A brain image, a cellular recording and a memory task measure different things. Neuroscience connects these levels, but a convincing explanation needs evidence for each connection. This guide introduces the terms most useful when deciding what a study measured and what remains an interpretation.
Neurons and their connections
Neuron research examines the cells that receive, process and transmit information in the nervous system. Axons and dendrites are parts of many neurons involved in these connections. Synapse research examines junctions where a neuron communicates with another cell. Neurotransmission and synaptic transmission describe the signalling involved in that communication.
Record whether a study measures electrical activity, chemical release, a molecular component or the structure of a connection. These observations can support different conclusions. A photograph showing a component does not automatically establish how information passed through the connection during a task.
Electrical activity and plasticity
An action potential is a rapid electrical event associated with signalling in excitable cells. Neuronal excitability concerns a neuron's responsiveness to inputs. Neuronal signalling includes several electrical and chemical processes, while synaptic plasticity concerns changes in the properties of connections.
The experimental preparation and stimulus are part of the result. An effect in a recorded cell does not directly establish an improvement in memory or behaviour. Keep the measured response, its time course and any separate behavioural test distinct when reading a proposed explanation.
Transmitters and receptors
Neurotransmitter signalling uses chemical messengers and responding systems. Neurotransmitter receptors detect particular signals and participate in downstream responses. Neurotransmitter metabolism concerns the processes producing, transforming or clearing these substances.
A concentration, receptor measurement and functional response are not the same endpoint. Avoid reducing a complex brain process to a universal claim that more or less of one messenger is better. The location, receptor type, timing and model can all affect the interpretation of an observation.
Imaging and recorded brain signals
Structural brain imaging examines anatomical features. Functional brain imaging examines signals related to activity or function using a stated method. MRI brain research can investigate different kinds of contrast and measurement; EEG research records electrical signals at the scalp. Neuroimaging research therefore covers several methods, not one direct readout of thought.
Read the signal definition, processing and comparison. A coloured statistical map is not a photograph of a feeling or diagnosis. An association between a task and a measured signal needs the study's analytical context, including how regions or time windows were selected.
Cognition and behaviour need defined tasks
Cognitive neuroscience examines relationships between brain systems and processes such as attention or memory. Behavioural neuroscience studies links between biological systems and behaviour. Memory research, attention research and learning neuroscience each use particular tasks, models and outcome definitions.
Identify the task rather than treating its score as a complete measure of a person's ability. Performance may involve several processes at once. A result on one task should retain that qualification unless the study also examines broader transfer. Note whether outcomes were assessed immediately or after a longer interval.
A model is a tool for a bounded question
Ask which part of the biological or behavioural process the model represents and which parts it leaves out. Compare observations at the same level before linking a molecular finding to an imaging result or a clinical claim. A proposed pathway should not become proof merely because it is drawn between two panels.
The cellular-research guide explains molecular measurements, and the sleep-research guide explains timing and sleep outcomes. This page does not diagnose a neurological condition or recommend a substance to alter cognition.
Sources and further reading
Sources checked 25 September 2026: NINDS: brain basics; NINDS: genes at work in the brain; NINDS: know your brain.
Separate an imaging signal from the claim made about it
Structural MRI and functional MRI answer different measurement questions. Structural images describe anatomical features, while commonly used fMRI methods use changes related to blood oxygenation to investigate activity. NIBIB's MRI overview explains the underlying distinction. A coloured statistical map is not a photograph of a thought.
For an invented experiment, compare a task condition with a specified control condition. A region differing between those conditions may reflect any feature that differs with them, including attention or movement. The experimental contrast therefore belongs beside the image. A region's association with a task does not establish that it is uniquely responsible for that task.
Read the acquisition method, comparison, motion handling, exclusions and statistical threshold before interpreting the map. Check whether the reported region was defined independently or selected after inspecting the same results. Retain the uncertainty and population studied. These questions support reading a brain-imaging paper; they do not diagnose a condition from a scan or establish an effect of a catalogue product.
An EEG cleaning decision is part of the result
EEG recordings can contain signals related to blinks, movement and other sources besides the activity being investigated. Preprocessing identifies or reduces some of these influences, but its decisions also affect the data retained. The MNE overview of artefact detection illustrates why inspection and documented processing are important.
Imagine a fictional comparison with 100 recorded trials in each group. If 10 are rejected in one group and 45 in the other, the final averages describe different retained sets. The reader needs the rejection rule and its consequences, not just the smooth final trace. Removing data because it disagrees with the preferred interpretation would create a further bias.
