VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Why Mathematics? | PET Scans, Coincidence Detection and Image Reconstruction

Three students reviewing open textbooks, notes and stationery together at a classroom table.

Why is mathematics important in a PET scan? Because the scanner does not take an ordinary photograph of the body. It detects pairs of gamma photons arriving around the same time, assigns each pair to a possible line through the patient, corrects imperfect measurements and reconstructs a three-dimensional estimate of where a radiotracer was distributed.

PET stands for positron emission tomography. It belongs to nuclear medicine and is used by qualified clinical teams for specific diagnostic, treatment-planning and research purposes. The image is built from physical events and statistical evidence, not from a camera lens. Geometry says where an event may have occurred. Probability describes uncertain counts. Calculus and linear algebra connect many lines of response to an image. Optimisation chooses an image that is compatible with the measurements and model.

This article explains those mechanisms for students and families. It does not interpret a scan, compare personal treatment options or replace advice from a medical professional. A bright region is not a diagnosis by itself; tracer biology, acquisition, correction, reconstruction and clinical context all matter.

> Did You Know? A single detected photon pair usually does not reveal one exact source point. It suggests a line of response. The image becomes useful only after many uncertain events are combined.


Quick Navigation


What PET Measures

A radiotracer contains a radionuclide that emits a positron. Its chemical form is chosen for a particular biological question. The tracer moves through the body according to biology; the scanner measures radiation produced after positron emission. The final image estimates tracer distribution over space and sometimes time.

The US National Institute of Biomedical Imaging and Bioengineering provides an official overview of nuclear medicine and positron emission tomography. That overview helps separate the established modality from simplified classroom language: PET is a nuclear-medicine technique, and its images reflect physiological or biochemical processes associated with the chosen tracer.

An image voxel is an estimate

A digital PET volume is divided into voxels, the three-dimensional counterparts of pixels. Each reconstructed voxel value is inferred from many detector events, system responses and corrections. It is not a tiny container that directly counted every decay occurring inside it.

This distinction is one reason mathematics matters. The scanner records detector positions, energies and times. Reconstruction infers a spatial distribution that could plausibly have produced those records.

Measurement and interpretation are separate steps

The same reconstructed intensity can have different meanings for different tracers, organs, times after administration and clinical questions. Image interpretation therefore requires protocol knowledge and medical context. Mathematics can state the measurement chain and its uncertainty; it cannot turn every bright region into one universal conclusion.


From Positron to Coincidence Pair

A positron is the antimatter counterpart of an electron. After travelling a short distance and losing energy in tissue, it can annihilate with an electron. The annihilation commonly produces two gamma photons travelling in nearly opposite directions, each with energy associated with the electron rest mass.

Detectors arranged around the patient convert arriving radiation into electrical signals. The electronics record properties such as detector location, deposited energy and time. When two qualifying detections occur within a short timing window, the system may treat them as a coincidence event.

Why “nearly opposite” matters

In an ideal sketch, the two photons leave at exactly 180 degrees. Real physical conditions introduce a small non-collinearity. The positron also travels some distance before annihilation. Detector elements have finite size, and their response is not perfect. These effects contribute to spatial resolution limits.

A school diagram is still useful, but it must be labelled as an ideal model. The strongest science writing tells the reader both what the model reveals and where it stops.

The coincidence window

Suppose one detector records an event at time \(t_1\) and another at \(t_2\). A simple coincidence rule might require

\[ |t_1-t_2|<\Delta t, \]

along with energy and system criteria. Here \(\Delta t\) is a timing window chosen by the instrument design and processing method.

A window that is too narrow could reject true pairs because measurements have finite timing uncertainty. A window that is too wide admits more unrelated detections that happen to occur close in time. The choice is therefore a signal-versus-background trade-off rather than a magical boundary between certainty and error.


Lines of Response and Geometry

If two detectors register a qualifying pair, draw the line joining their positions. Under the idealised back-to-back model, the annihilation occurred somewhere along this line of response.

Let detector positions be \(A=(x_1,y_1,z_1)\) and \(B=(x_2,y_2,z_2)\). A parametric line is

\[ r(u)=A+u(B-A),\qquad 0\le u\le1. \]

The parameter \(u\) moves from one detector to the other. Without additional timing information, the pair does not identify one value of \(u\).

Many lines reveal structure

One line is ambiguous. Thousands or millions of lines intersecting different regions create statistical evidence about the tracer distribution. If many well-corrected lines pass through one area more often than the model expects from background and neighbouring activity, the reconstructed value there can rise.

