{
  "id": "functional-magnetic-resonance-imaging",
  "title": "Functional Magnetic Resonance Imaging",
  "lang": "en",
  "source_lang": "en",
  "summary": "Functional magnetic resonance imaging maps brain activity indirectly through changes in blood oxygenation, blood flow, and related physiological signals.",
  "category": "health",
  "subcategory": "medicine",
  "type": "concept",
  "aliases": [
    "fMRI",
    "functional MRI",
    "functional MR imaging"
  ],
  "generated_at": "2026-11-02",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/functional-magnetic-resonance-imaging/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/functional-magnetic-resonance-imaging/",
      "json": "https://aiwiki.page/zh/functional-magnetic-resonance-imaging.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Functional magnetic resonance imaging (fMRI) is a family of magnetic resonance imaging methods used to investigate changes in physiological processes associated with activity in the brain. Its most widely used form measures blood-oxygen-level-dependent, or BOLD, contrast: small changes in the magnetic resonance signal caused by variations in blood oxygenation. By repeatedly imaging the brain during tasks or at rest, fMRI reveals patterns associated with perception, movement, memory, language, and other functions.\n\nUnlike methods that record electrical or magnetic activity directly, fMRI usually measures a vascular response to neural activity. This indirect relationship is central to both its usefulness and its limitations. Functional images can cover most or all of the brain with millimeter-scale spatial sampling, but the underlying blood response develops over seconds and is influenced by vascular physiology as well as neural events.\n\nSince the early 1990s, fMRI has become a major instrument of neuroscience and an established component of selected clinical brain-mapping procedures. Its interpretation depends on experimental design, image acquisition, statistical modeling, and knowledge of the biological processes connecting neural activity to blood circulation."
    },
    {
      "heading": "Historical foundations and development",
      "text": "The physiological basis of fMRI predates magnetic resonance imaging by about a century. In 1890, Charles Roy and Charles Sherrington reported experiments supporting a relationship between cerebral activity and local blood supply. Their work helped establish the principle that the brain regulates its circulation locally rather than receiving an entirely uniform supply. Later research clarified that active tissue can receive an increase in blood flow that exceeds the accompanying increase in oxygen consumption.\n\nA second foundation was the magnetic behavior of hemoglobin, the oxygen-carrying protein in red blood cells. In 1936, Linus Pauling and Charles Coryell demonstrated the different magnetic properties of oxygenated and deoxygenated hemoglobin. Deoxygenated hemoglobin is paramagnetic and alters the magnetic field around blood vessels. These effects eventually provided an intrinsic source of image contrast, without requiring an injected substance.\n\nThe development of spatially encoded magnetic resonance imaging during the 1970s made it possible to locate magnetic resonance signals within the body. Rapid acquisition methods were particularly important for functional imaging. Peter Mansfield’s work on echo-planar imaging provided a way to acquire an image after a single excitation, allowing repeated measurements on a timescale suitable for following hemodynamic changes. Early scanners and computing systems imposed substantial limitations, but improvements in gradient hardware, radiofrequency coils, and reconstruction progressively made time-series imaging practical.\n\nIn 1990, Seiji Ogawa and colleagues described oxygenation-sensitive contrast in animal magnetic resonance images and introduced the BOLD framework. In 1991, John Belliveau and colleagues demonstrated human functional mapping using an injected contrast agent to measure changes in cerebral blood volume during visual stimulation. In 1992, groups including those led by Kenneth Kwong and Ogawa, as well as Peter Bandettini and colleagues, published human functional imaging results using intrinsic magnetic resonance contrast. These studies helped establish noninvasive functional mapping without repeated contrast injections.\n\nThe field subsequently expanded from relatively simple visual and motor experiments to cognitive tasks, clinical mapping, and large-scale network analysis. Biswal and colleagues’ 1995 report of correlated low-frequency fluctuations in motor regions helped establish resting-state fMRI. Later developments included faster whole-brain acquisition, high-field imaging, standardized data-sharing practices, and population-scale projects. Functional imaging consequently evolved from a technique for detecting localized task responses into a broader approach for studying distributed brain organization, individual variation, and interactions between neural and vascular systems."
