortion properties of the functional imaging scans. Diffusion imaging scans have additional ‘eddy current’ distortions caused by switching on and off the strong magnetic field gradients used to generate diffusion weighting. The HCP developed a new algorithm for correcting eddy current distortions that outperforms previous methods and allows quality control and outlier detection (see Supplementary Topic #8;59-61). With field maps, advanced distortion correction methods, and anatomically informed registration algorithms, the HCP’s spatial preprocessing pipelines6 precisely align the T1w, T2w, fMRI, and diffusion data to   5 the subject’s physical space while minimizing interpolation-induced blurring by combining all spatial transformations in a single spline interpolation. Cleaning time-dependent artifacts from MRI timeseries data. In functional imaging, many processes cause time-dependent intensity fluctuations other than the BOLD-related fluctuations of interest. This structured temporal noise is especially problematic for resting-state functional connectivity because it uses correlations of noisy signals in different parts of the brain, rather than correlation of a predefined noiseless task design with noisy data. Temporal noise can either increase or reduce functional connectivity artifactually and lead to incorrect conclusions62- 64. Major sources of structured temporal noise include subject movement, subject physiology, and the physics of MRI. It is useful to consider the fMRI timeseries as containing fluctuations arising from 1) the BOLD signals of interest, 2) structured temporal noise, 3) and unstructured random ‘Gaussian’ temporal noise11. Temporal cleanup aims to remove temporal noise as selectively as possible while minimally impacting the BOLD signals of interest. The HCP devised a method based on Independent Components Analysis (ICA,65, which splits the fMRI timeseries into structured components and unstructured noise, representing all of the timeseries fluctuations. The structured components are then categorized by a machine learning classifier (FIX) into signal and noise7, and the noise components are removed by regressing them out of the data. For HCP data, the ICA+FIX approach removes spatially specific (i.e. affecting only part of the brain in a particular way) structured noise components and retains spatially specific BOLD signal components (i.e. resting state networks, RSNs) with better than 99% accuracy7. In contrast, traditional approaches that include bandpass filtering, WM and CSF regression, and censoring of high-motion frames63,64) are not as selectively at targeting structured noise while preserving BOLD signal. However, a controversy remains about "global" fluctuations in fMRI (see Box 1 and also Supplementary Topic #9 and Supplementary Figs. 5 and 6), as the spatial ICA+FIX approach is not designed to separate these ‘global’ fluctuations in the average fMRI signal into signal and noise. For dMRI, subject motion and physiology can cause signal loss. Such ‘outliers’ can be detected using a new approach that builds a model of the diffusion signal from all of the data, allowing artifactual measurements to be identified and replaced with estimates of what the signal should have been (see Supplementary Topic #8, 59-61). Box 1: Global fMRI Temporal Fluctuations. A hotly debated topic is how best to handle ‘global’ signals and artifacts present in fMRI data62,63,105-107. The core phenomenon involves slow and spatially widespread fluctuations in the fMRI timeseries. The average timeseries across the brain includes both neural signal (e.g. the average of all the resting state networks) and structured noise (e.g. physiological noise), and the debate centers on whether the average timeseries should be regressed out of the data to remove the effects of global noise62 or whether it should be preserved to avoid distorting neural signals by removing their average. Additive global noise will by definition artifactually increase ‘full’ correlations across the brain (especially problematic if it differs across groups), but removal of a global neural signal will artifactually decrease correlations across the brain and, more problematically, distort the pattern of correlations for particularly large/strong networks62,108,109.     What is needed to get past the current impasse is a method that selectively separates global signal and global noise. Though no such method is yet available, we offer several observations about global fMRI fluctuations. 1) In HCP data, the global signal (after ICA+FIX cleanup) is primarily a gray matter signal, with a near zero amplitude in white matter voxels and low-level anti-correlation with CSF (Supplementary Fig. 5). 2) Within grey matter, the mean grey time course has much higher amplitude in sensory regions (e.g. visual, auditory, vestibular, and somatosensory cortex, plus the lateral geniculate   6 nucleus and other thalamic nuclei) and is lower in the cerebellum and many neocortical cognitive regions (Supplementary Fig. 6). Thus, the global signal is not uniform across the brain, or even across grey matter, and these observations constrain the possible sources of the global fluctuations: (i) Physiological effects. Substantial fMRI timeseries variance is attributable to physiological variables, including rate and depth of breathing, heart rate, and end tidal pCO2110-112). Indeed, any physiological process that affects global perfusion pressure of the brain (i.e. the above three, or even yawning) may transiently alter the amount of deoxyhemoglobin in the brain and contribute to global, physiologically artifactual fluctuations of the BOLD signal. Different brain areas and tissue types may be differentially sensitive to such changes in perfusion pressure because of differing metabolic rates, which may explain regional