PET/CT Scatter Correction Explained

Scatter correction is what makes a PET number trustworthy. On a modern 3D PET/CT scanner, roughly a third or more of every coincidence that survives the energy and timing windows is a scattered event — a photon that changed direction inside the patient and now points the reconstruction to a line where no decay actually occurred. Left uncorrected, those events smear a low, diffuse haze across the entire image, sap lesion contrast, and bias standardized uptake values (SUVs). Scatter correction estimates that haze and removes it, and single-scatter simulation is the algorithm that does the heavy lifting on nearly every clinical system. ^{3, 4}

This article explains where PET scatter comes from, why the energy window cannot handle it alone, how single-scatter simulation reconstructs the scatter distribution from Compton physics, how scatter fraction is measured, where the method fails, and what is coming next. DRPS supports this work as part of its PET/CT and nuclear medicine physics services and medical physics consulting.

Introduction

Positron emission tomography measures a coincidence: two 511 keV annihilation photons detected within a few nanoseconds of each other, on opposite sides of the patient. The reconstruction assumes the annihilation happened somewhere on the straight line connecting the two detectors — the line of response (LOR). That assumption is what makes tomography possible, and it is exactly what Compton scatter breaks.

When one of the two photons scatters in the patient before reaching the detector, it arrives from a different direction, so the LOR the scanner records does not pass through the annihilation point at all. The event is still a genuine coincidence — it is not random — but it is mispositioned. Multiply that by the fact that modern fully-3D PET (with no interplane septa) accepts photons over a large solid angle, and scattered coincidences become a dominant background, not a nuisance. ^{1, 3}

The clinical stakes are high because PET is a quantitative modality. Oncologists stage disease, assess treatment response, and enroll patients in trials based on SUVs, and multicenter harmonization efforts stand or fall on whether a SUV measured in one city matches one measured in another. ¹⁰ None of that works if a variable, uncorrected scatter background is riding underneath every measurement. For the companion corrections that must all work together, see our posts on PET/CT attenuation correction and randoms, dead time, and NECR.

Topic Explanation

The three kinds of coincidence

Every "prompt" coincidence a PET scanner records falls into one of three categories:

True coincidence — both photons travel undeflected from a single annihilation to the two detectors. This is the signal.
Scattered coincidence — at least one photon Compton-scatters in the patient before detection, so the recorded LOR is wrong. This is a correlated background.
Random (accidental) coincidence — two photons from different annihilations happen to arrive within the timing window. This is an uncorrelated background, handled separately (see randoms and NECR).

The measured prompt sinogram is the sum of all three:

P = T + S + R

Randoms ( $R$ ) are estimated from delayed-window counts and subtracted. Scatter ( $S$ ) cannot be measured directly — it looks like signal in every respect except position — so it must be modeled. Recovering the trues is then:

T = P - S - R

Everything in scatter correction is about producing a trustworthy estimate of $S$ . ¹

Why the energy window is not enough

A natural first thought is: scattered photons lose energy in the Compton interaction, so just reject anything below 511 keV. This helps, but it cannot solve the problem, for two linked reasons.

First, the amount of energy lost depends steeply on scattering angle. The Compton relation for the scattered photon energy $E^{'}$ from an incident photon $E = 511$ keV is:

E^{'} = \frac{E}{1 + \frac{E}{m _{e} c ^{2}} ( 1 - cos θ )}

Since $m_{e} c^{2} = 511$ keV for the electron, this simplifies neatly at 511 keV to $E^{'} = 511/ (2 - cos θ)$ . Evaluating it shows how little energy small-angle scatter loses:

θ = 3 0^{\circ} : E^{'} = \frac{511}{2 - 0.866} = 451 keV

θ = 6 0^{\circ} : E^{'} = \frac{511}{2 - 0.500} = 341 keV

A photon scattered by 30° still carries 451 keV — comfortably inside any clinically usable energy window (for example 435–585 keV on many systems). Only large-angle scatter loses enough energy to be rejected outright. Second, real detectors have finite energy resolution (often 10–15% at 511 keV), so the photopeak is a broad bump, not a spike, and the lower edge of the window unavoidably admits scattered events while a too-tight window would start throwing away true photopeak counts. The energy window suppresses scatter; it does not eliminate it. A computational estimate is unavoidable. ^{1, 3}

