Loading field map
For most of neuroscience's history, the brain could only be studied after death or through invasive animal experiments. The desire to observe the living human brain—its structure, its activity, and the connections that bind it together—drove the development of neuroimaging. But the images themselves do not come with ready-made interpretations. Over the past century, researchers have adopted four distinct frameworks for turning raw scans into scientific explanations: structural, functional, connectivity, and quantitative-computational. Each framework answered a different question, relied on different methods, and reshaped what it meant to 'see' the brain.
The first framework was built on the conviction that brain anatomy holds the key to brain function. Structural neuroimaging emerged from clinical radiology: X-ray-based techniques such as pneumoencephalography and cerebral angiography gave way to computed tomography (CT) in the 1970s and then to magnetic resonance imaging (MRI) in the 1980s. These methods produced high-resolution images of brain tissue, allowing researchers to identify lesions, measure atrophy, and map the boundaries of cortical regions. The guiding assumption was localizationist: different mental faculties were thought to reside in distinct anatomical areas, so knowing the structure meant knowing the function.
Structural neuroimaging was enormously successful for clinical diagnosis—detecting tumors, strokes, and degenerative diseases—but it had a fundamental limitation. It captured only static anatomy, not the dynamic activity that underlies thought, perception, and action. A patient could have a structurally normal brain yet suffer profound cognitive deficits, and a structural scan could not explain why. This gap between anatomy and function created the pressure that the next framework would address.
Functional neuroimaging broke with the structural framework by aiming to measure brain activity directly. Positron emission tomography (PET), introduced in the 1970s, tracked radioactive tracers to map glucose metabolism or blood flow. Functional magnetic resonance imaging (fMRI), developed in the early 1990s, used the blood-oxygen-level-dependent (BOLD) signal to infer neural activity with much better spatial and temporal resolution. The core method became the activation map: participants performed a task while the scanner recorded signal changes, and statistical contrasts highlighted regions where activity increased relative to a baseline.
This framework extended the localizationist tradition of structural neuroimaging but transformed it into a dynamic, task-driven enterprise. Instead of asking 'Where is the lesion?', researchers asked 'Which brain regions are recruited when someone reads, remembers, or feels?' The result was a flood of maps linking cognitive functions to specific areas—the fusiform face area for faces, the parahippocampal place area for scenes, and so on. Yet functional neuroimaging carried its own limitations. It treated brain regions as independent processors, largely ignoring the interactions between them. A region might 'light up' during a task, but that told little about how it communicated with other regions or how the network as a whole produced behavior.
Connectivity and network neuroimaging emerged directly from the dissatisfaction with isolated activation maps. Diffusion tensor imaging (DTI) allowed researchers to trace white-matter tracts non-invasively, revealing the structural highways that connect distant regions. Resting-state fMRI, meanwhile, measured spontaneous low-frequency fluctuations in the BOLD signal while participants lay still, uncovering intrinsic functional networks—sets of regions whose activity fluctuated together even in the absence of a task. Graph-theoretic analysis then provided a mathematical language to describe these networks: nodes (brain regions) and edges (structural or functional connections), with measures such as modularity, hub centrality, and small-world organization.
Where functional neuroimaging treated regions as independent, connectivity neuroimaging emphasized that brain function arises from interactions across distributed networks. This framework did not replace functional neuroimaging; rather, it absorbed and reframed it. Many studies now combine task-based activation with connectivity analysis, asking not just which regions are active but how they coordinate. The network perspective also resonated with the broader neuroscience framework of Network Neuroscience, which views the brain as a complex system whose properties emerge from its wiring diagram. Connectivity neuroimaging narrowed the focus of earlier localizationist mapping by showing that even simple tasks engage widespread circuits, not isolated spots.
The most recent framework pushes beyond description toward mechanistic explanation. Quantitative and computational neuroimaging uses advanced MRI techniques—such as quantitative susceptibility mapping, relaxometry, and diffusion kurtosis imaging—to measure physical tissue properties like iron content, myelin density, or axon diameter. More importantly, it applies computational models to imaging data. In model-based fMRI, for example, researchers fit a cognitive or neural model (e.g., a reinforcement-learning algorithm) to a participant's behavior and then ask whether the model's internal variables—prediction errors, value signals—correlate with BOLD activity in specific regions. This approach treats the brain not as a map of regions but as a system that implements computations.
Unlike the earlier frameworks, which were primarily descriptive (this region activates, this network connects), quantitative and computational neuroimaging aims to infer the underlying mechanisms that generate the observed signals. It absorbs data from structural, functional, and connectivity frameworks but reframes them in terms of parameters that can be compared across individuals and linked to behavior or disease. For instance, a computational model might explain why a patient with schizophrenia shows altered prefrontal activation during a working-memory task: not because the region is 'broken,' but because the model's parameters for gain control or noise are abnormal. This framework coexists with the others; many labs use all four approaches in a single study, combining structural scans, functional activation maps, connectivity matrices, and computational models.
Today, no single framework dominates neuroimaging. Functional neuroimaging remains the most widely used, especially in cognitive neuroscience, because activation maps are intuitive and easy to communicate. Connectivity and network neuroimaging has become standard for studying large-scale brain organization and is central to the Human Connectome Project. Quantitative and computational neuroimaging is growing rapidly, driven by the desire for mechanistic understanding and by advances in machine learning that can fit complex models to high-dimensional data. Structural neuroimaging, once the research driver, has narrowed to a clinical baseline—every functional or connectivity study still acquires a structural scan for coregistration and segmentation, but anatomy alone is rarely the primary question.
The leading frameworks agree on several points: that brain function must be understood at multiple scales, that no single method captures the whole picture, and that combining structural, functional, and connectivity data is essential. They disagree, however, on what constitutes an explanation. Functional neuroimaging tends to treat a statistical map as an explanation: 'Region X is involved in memory.' Connectivity neuroimaging counters that involvement is meaningless without knowing the network context. Quantitative and computational neuroimaging goes further, arguing that even network descriptions are insufficient unless they specify the computations that the network performs. This tension—between mapping, networking, and modeling—is the driving debate of contemporary neuroimaging, and it ensures that the field will continue to evolve as new methods and frameworks emerge.