    },
    {
      "heading": "BOLD contrast and neurovascular coupling",
      "text": "BOLD imaging depends on differences in magnetic susceptibility between oxygenated and deoxygenated blood. Deoxygenated hemoglobin creates microscopic field variations around vessels, accelerating the loss of phase coherence among nearby nuclear spins. In a commonly used gradient-echo acquisition, this affects the effective transverse relaxation time, \\(T_2^*\\). A simplified signal relationship is\n\n\\[S(TE)=S_0 e^{-TE/T_2^*},\\]\n\nwhere \\(TE\\) is the echo time and \\(S_0\\) represents the signal before this effective transverse decay. Changes in deoxyhemoglobin concentration therefore alter the measured signal, even when the amount of tissue is unchanged.\n\nThe connection between neural activity and vascular response is called neurovascular coupling. Increased synaptic activity changes metabolic demand and engages signaling pathways involving neurons, astrocytes, vascular smooth muscle, and other components of the neurovascular unit. Local vessels dilate, increasing cerebral blood flow. During a typical positive BOLD response, this increase exceeds the rise in oxygen extraction, reducing the local concentration of deoxyhemoglobin and increasing the gradient-echo signal. The response reflects the combined effects of blood flow, blood volume, and oxygen metabolism rather than any one of them independently.\n\nBOLD signal changes are commonly only a few percent of the baseline signal in conventional human experiments, although their magnitude varies with field strength, brain region, task, and acquisition method. After a brief neural event, a typical response begins after a delay, reaches a peak roughly four to six seconds later, and then returns toward baseline. A post-stimulus undershoot may follow. This time course, termed the hemodynamic response function, is not identical across individuals, regions, or physiological conditions.\n\nExperiments combining electrophysiology and fMRI have shown that BOLD responses often relate strongly to local field potentials, which reflect aggregate synaptic and dendritic processing. The signal is not a direct count of action potentials. Excitatory and inhibitory activity both require energy and can both contribute to vascular responses. Consequently, an increase in BOLD signal cannot simply be translated into “more excitation,” nor can a decrease automatically be interpreted as neural inhibition.\n\nNegative BOLD responses are also biologically heterogeneous. They may accompany reduced neural activity, altered blood supply, or changes in the balance between oxygen delivery and consumption. Baseline vascular state matters: age, arterial carbon dioxide, medications, and vascular disease can alter response magnitude or timing. Measurements of carbon dioxide, perfusion, and cerebrovascular reactivity can help distinguish some of these influences, but conventional BOLD imaging alone does not provide an absolute measurement of neural activity or cerebral oxygen consumption."
