Thirty years of SPM-BrainMap synergy: making and mining coordinate-based literature

Abstract

Statistical Parametric Mapping (SPM) adheres to rigorous methodological standards, including: spatial normalization, inter-subject averaging, voxel-wise contrasts, and coordinate reporting. This rigor ensures that a thematically diverse literature is amenable to meta-analysis. BrainMap is a community database (www.brainmap.org; www.portal.brainmap.org) launched contemporaneously with SPM with the goal of efficiently sharing the results and methods of the literature compliant with SPM standards. The SPM-BrainMap symbiosis has motivated the development of coordinate-based meta-analytic methods and a substantial literature of secondary analyses. Collectively this corpus constitutes system-level probabilistic maps and models of the human brain, which details its functional organization, network architecture, and alterations by disease.

In “SPM—30 years and beyond”, Zeidman and colleagues—all prolific contributors to SPM’s panoply of analytic tools—laud SPM’s early adoption of an open-source development environment and the myriad technical advances achieved thereby (Zeidman et al. 2025). This commentary, by contrast, emphasizes SPM’s role in the creation of a truly massive, thematically broad literature that is nonetheless well-standardized and ripe for meta-analysis. SPM’s unprecedented scope of application was achieved by freely distributing an intuitive image-analysis suite to a multidisciplinary but largely nontechnical userbase. Methodological coherence was achieved by enforcing a common neuroanatomical space and outputting experimental results in a simple, standardized manner: x-y-z addresses of local maxima of significant regional change. Coordinate-based meta-analysis (CBMA) synthesizes this literature by assessing effect reproducibility, often discovering unanticipated convergences (reliable neurobiological effects) and, thereby, generating new hypotheses. A persistent shortcoming of SPM (and other neuroimaging analytics) is weak correction for multiple comparisons, leading to high rates of false-positive findings. A key role for CBMA, then, is explicitly identifying absences of replicable effects. A third contribution of CBMA is meta-connectomics: the discovery of brain network properties through cross-study, multiregional, co-occurrence patterns. In these ways, SPM and CBMA are symbiotic, jointly making and mining the extraordinary wealth of the neuroimaging literature (Fox et al. 1998, 2014; Fox and Friston 2012).

From SPM’s inception, Friston and colleagues did an impressive job of selecting existing and emerging neuroanatomical reporting standards and neuroimaging analysis strategies for adoption and amplification by SPM (Friston et al. 1994; Ashburner 2012). A significant part of the early genius of SPM was hardening, extending and “democratizing” robust constructs that were published, but not yet accessible to the community at large nor fully refined statistically and mathematically. The most fundamental of these standards were: spatial normalization, intersubject image-averaging, voxel-wise conditional contrasts, and local-maxima reporting. All four constructs were introduced to tomographic neuroimaging by the Washington University laboratory of Michel Ter-Pogossian and Marcus Raichle and rapidly adopted and extended by the fledgling neuroscience imaging community. Prior to SPM, though, none were provided as freely available, developer-supported software.

Spatial normalization—morphing neuroimaging data from “native space” (scanner-specific data addresses) into a common, atlas-referenced space—was inspired and informed by stereotactic neurosurgery. Invented at the advent of the 20^th century as a research tool (Horsley and Clarke 1908), stereotactic surgery was in widespread clinical use by the 1950s (Rahman et al. 2009). Stereotactic atlases intended to advance coordinate-reporting research followed (Talairach et al. 1967; Schaltenbrand and Warren 1977; Talairach and Tournoux 1988). Talairach’s 1967 atlas, a magnum opus, introduced the concept of probabilistic mapping, reporting spatial-probability contours for major sulci and a host of telencephalic functional locations as determined by intra-operative stimulation mapping. The 1967 atlas was the first to be used as a reference space for reporting functional neuroimaging results (Fox et al. 1985) and subsequently for inter-subject image averaging (Fox et al. 1986, 1987, 1988; Friston et al. 1989; Friston et al. 1995). Talairach’s 1988 atlas was specifically designed for imaging applications. One impactful, well-cited derivative was an automated anatomical labeling daemon (Lancaster et al. 2000). Overall, the 1988 atlas was embraced by the SPM team and the neuroimaging community at large; to date it has been cited > 23,000 times, reflecting its enormous impact on neuroscience imaging.