Key Technical Principles

Single-scatter simulation: the workhorse

Single-scatter simulation (SSS) is the standard clinical scatter-correction algorithm, and it works by directly computing the scatter distribution from first principles rather than measuring it. The method — introduced in its efficient, image-based form by Watson and by Ollinger in the mid-to-late 1990s — takes three inputs already available in a PET/CT study: the CT-derived attenuation map (which gives the electron density everywhere in the patient), a first-pass reconstructed emission image (the activity distribution), and a geometric model of the scanner. ^{1, 2, 3}

From these, SSS integrates over candidate scatter points in the patient. For each point, it uses the Klein-Nishina differential cross section — the exact quantum-electrodynamic probability that a photon scatters into a given angle — to compute how many coincidences will be created in which one of the two photons scattered once at that point:

\frac{d σ}{d Ω} = \frac{r _{e}^{2}}{2} (\frac{E ^{'}}{E})^{2} (\frac{E}{E ^{'}} + \frac{E ^{'}}{E} - sin^{2} θ)

where $r_{e}$ is the classical electron radius and $E^{'} / E$ is the energy ratio from the Compton relation above. Summing these contributions over all scatter points and all detector pairs produces a simulated scatter sinogram with the correct shape. ^{1, 3}

Tail fitting: getting the magnitude right

SSS accurately predicts the shape of the scatter distribution but not its absolute magnitude, because it models only single scatter and makes simplifying assumptions. The magnitude is fixed by a clever trick: outside the patient's body, in the tails of the sinogram, there can be no true coincidences (there is no activity and no attenuator on those lines of response). Any counts there must be scatter (plus randoms, already removed). So the simulated scatter is scaled to match the measured counts in the tails — "tail fitting" — which anchors the whole estimate to real data. ^{3, 4}

The scaled scatter sinogram is then incorporated into the iterative reconstruction (typically as an additive term in the forward model of an OSEM-type algorithm), so the reconstruction naturally accounts for it rather than crudely subtracting it. Because SSS depends on the emission image, and the emission image improves once scatter is accounted for, the process is iterated a small number of times. Watson's image-based reformulation is what made this fast enough for routine whole-body use — on the order of tens of seconds per bed position rather than the impractical times of earlier approaches. ³

Scatter fraction: measuring the size of the problem

How much scatter is there? The standardized answer is the scatter fraction (SF), defined as the ratio of scattered to total (true plus scattered) coincidences:

SF = \frac{S}{T + S}

Scatter fraction is measured under the NEMA NU 2 protocol using a line source inside a 20 cm diameter cylindrical phantom, and it is one of the headline acceptance-test numbers for any PET system. ⁸ On contemporary 3D PET/CT scanners, SF typically falls in the range of about 33% to 40% — meaning more than a third of the recorded coincidences are scattered. [4, and see representative NU 2 evaluations]

A worked illustration of what that ratio means: if a phantom acquisition records 100 arbitrary units of true-plus-scattered coincidences at a scatter fraction of 0.36, then

S = 0.36 \times 100 = 36, T = 100 - 36 = 64

so 36 of every 100 accepted events are pointing at the wrong line of response. Removing that 36% correctly is the difference between a quantitative image and a decorative one.

Scatter-correction method	Core principle	Speed	Accuracy / limitation
Energy window only	Reject low-energy events below the photopeak	Instant	Suppresses only large-angle scatter; insufficient alone ¹
Dual/triple energy window	Estimate scatter from a secondary lower window	Fast	Noisy; sensitive to energy calibration ¹
Single-scatter simulation (SSS)	Klein-Nishina forward model + tail fitting	~seconds/bed	Clinical standard; weak for multiple scatter and out-of-FOV activity ^{3, 4}
Monte Carlo	Full stochastic photon transport	Slow (GPU helps)	Most accurate; historically too slow for routine use ⁷
Deep learning / energy-based	Learned or spectral scatter estimate	Fast at inference	Emerging; aims to fix SSS failure modes ^{4, 6}