Record filter settings, reference choice, marked channels, rejected segments, any correction method and the resulting observation counts. Distinguish choices made before the analysis from later sensitivity checks. A cleaner-looking trace is not automatically a more valid result, and a preprocessing package cannot decide the scientific question for the researcher. Preserve the original data and processing provenance when evaluating reproducibility.
Keep PET tracer identity with the measurement
PET uses a radiotracer whose distribution is detected through emissions associated with radioactive decay. What the image can support depends on that tracer and the measurement approach. The NIBIB nuclear-medicine overview explains the method. A PET signal should not be rewritten as a direct measurement of every neurotransmitter or every aspect of brain activity.
In a hypothetical paper comparison, two studies use different tracers but both display a bright region. Similar appearance alone cannot establish that they measured the same biological target. The tracer, acquisition timing, reference and analysis determine the meaning of the displayed quantity.
For a reading record, retain tracer identity, intended target, participant group, acquisition window and the model used to derive the reported measure. Distinguish the observed signal from a proposed mechanism. Where a model includes assumptions about binding or a reference region, keep those assumptions visible. This allows a useful comparison between studies without turning a research image into a diagnosis or a recommendation for a scan.
Is a tractography streamline a directly observed nerve connection?
No. A streamline is a reconstructed path derived from diffusion measurements and modelling choices. It is not a direct tracing of an individual axon. In an original tractography challenge, algorithms recovered many intended bundles but also produced plausible-looking connections absent from the reference. An attractive reconstruction alone cannot establish anatomical truth.
In a reading example, two processing pipelines generate different numbers of streamlines from the same diffusion scan. The count difference is a result of the reconstruction procedures, not evidence that the participant acquired or lost axons between analyses. Ask which tracking rules, seed regions and stopping criteria produced each image before interpreting a difference as connectivity.
Keep reconstruction performance separate from a clinical or behavioural association. A benchmark with known reference bundles addresses one validation question; an association in human scans addresses another. Record which false-positive and missed-connection errors were evaluated and how the authors defined a successful match. The challenge is a particular study, not a measurement of every contemporary pipeline. Use its limitation to scrutinise the method reported in the paper you are reading, rather than treating all reconstructed paths as either proven or worthless.
Does activation of a brain region identify a particular thought?
Not by itself. Observing activity in a region during a task and inferring a particular mental process from activity in that region are different reasoning steps. Poldrack's analysis of reverse inference asks how much additional evidence an activation supplies, including how selectively that region is associated with the proposed process.
Imagine a fictional region that responds during both a memory task and several demanding tasks without an explicit memory requirement. Seeing its activity in a new experiment would not uniquely identify memory processing. The reader needs to consider alternative processes and the task comparison. The example does not assign a real brain region a single function; it illustrates why reversing an association requires more information.
Write down the task, comparator and measured signal before the psychological interpretation. Look for behavioural evidence supporting that interpretation and ask whether the report tested a specific prediction or offered a post-hoc explanation. A significant contrast does not by itself reveal a participant's intention, diagnose a condition or show that every person used the same strategy. Preserve that separation when summarising a cognitive-neuroscience result.
Can preprocessing create a negative brain-signal correlation?
A negative association between two recorded signals is not automatically evidence that one brain region inhibits another. The preprocessing pipeline can change the distribution of correlations before any biological explanation is considered. Global signal regression is a concrete example that needs to be reported explicitly.
An original fMRI methods study examined this issue using simulations and experimental recordings. It showed that removing a global signal can introduce or reshape negative correlations under the studied conditions. That finding does not mean that every negative correlation is artificial; it means the processing decision belongs in the interpretation.
For a hypothetical comparison, one analysis removes a whole-brain average signal and another does not. Applying the same correlation threshold to both maps does not make their retained connections equivalent. Their distributions and underlying nuisance-removal assumptions may differ even when they began with the same observations.
Read how the global signal was defined, which other nuisance variables were modelled, and whether the conclusion remained under justified alternative processing choices. Also preserve the distinction between statistical association and directional influence. A sensitivity analysis can reveal dependence on a choice, but it cannot by itself identify a unique biological mechanism. The useful conclusion states the observed relationship together with the analytical conditions supporting it.
Does changing the EEG reference change the apparent signal?
An EEG channel records a voltage difference relative to a reference. It is therefore incomplete to describe a scalp signal by electrode location alone. Changing the reference can change the appearance of the recorded differences without a new biological recording taking place.
An original comparison of EEG references reprocessed the same recordings using different reference approaches. The design is useful for interpretation because the participants and underlying recording were held constant while the reference definition changed. A resulting map difference cannot simply be assigned to a change in brain activity.
An average reference also depends on which channels enter that average. Missing, noisy or interpolated channels can affect what the reference represents. A reference intended to approximate a distant neutral point involves additional modelling assumptions; its name does not establish that every recording has become independent of reference choices.