Conditional contrasts of functional neuroimages were introduced by Lassen et al. (1978), displaying blood-flow data as task-induced percent change from rest. In this format, task-engaged regions appeared as “hot spots” of increased blood flow against a null background. Fox and colleagues adapted this “subtractive” format to standard-space mapping, performing voxel-wise, whole-brain contrasts (Fox et al. 1986; Petersen et al. 1988). The universally recognized SPM maximum-intensity projection within a bounding coordinate grid was an elegant refinement of Lassen’s format. From within this sparse-format data array, activation foci could be auto-extracted as local maxima and reported as x-y-z addresses of significant changes (Fox et al. 1988; Fox and Mintun 1989; Mintun et al. 1989). These display, analysis and reporting conventions were eagerly adopted by both the Positron Emission Tomography (PET) (Friston et al. 1991a, 1991b; Grafton et al. 1992) and functional Magnetic Resonance Imaging (fMRI) communities (Belliveau et al. 1991). This motivated their incorporation into SPM from its 1st release onward (Friston et al. 1994), solidifying these as standards for the field.

Early analysis methods created difference maps of physiological (blood flow) or physical (radiation counts) parameters (Fox et al. 1984, 1988, 1989). Change distributions of these parameters were analyzed using omnibus outlier detection metrics which tested the general hypothesis of a neurobiological difference between contrasted conditions. Local maxima were assigned statistical significance post-hoc relative to that change distribution, typically thresholding to Z-score ≥ 4. Challenging this approach, Friston argued that omnibus testing of effects at the change-distribution level was insufficient; rather, voxel-level significance testing with correction for multiple comparisons was needed (Friston et al. 1991). (Presumably each voxel was a unique hypothesis?) The severity of correcting for > 100,000 individual voxels was mitigated by incorporating image smoothness, which has ultimately proved insufficient. To this day, high false-positive rates remain problem for the field (Eklund et al. 2016). In this context, Eklund sagely observed: “meta-analysis can play an important role in teasing apart false positive findings from consistent results” (ibid pg 7904).

The BrainMap Database was launched in 1992 as an open-access, on-line, community database (client–server model) of the coordinate-reporting, task-activation (TA) neuroimaging literature (www.brainmap.org). To guide and promote this initiative, BrainMap workshops were held annually from 1992 to 1998, attended by leaders of all major neuroimaging laboratories, notably including Karl Friston, Keith Worsley, and Richard Frackowiak (Gibbons 1992; Fox and Lancaster 1994). As spatial normalization and coordinate reporting had been rapidly adopted by the international neuroimaging community, the logistical goal of BrainMap was to efficiently share the results (activation loci with statistical parameters) and methods (as coded experimental meta-data) of the peer-reviewed literature. BrainMap’s input data were rigorously quality controlled, ensuring compliance with the accepted standards (Fox and Lancaster 2002). BrainMap’s intended scientific impact was to inform and constrain the interpretation, design and analysis of future studies by ready access to those prior (Laird et al. 2005). A secondary, longer-term goal was to stimulate the development of formal meta-analytic methods for coordinate-based neuroimaging.

CBMA methods emerged and evolved as the coordinate-reporting corpus expanded. First to appear were quantitative (mean and variance), consensus localizations of reliably activated functions (Paus 1996; Picard and Strick 1996). This approach was refined into modeling functional zones as spatial probability contours scaled by the volume (n of studies and subjects) of input data (Fox et al. 1997, 1998, 1999, 2001). Collectively these early methods confirmed that brain functional areas were reliably located within Talairach 3D space, and provided a growing library of functionally and spatially characterized brain regions.

Activation Likelihood estimation (ALE) (Turkeltaub et al. 2002) moved CBMA a quantum leap forward. ALE combines neuroimaging results from multiple published studies to quantify interstudy concordance of those results. Input data are the array of activation-location coordinates from any number of conceptually related studies (eg, speech production studies). Location probability is modeled for each included activation focus (scaled by sample size). The union of these spatial probability distributions yields an ALE map for the brain. Significance was initially tested by permutation analysis of randomly generated sets of foci. Subsequently, improvements to the statistical model have enhanced the rigor of ALE, with only minor alterations to the core computation (Eickhoff et al. 2009; Turkeltaub et al. 2012). To promote community adoption of ALE CBMA, a virtual workshop was organized in which prominent neuroimaging investigators were assisted in performing ALE meta-analyses on coordinate-reporting literatures applying task paradigms they had introduced. The results were published in a special issue of Human Brain Mapping (Volume 25, Number 1, 2005; Fox et al. 2005a). Also reported in this issue was an expanded BrainMap taxonomy of experiment design (Fox et al. 2005b), social evolution of the BrainMap community (Laird et al. 2005) and meta-analytic network modeling strategies (Lancaster et al. 2005; Neumann et al. 2005). With the introduction of voxel-based morphometry (VBM; Ashburner and Friston 2002) a coordinate-reporting VBM literature emerged. As the reporting format and statistical assumptions for task and VBM data are the same, no change in the ALE algorithm was required. However, application of ALE to VBM is often termed Anatomical Likelihood Estimation or more generally Alteration Likelihood Estimation. As the VBM literature grew, BrainMap was expanded to include a VBM database sector (Vanasse et al. 2018). As anticipated, ALE has identified both highly impactful cross-study concordances (Laird et al. 2008; Kurth et al. 2010; Goodkind et al. 2015) and equally important failures to converge (Muller et al. 2017; Henn et al. 2023). At this time, ALE is by far the most widely adopted CBMA method, with > 1800 citations of the original publication and still more of the follow-on methods refinements.