This is tomography: information from many directions is combined to estimate an interior distribution. The principle connects with MRI k-space and Fourier image reconstruction, but the measurements and physics are different. MRI does not detect annihilation photon pairs, and PET does not simply use the same data pipeline under another name.

Projection coordinates and sinograms

In a two-dimensional idealisation, a line can be represented by its perpendicular distance \(s\) from the origin and angle \(\theta\):

\[ x\cos\theta+y\sin\theta=s. \]

Counts grouped by \((s,\theta)\) form projection data often visualised as a sinogram. A point source traces a sinusoidal pattern as the viewing angle changes, which gives the representation its name.

The equation is a useful meeting place for coordinate geometry and imaging. A line drawn in ordinary \((x,y)\) space becomes one point in projection space. Reconstruction reverses that relationship across many measurements.


Counts, Probability and Noise

Radioactive decay and photon detection are stochastic. Repeating a scan under nominally identical conditions would not produce exactly the same event list. Count data are often modelled with Poisson statistics when the assumptions are suitable.

For a Poisson random variable \(K\) with mean \(\lambda\),

\[ P(K=k)=\frac{e^{-\lambda}\lambda^k}{k!}, \]

and both the mean and variance equal \(\lambda\). The standard deviation is \(\sqrt{\lambda}\), so the relative standard deviation is approximately

\[ \frac{\sqrt{\lambda}}{\lambda}=\frac{1}{\sqrt{\lambda}}. \]

At 100 expected counts, this relative scale is about 10%. At 10,000 expected counts, it is about 1%. More counts improve statistical precision, but the square-root relationship means improvement is not proportional to acquisition time or activity.

Counts are not all true coincidences

Detected pairs can be grouped conceptually:

  • true coincidences: both detections originate from the same annihilation without a disqualifying interaction;
  • scattered coincidences: one or both photons change direction before detection;
  • random coincidences: unrelated photons happen to fall within the coincidence window;
  • lost or misclassified events: finite detector and processing behaviour prevents ideal recording.

The measured total is therefore not simply “number of tracer decays.” Corrections and statistical models estimate how different event classes contribute.

Signal-to-noise is not one universal number

Noise depends on count level, reconstruction, corrections, filtering, voxel size and region. A global signal-to-noise claim can hide local performance. Two images may contain the same events but look different because reconstruction settings trade spatial detail against visible noise.


Corrections Before Reconstruction

Photons can be absorbed or scattered in the body, detectors do not all respond identically, and geometry makes some lines easier to detect than others. Reconstruction that ignores these effects can place the wrong intensity in the wrong location.

Attenuation

For a narrow monoenergetic beam through a material, the Beer–Lambert model is

\[ I=I_0e^{-\int_L\mu(r)\,dr}, \]

where \(I_0\) is unattenuated intensity and \(\mu(r)\) is the position-dependent linear attenuation coefficient along path \(L\).

The line integral matters because a photon pair crosses different tissues over different distances. A line passing through more attenuating material is less likely to be detected. Attenuation correction estimates this path-dependent loss.

Normalisation and detector response

Detector pairs can differ in sensitivity. Calibration data help estimate normalisation factors so that one region does not appear unusually active merely because its detector geometry or efficiency is favourable.

Scatter and random correction

Scatter estimation seeks to account for events whose measured detector line no longer represents the original annihilation direction. Random-event estimation uses timing or other acquisition information to estimate unrelated coincidences. These are model-based corrections with uncertainty; they do not label every individual event with perfect knowledge.

Motion and registration

Breathing, heartbeat and patient movement can blur or misalign information. PET may be combined with computed tomography or magnetic resonance imaging for anatomical localisation and correction workflows. Registration uses transformations to align coordinate systems, but imperfect motion and timing remain possible.

This is why ultrasound echo time and image depth is related yet distinct. Both use timing and reconstruction, but ultrasound sends and receives acoustic pulses, whereas PET detects radiation originating from tracer events.


From Projections to an Image

Discretise the unknown tracer distribution into voxel values \(x_j\). Group measurements into detector bins \(y_i\). A system matrix \(A\) describes how strongly activity in voxel \(j\) contributes to measurement bin \(i\). A simple expectation model is

\[ \widehat{y}_i=\sum_j A_{ij}x_j+b_i, \]

where \(b_i\) represents estimated background such as scatter and random contributions.