    },
    {
      "heading": "Image acquisition and functional imaging methods",
      "text": "A functional experiment ordinarily collects a series of three-dimensional brain volumes, each assembled from multiple slices or acquired using a three-dimensional sequence. Many studies use scanners operating at 3 tesla and gradient-echo echo-planar imaging. Conventional whole-brain acquisitions often sample voxels approximately 2–3 millimeters across every one to three seconds, although protocols vary widely. Modern accelerated methods can obtain substantially shorter repetition times, while specialized high-resolution studies may use much smaller voxels with more restricted coverage.\n\nAcquisition involves trade-offs among spatial resolution, temporal sampling, coverage, and signal quality. Smaller voxels contain fewer contributing spins and therefore generally yield less signal. Faster imaging can reduce the time available for signal recovery or require acceleration that increases noise and reconstruction artifacts. Increasing the magnetic field can improve sensitivity and support finer spatial sampling, but it also increases some susceptibility effects and makes radiofrequency field uniformity and energy deposition more challenging.\n\nEcho-planar imaging is especially vulnerable to magnetic field inhomogeneity. Near air-filled sinuses and the ear canals, susceptibility differences can produce geometric distortion or signal loss in regions such as the orbitofrontal cortex and inferior temporal lobes. Field maps or acquisitions with opposite phase-encoding directions can support distortion correction, but severe signal loss cannot simply be reconstructed from missing information. Gradient-echo BOLD is also sensitive to draining veins, which may displace the measured response from its neural origin. Spin-echo approaches can offer different vascular weighting, particularly at high field strength, though usually with other sensitivity and acquisition trade-offs.\n\nBOLD is not the only functional magnetic resonance method. Arterial spin labeling magnetically labels incoming arterial blood water and compares labeled and control images to estimate tissue perfusion. It can provide quantitative cerebral blood flow, usually expressed in milliliters of blood per 100 grams of tissue per minute. Its signal is generally weaker than conventional BOLD, but its physiological interpretation and baseline stability make it useful for particular experimental and clinical questions. Vascular-space-occupancy methods instead emphasize changes related to cerebral blood volume.\n\nOther acquisition developments include simultaneous multislice imaging, which excites and separates several slices at once, and multi-echo imaging, which records signals at several echo times. Multi-echo data help distinguish components with BOLD-like echo-time dependence from some non-BOLD fluctuations. At very high field strength, including 7 tesla, submillimeter acquisitions can investigate variation across cortical depth. Such measurements do not automatically isolate individual cortical layers, because vascular anatomy, partial-volume mixing, and image alignment still influence the observed signal."
    },
    {
      "heading": "Task-based experiments and experimental design",
      "text": "Task-based fMRI examines signal changes associated with experimentally controlled events or conditions. A participant may view images, move a finger, remember a sequence, make a decision, or listen to speech. The scientific question is usually expressed as a comparison: viewing faces versus objects, recalling information versus a matched control task, or responding to increasing memory load. A functional map therefore represents a specified contrast rather than a complete inventory of all brain activity occurring during the experiment.\n\nTwo major designs are block and event-related designs. A block design groups repeated trials of the same condition into periods often lasting tens of seconds. The sustained response can produce strong statistical sensitivity, making this design useful for robust sensory, motor, or clinical mapping. However, long blocks may combine the process of interest with changes in expectation, strategy, adaptation, or sustained attention. The selection of a comparison condition is therefore as important as the selection of the main task.\n\nEvent-related designs model individual trials, allowing different trial types to be intermixed and sorted according to behavior. For example, successfully remembered items can be compared with forgotten items, or correct responses with errors. Because vascular responses overlap in time, the sequence and spacing of events matter. Variable intervals, often called temporal jitter, can improve the separation of modeled responses. Rapid event-related experiments exploit approximate linearity of the hemodynamic response, although this approximation becomes less reliable under some closely spaced or strongly adapting stimulation conditions.\n\nA typical analysis predicts the measured signal by convolving a sequence of experimental events with a hemodynamic response function:\n\n\\[x(t)=(u*h)(t),\\]\n\nwhere \\(u(t)\\) describes event timing and \\(h(t)\\) represents the assumed vascular response. Separate predictors can represent different conditions, stimulus durations, or continuously varying properties such as difficulty. Behavioral measurements, including accuracy and reaction time, establish whether participants performed the intended task and can help distinguish alternative interpretations of a response.\n\nDesign also determines what can be inferred. A language task that is substantially harder than its control condition may reveal effort-related differences as well as language processing. Repeated exposure may change familiarity and learning. Head movement can become correlated with overt speech or motor responses. Studies address these problems through matched control conditions, training, counterbalancing, response monitoring, and models that represent relevant confounds. The scanner environment itself can alter performance through noise, restricted movement, and unusual viewing arrangements. These influences make task implementation and behavioral validation integral parts of the measurement rather than secondary procedural details."