differences in global signal amplitude. (ii) Head movements (direct, biophysical). Transient head movements are sometimes followed by global fluctuations in the fMRI signal, leading to suggestions that head movements could directly cause these fluctuations62,64,106. However, head movements are unlikely to cause global fluctuations directly, given the lack of a plausible biophysical mechanism for how head movement might induce tissue-specific and brain-area-specific effects on the fMRI timeseries, let alone globally uniform changes in signal intensity short of complete head coil exit and reentry (See Supplementary Topic #9). (iii) Head movements (indirect, neural). Head movements may induce neural responses in somatosensory, vestibular, and/or visual systems that may in turn modulate wider brain activity in a pattern very consistent with that mentioned above. (iv) Widespread/High amplitude RSNs, average of the RSNs. As discussed above, a portion of the global fluctuations will represent the average of the RSNs with particular weighting to more widespread or stronger networks. It will be important to separate global physiological effects from spatially widespread neural effects (using either extrinsic physiological monitoring or intrinsic fMRI-data-driven means yet to be developed). Short of a complete solution to the problem of global fluctuations, we note that some approaches to analysis (such as the use of partial correlation when estimating functional connectivity, and multi-variate measures in general) are much more immune to the effects of global fluctuations than others (e.g., full correlation measures and uni-variate measures in general)113.   The traditional approach to reducing unstructured Gaussian noise involves spatial and temporal smoothing, which unfortunately affect both signal and noise unselectively. A common problem in traditional neuroimaging is that smoothing mixes the grey matter functional signals of a given brain area with signals from white matter, CSF, and nearby brain areas, markedly reducing the spatial fidelity of neuroimaging data. Parcellation (tenet 6) enables averaging of data without mixing signal across different tissue types or between different brain areas in so far as the parcellation is neurobiologically accurate, and acts as a neuroanatomically informed method of nonlinear smoothing. Another option is to selectively smooth unstructured noise more than structured BOLD signal (see Supplementary Topic #10). This enables reduction of certain artifacts (see Supplementary Fig. 7) and enhancement of SNR without concomitant spatial blurring and without invoking a brain parcellation.   7 Preserving precision from voxels to CIFTI “grayordinates.” It is vital that the preprocessing and analysis techniques used on high resolution HCP-style data maintain this hard-won spatial resolution. Besides removing distortions and accurately aligning data across modalities within each subject, the data must be put into an analysis framework that enables accurate cross- subject and cross-study comparisons (tenet 5 below). The traditional neuroimaging paradigm represents spatial data in a regular 3D array of voxels (an image volume) that reflects the output of the MRI image reconstruction process. Because the cerebral cortex has a sheet-like folded geometry, it is advantageous to visualize and analyze cortical data on surface mesh vertices, instead of volume voxels15,25,26,38. However, subcortical grey matter structures are mostly globular in shape and remain best represented using volume voxels. To represent the inherently dual geometry of the brain, the HCP developed the CIFTI file format6, which integrates surface vertices and subcortical voxels into a single file and represents each major brain gray matter structure using the appropriate geometry (see Figure 1). CIFTI gray matter vertices and voxels are called ‘grayordinates.’ Analyses focusing only on grayordinates (e.g., fMRI) dramatically reduce file sizes and computational burdens (see Figure 1 legend). For dMRI, surface representations allow more accurate anatomical constraints in tractography (e.g., the white surface can be used as the destination for tractography). The final common outputs of the HCP preprocessing and analysis Figure 1. Combined representation of cortical surface vertices and pipelines are datasets in the CIFTI standard space6. CIFTI is natively subcortical voxels in the CIFTI grayordinates standard space. Left and right cerebral cortices contribute about 30k surface vertices each (fewer supported by Connectome than the 32k vertex standard meshes for each hemisphere, because Workbench, and other tools are the non-cortical medial wall is not included). Additionally, 19 subcortical currently implementing CIFTI grey matter structures combine to contribute about 30k volume voxels. compatibility (e.g., FSL, FreeSurfer, In total, there are 91,282 grayordinates corresponding to all of the grey FieldTrip, nibabel). In particular, matter sampled at a 2 mm average vertex spacing on the surface and FSL’s PALM software enables as 2 mm voxels subcortically. The HCP’s minimal preprocessing permutation-based statistical pipelines ensure that each subject has 91,282 aligned grayordinates, inference on CIFTI datasets thereby facilitating cross-subject comparisons of data within this coordinate system. This entails more than a two-fold reduction in file (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PA LM 66). The CIFTI grayordinates size relative to the >200,000 voxels needed for an equivalent 2 mm isotropic volume representation. For the 1.6mm 7T HCP data, we have developed a 1.6mm standard grayordinates space with 170,494 grayordinates using “59k” surface meshes. Reproduced with   permiss8io  n from Reference 6. analysis framework allows each brain structure to be registered using