Meta-connectomic CBMA began modestly as within-paradigm, co-occurrence network analyses using ALE-derived convergences as nodes (Lancaster et al. 2005; Neumann et al. 2005, 2010.) This was soon extended to region-to-whole-brain functional co-occurrence analyses using large collections of studies without constraints on activation-paradigm type (Postuma and Dagher 2006; Robinson et al. 2010, 2012). Independent components analysis (ICA) also proved applicable to entire BrainMap sectors, both functional and structural. Applying ICA to the BrainMap task sector, Smith demonstrated the correspondence of the functional architectures of task and rest (Smith et al. 2009). In a like manner, Vanasse applied ICA to the BrainMap VBM sector, finding a clear demarcation between neurological and psychiatric disorders in network architecture (Vanasse et al. 2018). Extending this line of inquiry to a comparison of the task and VBM sectors, Vanasse et al. (2021) found highly corresponding canonical architectures in both (Fig. 1). Jointly these ICA studies indicate that network architecture is a brain-organizational motif constraining both function and structure. Further, these studies suggest that structural pathology of both neurological and psychiatric disorders recapitulates the normal, physiological network architecture observed in both rest and task states. That is, they demonstrate that network degeneration characterizes far more disorders than the severe dementing neurodegenerations noted by Seeley when proposing the Network Degeneration Hypothesis (Seeley et al. 2009). However, a striking property of the Vanasse analyses was that no disorder loaded on a single network and no network loaded on a single disorder. Rather, each disorder loaded on canonical networks in a unique pattern. This argues that canonical networks do not adequately characterize the network architecture of individual disorders. To this end, using CBMA to generate disorder-specific network models as candidate biomarkers seems a promising way forward (Glahn et al. 2008; Barron et al. 2013; Chiang et al. 2019; Gray et al. 2020). Disorder networks models are amenable to individual fitting, providing a sophisticated strategy for personalized targeting of therapeutics, including neuromodulation and stereotactic surgery. Out-of-sample validations in resting-state fMRI can assay model goodness of fit at the group and individual levels (Barron et al. 2015; Chiang et al. 2021; Gray et al. 2023, 2025).

Fig. 1.

Co-activation and TA network correspondence (D = 20). (left) thirteen TA networks derived from the BrainMap-TA sector (n = 7865 TA experiments among healthy subjects). (right) fourteen structural co-alteration networks derived from the BrainMap voxel-based morphometry sector (n = 2002 experiments across n > 40 diseases). Clear network matches across datasets are shown by connecting bars, the width of which are proportional to the whole-brain, spatial correlation coefficient. Some networks matched to two, separate opposite-modality networks (red and green). Color scale is shown at the bottom of the image. Independent component numbers are ordered by explained variance within the respective dataset. All component matches are at or below a significance threshold of P = 0.01, family-wise error rate corrected. Reproduced with permission from Vanasse et al. (2021).

In conclusion, the neuroanatomically standardized literature generated by SPM over the past 30 years has served as source data for secondary analyses—CBMA—that collectively constitute system-level probabilistic maps and models of the human brain characterizing its functional organization, its network architecture, and its alterations by disease at the systems level. This is a remarkably productive symbiosis. Long may it continue.

Biography

Peter T. Fox is Director of the Research Imaging Institute and Professor of Radiology, Neurology, Psychiatry and Physiology at UT Health Science Center, San Antonio, Texas. Clinically trained in Neurology, Fox develops and applies noninvasive neuroimaging methods to map human brain function, structure, networks and their disease-related alterations. He developed the BrainMap databases (brainmap.org; portal.brainmap.org), was founding editor-in-chief of Human Brain Mapping, and received the Glass Brain Award from the Organization for Human Brain Mapping.