    },
    {
      "heading": "Resting-state imaging and brain connectivity",
      "text": "Resting-state fMRI records spontaneous signal fluctuations without requiring a repeated experimental task. Participants may keep their eyes open, view a fixation point, or close their eyes, depending on the protocol. The resulting data are often used to estimate functional connectivity, meaning statistical dependence between signals from different brain locations. This dependence is commonly measured with correlation, although many other measures are used. Connectivity in this sense does not necessarily imply a direct anatomical connection or a causal interaction.\n\nMany resting-state analyses focus on slow BOLD fluctuations, often in a frequency range approximately between 0.01 and 0.1 hertz. Correlated patterns reveal networks encompassing visual, somatomotor, attentional, and other systems. The default mode network, including regions of medial prefrontal, posterior cingulate, and lateral parietal cortex, became particularly influential because its components show coherent fluctuations and often decrease their activity during externally demanding tasks. Neither resting-state networks nor the default mode network are inactive background structures; their signals vary with ongoing mental and physiological states.\n\nSeveral analytical strategies characterize these patterns. Seed-based analysis correlates a selected region’s time series with signals elsewhere. Independent component analysis separates statistically independent components that may correspond to neural networks, artifacts, or mixtures of both. Brain parcellation divides the brain into regions, allowing construction of a connectivity matrix whose entries describe pairwise relationships. Network measures can then quantify properties such as segregation into communities or integration across systems, but the results depend on parcellation, preprocessing, and the mathematical definition of an edge.\n\nResting-state measurements are particularly sensitive to confounding. Small head movements can increase some short-distance correlations while reducing some long-distance relationships. Respiration, cardiac activity, changes in arterial carbon dioxide, and fluctuations in alertness can produce spatially structured signals. Removing the average whole-brain signal changes the correlation structure and can introduce or alter negative correlations. Conversely, retaining it leaves a mixture of physiological and potentially neural contributions. There is no preprocessing choice that resolves every question independently of the study’s aims.\n\nThese methods have made it possible to study network organization in infants, patients with limited task performance, and large population cohorts. However, an absence of an explicit task does not make the acquisition psychologically standardized: participants may differ in thought content, drowsiness, and eye behavior. Estimates of individual connectivity also become more stable with additional usable data. Claims about rapid changes in connectivity require particular care because short analysis windows produce noisy estimates even when the underlying statistical relationships are constant. Resting-state fMRI therefore measures structured spontaneous dynamics under specified acquisition conditions, not an unqualified map of fixed brain wiring."
    },
    {
      "heading": "Preprocessing and statistical inference",
      "text": "Functional images require several processing steps before an experimental effect can be evaluated. These commonly include correction for head motion, correction of susceptibility-related distortion, alignment with a structural image, and transformation into a common anatomical coordinate system. Slice-timing correction may compensate for different acquisition times within a volume when relevant to the protocol. Spatial smoothing can improve signal-to-noise ratio and accommodate anatomical variation, but it also reduces spatial specificity and can blur signals across neighboring structures.\n\nMotion correction estimates changes in head position and resamples the images. It cannot fully repair all consequences of movement, because movement also changes susceptibility fields, spin history, and the mixture of tissues within a voxel. Analyses may therefore include movement estimates as nuisance variables, remove heavily contaminated time points, or use component-based artifact detection. Quality control examines both images and time series; an apparently plausible activation map does not demonstrate that the underlying data are artifact-free.\n\nTask analyses commonly use the general linear model, written as\n\n\\[\\mathbf{y}=\\mathbf{X}\\boldsymbol{\\beta}+\\boldsymbol{\\varepsilon}.