The matrix can be enormous and sparse. Its entries may include geometry, attenuation, detector response, resolution and other effects. Reconstruction asks for non-negative voxel values whose predicted measurements agree reasonably with the observed counts under the statistical model.

Why direct inversion is not enough

Writing \(y=Ax\) tempts students to calculate \(x=A^{-1}y\). In practice, the system may not have a simple square invertible matrix. Data are noisy, measurements incomplete, and tiny perturbations can cause unstable changes. Physical constraints and statistical likelihood matter.

Iterative methods start with an image estimate, predict measurements, compare them with observations and update the image. Maximum-likelihood expectation-maximisation and ordered-subsets variants are important examples in emission tomography. The algorithm name should not be mistaken for guaranteed truth: it optimises a stated objective under a stated model.

A simplified multiplicative update

One conceptual maximum-likelihood emission update has the form

\[ x_j^{(k+1)}=x_j^{(k)} \frac{\sum_i A_{ij}\,y_i/\widehat{y}_i^{(k)}}{\sum_i A_{ij}}, \]

with appropriate treatment of backgrounds and implementation details. The ratio \(y_i/\widehat{y}_i\) says whether the current image underpredicts or overpredicts a measurement. The matrix redistributes that information back toward contributing voxels.

Iterations can increase data agreement while also amplifying noise or model mismatch. Stopping rule, regularisation, subsets, filtering and calibration therefore affect the displayed image.

Filtered backprojection as a geometric lesson

In an ideal continuous setting, projection data can also be reconstructed through filtered backprojection. Plain backprojection smears each measured line back across the image, producing blur. Frequency-domain filtering compensates for that blur before or during backprojection.

The method is pedagogically valuable because it makes the inverse problem visible. Each projection spreads evidence along its possible origin line; many angles intersect to form structure. Noise and missing angles create artefacts because the inverse operation cannot invent reliable measurements that were never made.


Time-of-Flight PET

If one annihilation photon arrives slightly earlier than the other, the timing difference gives information about where along the line the event occurred.

Let \(\Delta t=t_1-t_2\). If the annihilation is displaced by distance \(d\) from the line midpoint, one photon travels \(d\) less while the other travels \(d\) more, a path difference of \(2d\). With light speed \(c\),

\[ d=\frac{c\Delta t}{2}. \]

This does not normally locate the event at an exact mathematical point because timing has finite resolution. It narrows the probability distribution along the line.

Worked timing scale

For \(\Delta t=200\) picoseconds,

\[ d=\frac{(3.0\times10^8\ \text{m/s})(200\times10^{-12}\ \text{s})}{2} =0.030\ \text{m}. \]

The displacement scale is about 3 cm. This illustrates why picosecond timing matters and why dividing by 2 is essential. The time difference corresponds to twice the displacement in path length.

The calculation is an ideal scale, not a claim that every event is located to exactly 3 cm. System timing response is a distribution, and reconstruction uses it statistically.


A Worked Coincidence Example

Imagine a circular detector ring with two opposing detectors at \(A=(-40,0)\) cm and \(B=(40,0)\) cm. A qualifying coincidence defines the x-axis segment between them:

\[ r(u)=(-40,0)+u(80,0),\quad 0\le u\le1. \]

Without time-of-flight information, any position along that segment remains possible under the simplified geometry. Suppose detector A receives its photon 100 ps earlier than B. The event is closer to A. The displacement magnitude from the midpoint is

\[ |d|=\frac{(3.0\times10^8)(100\times10^{-12})}{2}=1.5\ \text{cm}. \]

Under a sign convention where negative x points toward A, the most likely location along this one-dimensional example shifts toward \(x=-1.5\) cm. A different sign convention changes the algebraic sign but not the physical conclusion.

Audit the calculation

  • Time must be converted from picoseconds to seconds.
  • Light speed and distance units must match.
  • The factor of 2 comes from opposite path changes.
  • Earlier arrival identifies the nearer detector only when timing channels and convention are correctly calibrated.
  • Finite timing resolution means a probability region, not an exact pin.

This audit is the imaging version of checking work before submission: verify units, direction, model and limiting cases before trusting a polished answer.


Resolution, Sensitivity and Dose

Imaging design contains trade-offs. Smaller detector elements can support finer sampling, but sensitivity, cost, timing and electronic complexity matter. Larger voxels collect more counts per voxel but display less spatial detail. Longer acquisition can increase counts, but patient movement and practical constraints may grow.