Clinical Impact

What uncorrected scatter does to an image

Scatter's signature is a low, spatially smooth background that fills the whole field of view. Clinically this does three damaging things: it reduces the contrast of hot lesions (the peak sits on a raised pedestal), it puts apparent activity into truly cold regions (a lung or an air cavity appears to glow faintly), and it biases every SUV upward or downward depending on the local geometry. Because the effect is diffuse and plausible-looking, it is far more insidious than a sharp artifact — the image looks fine while the numbers lie. ^{1, 3}

The quantitative payoff — and the harmonization stakes

Scatter correction is one of the non-negotiable prerequisites for quantitative PET. The EANM tumor-imaging guidelines and the broader harmonization enterprise exist precisely so that a SUV measured on one scanner equals a SUV measured on another; that equivalence is impossible if scatter — which varies with patient size, activity distribution, and scanner geometry — is not consistently removed. ¹⁰ Modern method comparisons quantify how good the correction now is: in NEMA-style phantom evaluations, well-tuned SSS leaves only a small residual scatter signal (on the order of a few percent), and newer energy-based estimation has been reported to trim that residual further while producing lesion SUVs that agree with SSS to within about 0.1 SUV units. ⁴ The takeaway is not that one method wins, but that the floor of accuracy is now set by how well scatter is handled.

When scatter correction goes wrong

The failure modes are clinically recognizable and worth knowing:

CT–PET misregistration from motion. SSS relies on the CT attenuation map aligning with the emission activity. If the patient moves — or breathes differently — between the CT and the PET, the scatter estimate is computed on the wrong geometry, producing cold artifacts next to intense sources. This is exactly the failure Magota and colleagues documented for high-activity sources such as a face mask in brain PET, where conventional tail-fitted SSS underestimated activity by tens of percent and a Monte Carlo–scaled variant fixed it. ⁵
Activity outside the field of view. A full bladder or an injection-site hot spot just beyond the axial FOV contributes scatter that SSS, which only models activity it can "see," misestimates. ^{4, 6}
Truncation and large patients. When the patient extends beyond the CT or PET transverse FOV, the attenuation and activity maps are truncated, and the tails used for scaling may be corrupted. ⁴

Each of these is a reason the medical physicist should review scatter estimates and reconstructed uniformity as part of routine PET quality control, alongside the daily QC and calibration checks.

Practical Optimization Tips

1. Keep the patient still between CT and PET

The single most preventable scatter-correction failure is CT–PET misregistration. Coach breathing, immobilize where appropriate, and minimize the delay between the CT and the emission scan so the attenuation map used by SSS actually matches the activity. ⁵

2. Watch for out-of-FOV hot sources

Have patients void before scanning to remove a bladder that sits at the edge of the FOV, and be alert to injection-site extravasation and paravenous activity that can distort the scatter estimate. ^{4, 6}

3. Review uniform regions on every study

Cold structures — lung, trachea, air — should read near zero. A faint diffuse fill in these regions is the fingerprint of imperfect scatter correction and warrants a second look before SUVs are trusted. ^{1, 3}

4. Confirm scatter fraction at acceptance and after major service

Measure scatter fraction per NEMA NU 2 at acceptance, and re-check after detector or electronics service. A scatter fraction that has drifted well outside the expected 33–40% band is a signal that something in the energy calibration or detector chain has changed. ⁸

5. Don't tighten the energy window blindly

A narrower photopeak window rejects more scatter but also discards true photopeak counts and can destabilize the energy-dependent corrections. Any window change should be validated against phantom scatter fraction and sensitivity, not adjusted by feel. ¹

6. Know your scanner's algorithm

SSS implementations differ by vendor in how they handle multiple scatter, out-of-FOV activity, and time-of-flight information. Understanding which variant your system runs — and its known weak spots — is part of interpreting borderline studies correctly. ^{3, 4}

Regulatory Considerations

Scatter correction is not the subject of a standalone regulation, but it lives inside the performance-standard and accreditation framework that governs clinical PET. The relevant anchors are standards and guidance rather than statutes:

NEMA NU 2-2024 — the current edition of the standard that defines how PET performance, including scatter fraction and count-rate performance, is measured and reported; it supersedes NEMA NU 2-2018. Manufacturers cite it in specifications, and physicists use its procedures for acceptance testing. ⁸
IAEA Human Health Series No. 1 — international guidance on quality assurance for PET and PET/CT systems, including the corrections that underpin quantitation. ⁹
EANM FDG PET/CT tumour imaging guidelines, version 2.0 — the widely used procedure and harmonization guidance that makes accurate scatter (and attenuation) correction a prerequisite for comparable SUVs. ¹⁰
AAPM acceptance-testing and QC guidance for PET/CT — the professional-society framework U.S. medical physicists follow for commissioning and periodic testing. ^{8, 9}

On the regulatory side that does carry the force of law: the radioactive material used in PET (F-18, Ga-68, and others) is governed by NRC 10 CFR Part 20 and Part 35 or the equivalent Agreement State program, while the CT subsystem's x-ray output is regulated by the FDA and the state radiation-control program. Of the states DRPS serves, Florida, Maryland, Virginia, California, Nevada, Pennsylvania, New York, and New Jersey are NRC Agreement States, while Washington DC and Delaware are regulated directly by the NRC for radioactive material. Scatter correction quality is what the accreditation and quantitation layer cares about; the licensing layer governs the isotopes that make the scan possible. DRPS integrates both through its PET/CT and nuclear medicine physics service and accreditation support. ^{8, 9, 10}

Frequently Asked Questions (FAQs)

What is scatter correction in PET?

Scatter correction is the process of estimating and removing the contribution of Compton-scattered photons from PET data before or during image reconstruction. Scattered coincidences are mispositioned events that add a diffuse background to the image; correcting for them restores image contrast and the quantitative accuracy of standardized uptake values (SUVs).

What is single-scatter simulation (SSS)?

Single-scatter simulation is the standard PET scatter-correction algorithm. It uses the CT-derived attenuation map and a first-pass emission image to directly compute, from the Klein-Nishina Compton physics, the expected distribution of coincidences in which one of the two photons scattered once. The simulated scatter is then scaled to the measured data using the counts in the sinogram tails and subtracted.

Why is scatter correction important for SUV accuracy?

Uncorrected scatter adds a low, spatially smooth background across the field of view, which inflates apparent activity in cold regions and reduces contrast in hot lesions. Because the SUV is a ratio of measured activity concentration to injected dose per body weight, that background biases the number. Accurate scatter correction is a prerequisite for SUVs that are comparable across scanners, sites, and time points.

What is scatter fraction in PET?

Scatter fraction is the ratio of scattered coincidences to the total of true plus scattered coincidences, measured with a standard line source in a cylindrical phantom under the NEMA NU 2 protocol. On modern 3D PET/CT systems it is typically in the range of about 33% to 40%, meaning more than a third of the events that pass the energy and timing windows are scattered.

Does the energy window remove scatter?

Only partially. A tighter energy window around the 511 keV photopeak rejects large-angle scatter, which loses the most energy, but small-angle Compton scatter loses very little energy and falls inside any clinically usable window. Detector energy resolution is finite, so the energy window alone cannot separate all scatter from true events, which is why a computational scatter estimate is still required.

What happens if scatter correction fails?

When scatter correction fails — for example from patient motion between the CT and PET scans, high activity just outside the field of view, or truncation — the image can show cold artifacts, halo effects around intense sources, and biased SUVs. This is why scatter estimates should be reviewed as part of PET quality control, not assumed to be correct.

Is single-scatter simulation being replaced?

SSS remains the clinical workhorse, but its known limitations with multiple scatter and activity outside the field of view have driven research into Monte Carlo scatter estimation, energy-based methods, and deep-learning approaches. These aim for higher accuracy in the cases where SSS struggles, but SSS is still the default on most installed PET/CT systems.