\\]\n\nHere, \\(\\mathbf{y}\\) is a voxel’s measured time series, \\(\\mathbf{X}\\) contains experimental and nuisance predictors, \\(\\boldsymbol{\\beta}\\) contains their estimated coefficients, and \\(\\boldsymbol{\\varepsilon}\\) represents residual variation. Contrasts of the coefficients test particular hypotheses. Because successive fMRI measurements are temporally correlated, statistical procedures must address this dependence rather than treating every volume as an independent observation. Individual-level estimates can then enter group analyses that account for variation among participants.\n\nA whole-brain analysis may test tens or hundreds of thousands of voxels. Without multiple-testing correction, many apparently significant results can arise by chance. Family-wise error control limits the probability of at least one false positive within a specified family of tests; false discovery rate procedures instead control an expected proportion of false discoveries under their assumptions. Cluster-based inference evaluates spatially contiguous effects, but its validity depends on the statistical model and thresholding procedure. Eklund, Nichols, and Knutsson’s 2016 evaluation highlighted problems with some widely used parametric cluster-inference settings, encouraging greater use of validated methods and permutation testing.\n\nReliable inference also requires separation of discovery from evaluation. Selecting voxels because they show a strong effect and then estimating the same effect within those voxels can inflate the reported magnitude. Predictive analyses can similarly overestimate accuracy if information from test participants influences feature selection or preprocessing. Transparent reporting specifies exclusions, contrasts, thresholds, effect estimates, and uncertainty. Standardized formats such as the Brain Imaging Data Structure support reproducible organization, while sharing unthresholded statistical maps preserves information that a colored illustration alone cannot convey."
    },
    {
      "heading": "Spatial specificity, timing, and interpretive limits",
      "text": "The spatial resolution of an fMRI acquisition is not the same as the biological precision of its result. A voxel samples a volume containing large numbers of neurons, glia, capillaries, and larger vessels. Its signal reflects a mixture of local tissue properties and blood arriving from or draining nearby regions. Spatial smoothing, motion correction, and interpolation introduce additional mixing. A map displayed with sharp boundaries can therefore look more anatomically precise than the measurement warrants.\n\nTiming presents a different limitation. Electrical signaling occurs over milliseconds, whereas the BOLD response evolves over seconds. Rapid image sampling can characterize the vascular response more densely and improve noise modeling, but it does not remove the vascular delay. Under favorable experimental conditions, differences in response timing can be estimated with precision finer than the duration of the full response. Such estimates nevertheless depend on assumptions about regional vascular responses and cannot ordinarily establish the sequence of individual neural computations.\n\nA central inferential problem is reverse inference: concluding that a particular mental process occurred because a brain region associated with that process became active. A region implicated in memory may also participate in attention, decision-making, or perception. In probabilistic terms, observing a high probability of activation given a process does not by itself establish a high probability of that process given activation. The selectivity of the response and the prevalence of alternative explanations matter. Anatomical labels such as “fear center” or “reward center” obscure the distributed and multifunctional organization revealed by many experiments.\n\nStatistical significance also differs from biological magnitude. A small effect can be highly significant in a sufficiently large dataset, while a substantial but variable effect may remain uncertain in a small sample. Likewise, a significant effect in one condition and a nonsignificant effect in another does not demonstrate that the conditions differ; the difference must itself be tested. Group-average maps can conceal substantial individual variation in anatomy and functional organization, particularly near the boundaries of cortical areas.\n\nFinally, most conventional fMRI evidence is observational with respect to neural mechanisms. Changing a task can establish that the experimental manipulation influenced the measured signal, but activation alone does not show that the identified region is necessary or sufficient for the behavior. Lesion studies, intracranial recordings, and interventions such as transcranial magnetic stimulation provide complementary evidence. Likewise, statistical methods labeled “effective connectivity” estimate directed relationships within particular models; their conclusions depend on model specification and cannot be treated as direct observations of information flowing between neurons."