Radiation protection and clinical benefit are professional responsibilities governed by justified protocols, optimisation and relevant standards. Mathematics can compare expected information and uncertainty, but students should never convert simplified count formulas into personal dose advice.

Resolution is not only pixel size

Making a display grid finer does not create physical resolution beyond the scanner and reconstruction. Positron range, photon non-collinearity, detector size, depth of interaction, motion and reconstruction all contribute. A 1 mm voxel can hold a smoothly interpolated estimate even when the effective spatial resolution is coarser.

Sensitivity is spatially variable

Detector coverage and acceptance vary with position and line angle. A system model accounts for this non-uniformity. Reporting one sensitivity number without geometry, energy window, acquisition mode and test definition can be misleading.

Quantitative values need calibration

PET can support quantitative measures when acquisition, calibration, corrections and analysis are appropriate. Yet partial-volume effects can make small objects appear less intense because activity is spread across neighbouring voxels. Region definition, timing and biological variability also influence measurements.


A Safe Student Simulation

Students can model tomography without radiation or patient data.

Step 1: make a hidden grid

Create a 20-by-20 grid with two circular regions assigned higher activity. Treat each cell value as an expected event contribution.

Step 2: generate line sums

For several angles, calculate approximate sums along parallel lines. These values form ideal projections. Add Poisson random noise by sampling each projection bin around its expected count.

Step 3: backproject

Spread each noisy projection value back across the cells along its line. Normalise and display the result as a heat map. The circles should emerge, but blur and streaks will remain.

Step 4: test sample count

Repeat with expected counts multiplied by 4. Relative Poisson noise should fall by about a factor of 2 because \(1/\sqrt{4\lambda}=1/(2\sqrt\lambda)\).

Step 5: remove angles

Reconstruct after deleting a block of projection angles. Observe directional artefacts. The exercise shows that an inverse problem depends on measurement coverage, not only total count.

Students should record the hidden truth, random seed, angle set, noise model and reconstruction settings. Better note-taking turns a pretty heat map into a reproducible experiment.


Common Misconceptions

“A PET scanner photographs organs directly”

No. It detects radiation events and reconstructs an estimate of tracer distribution. Anatomical detail may come from combined imaging and registration.

“Every coincidence pair gives the source point”

Without time-of-flight information, it primarily gives a line of response. Even with time-of-flight, finite resolution produces a probability distribution along that line.

“More iterations always make the image better”

More iterations can improve agreement with measured data while also amplifying noise or model mismatch. Image quality depends on purpose and validated reconstruction settings.

“A brighter voxel always means disease”

Tracer uptake can have normal and abnormal causes, and interpretation depends on tracer, anatomy, timing, artefacts and clinical context. A voxel is not a diagnosis.

“A finer voxel size guarantees finer resolution”

It changes sampling and display. Physical resolution still depends on the detector, positron physics, motion and reconstruction.

“All detected pairs are true events”

Scatter, random coincidences and other effects contribute. Corrections estimate these components rather than perfectly labelling every event.


Sampling, Calibration and Uncertainty

Reconstruction begins with a finite sample of events. If the same acquisition could be repeated under identical conditions, the individual detections would differ even though the underlying tracer distribution was unchanged. This sampling variation is one reason a small image fluctuation should not automatically be interpreted as biology.

Binning and list-mode data

A scanner may retain events in list mode, where each accepted event carries measured attributes such as detector pair, energy and time, or group them into bins. A bin represents a small range of measurement geometries. Binning reduces a huge event list to counts, but the bin size matters.

Broad bins combine events that are not geometrically identical. Very narrow bins create many cells with few counts. This is another resolution-versus-variance trade-off. A finer grid does not automatically produce a more reliable estimate when the sample is limited.

Detector normalisation

Detector elements do not all respond identically. Normalisation estimates relative efficiencies so that a difference in recorded counts is not automatically mistaken for a difference in tracer activity. In a simplified model, if detector pair (i) has efficiency (e_i), a raw count (r_i) may satisfy

\[ r_i\approx e_i s_i, \]

where (s_i) is the idealised signal. Calibration estimates (e_i). Dividing by a small or uncertain efficiency can amplify noise, so real systems use controlled procedures and quality checks rather than a casual spreadsheet correction.

Random and scattered coincidences

A true coincidence comes from the same annihilation and follows the expected geometry closely enough to be accepted. A random coincidence joins unrelated detections that fall within the timing window. A scattered coincidence includes at least one photon whose direction changed before detection.