the approach best suited to it (see Tenet 5). Tenet 5: Accurately Aligning Brain Areas Across Subjects and Studies. Cross-subject comparisons. Analyses that combine neuroimaging data obtained from different individuals benefit greatly from accurately aligning corresponding brain areas, so that like is compared with like14, Alignment of subcortical and cortical structures is best achieved through differing means. Subcortical structures are less variable in size, shape, and position across subjects, and reasonable alignment is achieved by nonlinear 3D registration to a standard template, as in the traditional paradigm. The HCP currently uses FSL’s FNIRT for nonlinear registration of T1w images to MNI space, but registrations that also incorporate dMRI fiber orientations in the white matter may improve alignment of subcortical white matter fiber tracts (See Supplementary Topic #1167,68). For human cerebral cortex, the complexity and variability of folding patterns and of areas relative to folds38,69-71 poses a profound challenge that is an ill-posed problem for nonlinear 3D registration. Though alternative approaches such as hyperalignment (where alignment is not spatially constrained) are useful in some contexts, many neuroscientific questions are best addressed by aligning subjects to a common space. Registering cortical surfaces rather than brain volumes enables better alignment because accurate cortical segmentation of each subject reduces the dimensionality of the registration problem from 3D to 2D—alignment need only occur tangential to the cortical sheet as opposed to a 3D alignment of the folded sheet itself26,38,72-76. Though surface-based registration of the cerebral cortex to an atlas is not a part of the traditional paradigm, it has been adopted by a growing number of investigators. Most use surface registration based on cortical folding patterns26,38,72-76, which achieves ‘geographic correspondence’ in regions where folding patterns are consistent across individuals and ‘functional correspondence’ in regions where areal boundaries are consistently correlated with folding. Because cortical folding patterns are highly variable in many regions of cortex and their relationships to cortical areas are similarly variable, folding patterns alone are not enough to achieve alignment of (functional) cortical areas across most of cerebral cortex. Registration needs instead to be driven by features more closely tied to cortical areas, such as architecture, function, connectivity, and topography77. Such areal-feature-based cortical surface registration can dramatically improve the sharpness of group average maps and enhance group statistics, as it largely compensates for differences across subjects in cortical areal size, shape, and position. Indeed, alignment based on T1w/T2w myelin maps, resting state network maps, and visuotopic connectivity maps yields sharper task fMRI contrast maps with higher statistics, even though the task fMRI data were not used in the registration (Figure 2)8,9. Areal-feature-based aligned data was recently released by the HCP (dubbed ‘MSMAll’ because it uses the Multi- modal Surface Matching algorithm,9 and all of the useful modalities listed above). We believe that many studies will benefit from using areal-feature-based surface registration as a starting point, even if they choose to make final comparisons across subjects using other means (e.g. hyperalignment78, or individual subject parcellation (see below and ref.8). Cross-study comparisons. Traditionally, cross-study comparisons based on 3D stereotaxic coordinates were assumed to align if everyone used a common standard volume space (e.g., MNI). These 3D coordinates were typically derived from thresholded statistical maps of the ‘significant’ activations (or connectivity seed locations). Though such standard space coordinates allow for gross comparisons across studies at the gyral level (e.g., ref.79, questions   9 of whether a particular cortical area was activated in particular task contrast typically reveal clouds of points whose distribution may be suggestive, but often inconclusive (e.g., ref.19). Stereotaxic coordinates have high spatial uncertainty (relatively small changes in the thresholded statistical maps used to produce them can have large effects on the resulting coordinate locations, see Supplementary Topic #12 and Supplementary Fig. 8), and they do not contain information about the spatial extent or boundaries of activation. Conclusive cross-study comparisons require assessment of the overlap of activated regions and whether their boundaries align or not. Cross-study comparisons also rely on the absence of any “drift” in alignment between the studies (see Supplementary Topic #13 and Supplementary Fig. 9). When drift is eliminated and boundaries are compared, direct and conclusive assessments of cross-study agreement are possible (see Supplementary Fig. 10, also Figure 6). F igure 2. Improved intersubject registration using information based on areal features in addition to cortical f olding. The top row shows group task-fMRI z-stat maps (“Story vs Baseline” contrast from the language task o n the left and the two-back vs zero-back contrast in the working memory task on the right) from 120 Q1 and Q2 HCP subjects after intersubject registration using the Multimodal Surface Matching (MSM) method constrained only by folding (FreeSurfer’s ‘sulc’ maps). Bottom row shows sharper group task-fMRI maps and higher z-statistics when using resting-state networks (RSN) along with myelin maps to constrain the registration. For the data in this figure and figure 6, subject recruitment procedures and informed consent forms, including consent to share de-identified data, were approved by the Washington University institutional review board.  Data at http://balsa.wustl.edu/97V4.     10