    },
    {
      "heading": "Neuroscientific applications and complementary measurements",
      "text": "Functional imaging has reshaped accounts of brain organization by demonstrating both specialized responses and distributed cooperation. Visual experiments map orderly representations of the visual field and characterize regions with differential responses to categories such as faces, scenes, and bodies. Motor experiments reveal somatotopic organization while also showing involvement of premotor, parietal, cerebellar, and subcortical systems. These findings link behavior to interacting circuits rather than assigning complex abilities to isolated locations.\n\nResearch on language illustrates this approach. Listening, reading, speaking, and comprehending engage partly overlapping networks whose activity depends on linguistic content, task demands, and prior experience. Memory studies distinguish contributions of medial temporal structures, prefrontal regions, and distributed cortical representations. Studies of pain similarly show that nociceptive stimulation, attention, expectation, and affect influence multiple brain systems. In each case, fMRI contributes a spatially organized account of responses, while behavioral experiments establish what the participant actually perceived or performed.\n\nMultivoxel pattern analysis examines information distributed across groups of voxels rather than only the average response in each region. A classifier may distinguish stimulus categories from patterns that show little difference in mean activity. Representational analyses compare the similarity structure of brain responses with properties of stimuli, behavior, or computational models. These approaches have connected fMRI with machine learning, but successful prediction does not necessarily identify the biological computation used by the brain. Models can exploit correlated stimulus features or acquisition artifacts unless the experimental and validation design excludes such alternatives.\n\nNaturalistic experiments use films, stories, and other extended stimuli to investigate responses under conditions closer to everyday perception. They can reveal shared temporal structure across participants and responses at multiple timescales. Their richness, however, creates difficulty in separating correlated features: a dramatic scene may change visual motion, sound, emotion, and attention simultaneously. Large datasets and detailed stimulus models expand the questions that can be studied, but do not eliminate the need for interpretable comparisons.\n\nComplementary techniques address aspects of neural function that BOLD imaging does not directly measure. Electroencephalography and magnetoencephalography provide millisecond-scale measurements of aggregate electrical activity, while functional MRI generally provides more detailed whole-brain spatial localization. Positron emission tomography can investigate molecular targets and metabolism using radiotracers, and magnetic resonance spectroscopy measures selected metabolites. Diffusion MRI characterizes tissue microstructure and supports estimates of white-matter pathways. Combining these approaches can constrain explanations more effectively than any single measurement, although alignment across distinct signals requires explicit assumptions about their biological origins and spatial scales."
    },
    {
      "heading": "Clinical mapping and translational research",
      "text": "One established clinical application is mapping function before selected neurosurgical procedures. In patients with brain tumors or drug-resistant epilepsy, task-based fMRI can help locate motor, sensory, or language-related responses relative to a proposed surgical region. The images are interpreted alongside structural imaging, clinical examination, and other investigations. They may contribute to surgical planning and selection of further mapping procedures, but a BOLD response is not equivalent to proof that tissue is indispensable for a function.\n\nClinical interpretation is particularly demanding near lesions. Tumors and vascular abnormalities can disrupt neurovascular coupling, producing reduced or absent BOLD responses even where functioning neural tissue remains. This phenomenon can create false reassurance if an apparently inactive region is assumed to lack function. Movement, impaired comprehension, altered anatomy, and difficulty performing tasks can further complicate the result. Direct electrical stimulation during appropriate neurosurgical procedures provides different evidence by testing the effects of perturbing tissue, and fMRI does not universally replace it.\n\nLanguage lateralization is another important application. Patterns of activation can help characterize the relative contribution of the two hemispheres and can inform evaluation in selected epilepsy surgery candidates. Interpretation depends on the language task, analysis method, and clinical population. Memory fMRI has also been studied for estimating postoperative cognitive outcomes, but the strength and generalizability of predictions vary. Resting-state approaches are being investigated for patients unable to perform reliable tasks, although their clinical interpretation requires careful separation of functional organization from disease-related vascular and movement effects.\n\nBeyond presurgical mapping, fMRI is widely used to investigate stroke, dementia, depression, schizophrenia, and other neurological or psychiatric conditions. Group studies may identify altered task responses or network relationships, but overlap between affected and unaffected individuals can be substantial. A reproducible group difference does not automatically become a clinically useful biomarker. Individual diagnostic use requires validation of accuracy, calibration, robustness, and added clinical value in the populations where the test would be applied.\n\nResearch also uses fMRI to study treatment mechanisms, recovery, and covert responses in some patients with disorders of consciousness. In carefully designed paradigms, following an imagery instruction can produce task-specific brain activity despite an absence of observable movement. Such results require rigorous exclusion of alternative explanations, while a negative result cannot establish absence of awareness. Real-time fMRI neurofeedback, in which participants receive information derived from their own signal, has been investigated experimentally. Across these translational areas, feasibility, biological interpretation, and demonstrated patient benefit remain separate questions."