These categories affect the measurement differently. Randoms add a background related to singles rates and timing. Scatter redistributes events into incorrect lines of response. Correction models estimate both contributions, but those estimates have uncertainty. Subtracting an estimate does not erase the statistical consequences of having measured background.

Attenuation is directional

Photons travelling through more material are less likely to reach detectors without interaction. A simple attenuation model along path (L) is

\[ P(\text{no interaction})=\exp\left(-\int_L \mu(r)\,dr\right), \]

where \(\mu(r)\) is a position-dependent attenuation coefficient. The line integral explains why equal source activity can produce different detected counts along different paths. Hybrid scanners may use anatomical imaging and calibration information in attenuation correction, subject to alignment and modelling limits.

Partial-volume effects

If an active object is comparable to or smaller than the system's effective spatial resolution, its reconstructed activity can spread across neighbouring voxels and its peak value can be underestimated. This partial-volume effect shows why the voxel size printed in a file is not the same as independently resolving objects of that size.

Discretising space finely creates more unknowns, but the measurements may not contain enough independent information to determine every unknown sharply. More pixels do not create more physical evidence.


Dynamic PET and Time-Activity Curves

Some PET studies divide an acquisition into time frames. Instead of one static volume, the result is a sequence of reconstructed distributions. A defined region can then have a time-activity curve (C(t)).

From frames to a curve

Short frames preserve rapid change but contain fewer counts. Long frames improve count precision but average change within the frame. If a frame runs from (t_a) to (t_b), its value is closer to

\[ \bar C=\frac{1}{t_b-t_a}\int_{t_a}^{t_b}C(t)\,dt \]

than to a perfect instantaneous sample. Frame design is therefore a sampling decision.

Compartment models are hypotheses

In research and selected clinical settings, kinetic models describe movement of tracer between conceptual compartments such as blood and tissue. Differential equations connect transfer rates to measured curves. Parameters are fitted to data, but the compartments are modelling choices rather than little physical boxes visible in the image.

Two models can fit noisy observations similarly while giving parameters different interpretations. Responsible analysis checks identifiability, residuals, assumptions and sensitivity instead of reporting a fitted number without context.

Area under a curve

The area under a time-activity curve is an integral. With discrete frames, trapezoids can approximate it. For neighbouring measurements ((t_i,C_i)) and ((t_{i+1},C_{i+1})), the contribution is

\[ \frac{C_i+C_{i+1}}{2}(t_{i+1}-t_i). \]

Units must be carried carefully. If concentration is measured per millilitre and time in minutes, the area has concentration-minutes, not concentration alone.


Quality Control and Residuals

Quality control uses repeatable tests to detect changes in calibration, uniformity, sensitivity, timing and spatial response before an obvious failure.

Uniform phantom reasoning

Imagine scanning an object designed to have uniform activity. The expected reconstruction should be approximately uniform after accounting for noise and known physical effects. A systematic ring, edge bias or regional deficit can point toward detection, calibration, correction or reconstruction problems.

One summary is the coefficient of variation,

\[ CV=\frac{s}{\bar x}, \]

where (s) is a sample standard deviation and \(\bar x\) is a mean over a defined region. It is meaningful only with the protocol, region, sample size and acceptance criteria that produced it.

Residuals in measurement space

After reconstructing (x), the forward model predicts \(\hat y=Ax+b\). Residuals are \(y-\hat y\). A residual pattern can reveal structure the model does not explain. Random-looking residuals are not proof that the model is correct, but a strong systematic pattern is a reason to investigate.

Repeatability and reproducibility

Repeatability asks how measurements vary under closely matched conditions. Reproducibility asks about changes across scanners, sites, operators, protocols or times. A quantity can be precise within one setting yet disagree across settings because calibration or protocol differs.

Students can transfer this distinction to every science experiment. Getting the same answer twice is useful evidence, but it is not the same as showing another team with another instrument will obtain a comparable result.


A Second Worked Example: Counts and Precision

Suppose a simplified region has an expected 10,000 true-equivalent counts. Under a Poisson model, the standard deviation is approximately

\[ \sqrt{10{,}000}=100, \]

so the relative standard deviation is about one per cent. If another region has 400 expected counts, the standard deviation is about 20 and the relative standard deviation is five per cent.

The lower-count region has fewer absolute fluctuations but much larger fractional uncertainty. Comparing raw standard deviations without considering the mean can therefore mislead.