For a comparison, record the acquisition reference, any later re-referencing, the channel set and the order of relevant processing steps. Check whether both groups and visits were treated consistently. Reference choice is distinct from removing eye movements or other artefacts, although both can influence later analysis. Preserving those details supports a clear comparison of spectral or topographic findings without converting an EEG map into an individual diagnosis.
Why does recognition accuracy need both hits and false alarms?
In a recognition task, correctly identifying a previously presented item is only one outcome. A participant can also identify a new item incorrectly as familiar. Looking only at correct recognitions can confuse a greater tendency to answer yes with better discrimination between old and new material.
The original methods paper on signal-detection measures distinguishes discriminability from the decision criterion used to make responses. This framework retains both the hit rate and false-alarm rate rather than treating the proportion of yes responses as a complete memory measure.
Imagine two hypothetical participants with equal hit rates. If one also accepts many new items as familiar, their performances differ in a way that the hit rate alone conceals. Instructions, expected item frequencies or the consequences of an error can influence response strategy. Those task conditions should therefore remain visible alongside the score.
Check how old and new items were sampled, how nonresponses were treated and which correction was used if a rate was zero or one. A derived sensitivity measure also has model assumptions; it is not a direct quantity read from the brain. Report the task, component rates and chosen statistic before generalising to memory in daily life. This distinction complements, rather than replaces, checks for practice effects across repeated testing.
Has a brain-data model been tested on genuinely unseen people?
A model can recognise features of its development data without generalising to new participants. In neuroscience, many measurements from one person can make a dataset look large while leaving the number of independent people small. The intended prediction target determines which observations must be kept separate.
An original investigation of machine-learning validation used simulations to examine performance estimates with limited samples. Its methodological message is relevant to brain-data studies: selecting features or tuning a model using information from the evaluation set can give an optimistic estimate of future performance.
Suppose repeated scans from the same participant appear in both training and testing data. The model may exploit stable person-specific features rather than the general relationship the study claims to predict. A test intended to represent new patients or research participants therefore needs separation at the relevant participant level, not simply a random split of individual scans.
Look for the timing of feature selection, model tuning and final evaluation. Report whether preprocessing decisions were learned within training data and whether a genuinely independent dataset was used. A high accuracy figure alone does not establish clinical usefulness, portability between scanners or performance in an unrepresented population. The validation design must match the proposed use.
Is every spike-sorting cluster a separate neuron?
Extracellular electrodes record overlapping electrical activity. Spike sorting groups detected events into estimated units, but a cluster label is an analytical assignment rather than a direct count of anatomically identified neurons. Errors can split one unit into several clusters or combine events from different units.
The original Kilosort4 paper describes challenges from changing waveforms and movement of the recording probe relative to tissue. Its validation uses simulated and hybrid data with known events, allowing missed and incorrectly assigned spikes to be assessed separately. A larger output unit count is therefore not, by itself, proof of a better reconstruction.
Imagine a recording in which one neuron's waveform shifts during the session. If the shift is mistaken for a new identity, the output may appear to contain two units active at different times. That could distort an interpretation of changing population activity even though the underlying biological source was continuous.
Retain the sorter and version, drift handling, quality criteria and manual curation record. Ask which ground-truth conditions resemble the actual recording and which do not. Unit isolation, event detection and downstream firing-rate analysis are related but separate checks. This is a limitation of interpreting neural recordings, not a claim that one algorithm recovers every neuron correctly.
What distinguishes an axon from a dendrite in a research figure?
An axon is a neuronal projection that carries signals away from the cell body towards terminals communicating with other cells. Dendrites are usually branching projections that receive inputs. The NINDS glossary distinguishes these roles. Calling both structures parts of a neuron leaves out the difference that makes their identification meaningful.
In a research figure, the cell body, axon, dendritic branches and communication junctions describe different locations. A measured length or branch count is a structural observation. An electrical recording describes activity under the recording conditions. One does not automatically supply the other, even when both observations concern the same cell.
Imagine a hypothetical image analysis reporting more dendritic branches in one preparation. That result concerns the structures identified and counted. Without a separate functional measurement, it does not establish that the cell received more effective input, that an axon conducted faster, or that a person would learn better. The endpoint should remain attached to the measured structure.
When comparing papers, retain the neuronal type, preparation, identification method and the part of the cell analysed. Check whether an image shows a complete projection or only the portion within the field of view. These details distinguish a bounded axon or dendrite result from a general claim about neuronal communication; they are not a diagnosis or evidence for a cognitive-enhancement product.
Top comments (0)