    },
    {
      "heading": "Safety, ethics, reproducibility, and emerging directions",
      "text": "Standard BOLD fMRI does not use ionizing radiation and ordinarily requires no injected contrast agent. Its physical risks are those of the magnetic resonance environment: a strong static magnetic field, changing gradient fields, radiofrequency energy, and substantial acoustic noise. Ferromagnetic objects can become dangerous projectiles, some implants can malfunction or heat, and conductive materials can create burn hazards. Implant compatibility is specific to the device and scanning conditions rather than determined by a simple distinction between metal and nonmetal.\n\nParticipant experience also affects both ethics and data quality. Confined space, noise, prolonged immobility, and demanding tasks can cause discomfort or anxiety. Studies involving children, people with cognitive impairment, or patients with serious illness require procedures appropriate to their circumstances and decision-making capacity. Institutional review board oversight and informed consent address research risks, recruitment, and handling of unexpected findings. Structural scans obtained alongside fMRI may reveal incidental abnormalities, making plans for review and communication an important part of study governance.\n\nBrain imaging data raise distinctive data privacy concerns. Structural images can contain facial information, and imaging datasets may be linked to genetic, behavioral, or clinical records. Removing names is therefore not always sufficient to prevent reidentification. At the same time, claims that routine fMRI can read unrestricted private thoughts exceed the evidence. Decoding studies operate under defined training and testing conditions, often requiring extensive individual data and constrained stimulus sets. Their performance cannot be assumed to transfer to uncooperative participants or unrestricted real-world settings.\n\nReproducibility depends on sample size, measurement reliability, analytical flexibility, and independent validation. An effect that is robust across participants need not distinguish individuals reliably, because reliability depends on the ratio of stable individual variation to measurement noise. Small studies of brain–behavior associations can yield unstable estimates and exaggerated correlations. Large multisite datasets improve statistical power and representation, but introduce scanner, protocol, and population differences that require explicit modeling. Preregistration, shared processing code, independent replication, and clear separation of exploratory from confirmatory analysis help make the evidential basis of results assessable.\n\nCurrent technical work addresses different limitations rather than a single universal bottleneck. Faster acquisition improves temporal sampling; higher fields and improved coils support finer spatial measurements; multi-echo methods assist artifact separation; and calibrated approaches combine BOLD with perfusion and physiological challenges to constrain vascular interpretation. Cortical-depth imaging seeks information about responses across the thickness of the cerebral cortex, while individualized mapping characterizes organization that can be obscured by group averaging. Each approach introduces its own demands on acquisition, modeling, and validation, particularly when the intended inference concerns neural mechanisms rather than the measured hemodynamic signal."
    }
  ],
  "links": [
    "magnetic-resonance-imaging",
    "brain",
    "contrast-agent",
    "astrocyte",
    "carbon-dioxide",
    "multiple-testing",
    "language",
    "pain",
    "machine-learning",
    "electroencephalography",
    "magnetoencephalography",
    "epilepsy",
    "stroke",
    "dementia",
    "depression",
    "schizophrenia",
    "biomarker",
    "institutional-review-board",
    "data-privacy",
    "cerebral-cortex"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}