If acquisition time doubles under an idealised constant rate, expected counts rise from 400 to 800. Relative noise changes from (1/\sqrt{400}=5\%\) to (1/\sqrt{800}\approx3.54\%\). Doubling counts does not halve the relative Poisson noise; half the relative noise requires roughly four times the counts.

Real protocols cannot be chosen from this rule alone. Motion, tracer kinetics, patient needs, equipment and radiation-protection principles all matter. The example shows how quantitative trade-offs replace the vague claim that longer is always better.


Communicating Results Responsibly

Numbers such as uptake values, dimensions or changes between scans can look precise. Their digits do not remove biological and measurement uncertainty. Comparisons may depend on uptake time, normalisation, reconstruction settings, calibration, motion and region definition.

A responsible explanation separates four layers:

  • detected events;
  • corrections and reconstructed image;
  • a derived measurement made from the image;
  • clinical interpretation made with other evidence.

This separation prevents a common reasoning error: treating a derived value as if the scanner observed it directly. Mathematics helps by making the chain of inference visible and testable, not by replacing professional judgement.

When comparing two scans, subtraction is not enough. Registration, acquisition timing, reconstruction settings and uncertainty must be compatible before a difference is treated as meaningful. A visible change can combine biology with motion, calibration and sampling variation. The honest question is not only “Did the number change?” but “Is the change larger than plausible measurement variation under comparable conditions?”


Guidance and Pathways

For students, PET is a meeting point between physics, mathematics, computing, chemistry and biology. Useful foundations include coordinates, vectors, exponentials, probability, matrices, calculus, optimisation and programming.

Ask five questions when reading any imaging result:

  • What physical signal was measured?
  • How was position inferred?
  • Which corrections changed the raw data?
  • What uncertainty or resolution limit remains?
  • Which conclusion requires expert biological or clinical context?

For parents and teachers, focus on the reasoning chain rather than medical drama. A student can explore line geometry, Poisson counts and reconstruction safely through simulation. Avoid asking learners to diagnose example scans.

Careers connected to this mathematics include medical physics, nuclear medicine, biomedical engineering, radiochemistry, imaging science, detector engineering, data science and clinical research. Entry routes and professional responsibilities differ, so the Mathematics Pathways guide should be paired with current official institution and professional requirements.


Useful Next Reading

qPCR, cycle thresholds and exponential amplification follows a different biomedical measurement chain from fluorescence to quantity. Clinical trials, randomisation and sample size moves from an instrument-level measurement to evidence across participants.

For a broader learning map, visit the eduKate Sengkang Mathematics Hub and the Mathematics Year 0 to Adulthood Engineer Series.


Frequently Asked Questions

What does PET stand for?

PET stands for positron emission tomography, a nuclear-medicine imaging method that estimates radiotracer distribution from detected radiation events.

Why are two photons detected?

Positron–electron annihilation commonly produces two photons travelling in nearly opposite directions. A qualifying pair helps define a line of response.

What mathematics is used in PET?

Coordinate geometry, line integrals, exponentials, Poisson statistics, matrices, optimisation, calculus and numerical computation all appear.

What is a sinogram?

It is a representation of projection data indexed by line position and angle. A point object traces a sinusoidal path in an ideal two-dimensional sinogram.

What does time-of-flight add?

Arrival-time difference narrows the likely annihilation location along a line of response. Finite timing resolution means it remains probabilistic.

Does PET show anatomy or function?

PET primarily reflects distribution of a chosen tracer and associated physiological or biochemical processes. Combined CT or MRI can provide anatomical context.

Why are PET images noisy?

Decay and detection are stochastic, counts are finite, and corrections and reconstruction propagate uncertainty. Relative count noise often improves roughly with the square root of count.

Can students reconstruct an image safely?

Yes. Simulated line sums and Poisson counts teach the mathematics without radiation, clinical data or diagnostic interpretation.


Final Perspective: Mathematics Turns Coincidences Into Evidence

A PET image begins as a stream of detector events: where, when and with how much deposited energy. Geometry converts pairs into possible lines. Physics describes attenuation and timing. Probability explains count variation. Linear algebra and optimisation reconstruct a spatial estimate. Calibration and professional interpretation keep that estimate connected to reality.

That is why mathematics matters in PET. It does not remove uncertainty or make clinical decisions automatic. It makes the measurement chain explicit, correctable, testable and useful—so that many ambiguous coincidences can become a coherent image while their limits remain visible.