Key Takeaways

Scatter is a dominant background, not a nuisance. On modern 3D PET/CT, scatter fraction is typically about 33–40%, so more than a third of accepted coincidences are mispositioned. ^{4, 8}
The energy window cannot fix it. Small-angle scatter loses too little energy (a 30° scatter still carries 451 keV), and finite detector energy resolution guarantees scatter inside the photopeak window. ¹
Single-scatter simulation is the clinical standard. It computes the scatter shape from Klein-Nishina physics using the CT attenuation map and emission image, then fixes the magnitude by fitting the sinogram tails. ^{2, 3}
Scatter correction is a quantitation prerequisite. Without it, SUVs are neither accurate nor comparable across sites — undermining staging, response assessment, and multicenter harmonization. ¹⁰
Know the failure modes. CT–PET motion misregistration, out-of-FOV activity, and truncation are the classic ways scatter correction breaks and produce recognizable artifacts. ^{4, 5}
Newer methods are coming, but SSS is the default. Monte Carlo, energy-based, and deep-learning scatter estimation target SSS's weak spots, yet SSS still runs on most installed systems. ^{4, 6, 7}

Conclusion

Scatter correction is one of the quiet corrections that separates PET as a quantitative science from PET as a picture-making device. A third or more of the coincidences a scanner accepts are scattered, and because those events masquerade as signal in every dimension except position, only a physics-based estimate — single-scatter simulation, anchored to the sinogram tails — can strip them out. Done well, it delivers clean images and SUVs that mean the same thing everywhere. Done poorly, it leaves a plausible-looking image sitting on a lie.

For the medical physicist, the job is to verify that scatter correction is working: measure scatter fraction at acceptance, watch cold regions on clinical studies, guard against the motion and out-of-FOV conditions that break it, and understand the specific algorithm the scanner runs. Quantitative PET is only as trustworthy as its corrections, and scatter is one of the largest.

How DRPS Can Help

Diagnostic Radiation Physics Services provides PET/CT physics support that keeps quantitative imaging accurate and defensible: NEMA NU 2 acceptance testing including scatter fraction and count-rate performance, SUV calibration and harmonization checks, review of scatter and attenuation correction on clinical protocols, artifact investigation, and periodic QC. DRPS delivers this through its PET/CT and nuclear medicine physics service, with accreditation support and medical physics consulting as needed.

DRPS supports facilities across our service locations, including Florida, Maryland, Virginia, Washington DC, California, Nevada, New York, Pennsylvania, New Jersey, and Delaware.

Accurate SUVs start with corrections that are verified, not assumed.

Related Resources

References

Zaidi H, Koral KF. Scatter modelling and compensation in emission tomography. Eur J Nucl Med Mol Imaging. 2004;31(5):761-782. doi:10.1007/s00259-004-1495-z. PubMed
Ollinger JM. Model-based scatter correction for fully 3D PET. Phys Med Biol. 1996;41(1):153-176. doi:10.1088/0031-9155/41/1/012. PubMed
Watson CC. New, faster, image-based scatter correction for 3D PET. IEEE Trans Nucl Sci. 2000;47(4):1587-1594. doi:10.1109/23.873020. doi.org
Hamill JJ, Cabello J, Surti S, Karp JS. Energy-based scatter estimation in clinical PET. Med Phys. 2024;51(1):54-69. doi:10.1002/mp.16826. PubMed
Magota K, Shiga T, Asano Y, et al. Scatter correction with combined single-scatter simulation and Monte Carlo simulation scaling improved the visual artifacts and quantification in 3-dimensional brain PET/CT imaging with 15O-gas inhalation. J Nucl Med. 2017;58(12):2020-2025. doi:10.2967/jnumed.117.193060. PubMed
Laurent B, Bousse A, Merlin T, Nekolla S, Visvikis D. PET scatter estimation using deep learning U-Net architecture. Phys Med Biol. 2023;68(6):065004. doi:10.1088/1361-6560/ac9a97. PubMed
Galve P, Arias-Valcayo F, Villa-Abaunza A, Ibáñez P, Udías JM. UMC-PET: a fast and flexible Monte Carlo PET simulator. Phys Med Biol. 2024;69(3):035006. doi:10.1088/1361-6560/ad1cf9. PubMed
National Electrical Manufacturers Association. NEMA Standards Publication NU 2-2024: Performance Measurements of Positron Emission Tomographs (PET). 2024. nema.org
International Atomic Energy Agency. Quality Assurance for PET and PET/CT Systems. IAEA Human Health Series No. 1. 2009. iaea.org
Boellaard R, Delgado-Bolton R, Oyen WJG, et al. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. Eur J Nucl Med Mol Imaging. 2015;42(2):328-354. doi:10.1007/s00259-014-2961-x. PubMed