PET PSF Reconstruction: Resolution Recovery

PSF reconstruction, also called resolution modeling, sharpens PET images by building the scanner's measured blur directly into the reconstruction — improving contrast recovery and lesion detectability, but at the cost of edge overshoot artifacts and reduced small-lesion SUV reproducibility. Understanding that trade-off is what separates a sharper picture from a defensible quantitative measurement.

Modern PET/CT scanners from every major vendor ship with a point spread function (PSF) reconstruction option, often bundled with time-of-flight (TOF) and marketed under names such as HD·PET or TrueX, SharpIR, or similar. Turned on by default, it makes images look crisper and lesions pop. Turned on without understanding, it can inflate a small-lesion SUV, plant a bright rim that mimics disease, and quietly break the comparability of longitudinal scans. This guide explains what resolution modeling actually does, where it helps, where it hurts, and how a medical physicist should validate and govern it. DRPS provides this analysis as part of its PET/CT and nuclear medicine physics services across Florida, Maryland, Virginia, Washington DC, California, and Nevada.

Introduction

Spatial resolution has always been the fundamental limit of PET. A whole-body clinical scanner resolves objects on the order of 4–6 mm, which means that a large fraction of clinically important lesions — sub-centimeter lymph nodes, small pulmonary nodules, early bone metastases — are smaller than the system can faithfully depict. When an object is comparable to or smaller than the resolution element, its measured activity concentration is spread into surrounding voxels and diluted: the partial volume effect. ¹

For two decades, PET reconstruction moved from analytic filtered back projection to statistical iterative methods (OSEM), then added TOF information, and then added resolution modeling. PSF reconstruction is the step that tries to undo some of the blurring that the detector physics imposes, by telling the reconstruction algorithm exactly how the scanner smears a point of activity across the image. ¹²

The appeal is obvious: better resolution, better contrast, better small-lesion detection, all in software, on hardware the facility already owns. The catch is equally real: modeling resolution changes the noise texture, introduces edge artifacts, and alters the quantitative values (SUVs) that oncologists increasingly rely on for staging and response assessment. This is not a reason to avoid PSF — it is a reason to understand it. ¹³

Topic Explanation

What is the point spread function?

The point spread function is the image a scanner produces from an idealized point source of activity. In a perfect system, a point would map to a point. In a real PET scanner, that point is blurred into a roughly Gaussian blob several millimeters wide, and the width and shape of that blob change with position in the field of view. Several physical effects contribute:

Positron range — the positron travels a short distance before annihilating, so the annihilation photons originate away from the decay. For our overview of this limit, see PET spatial resolution and positron range.
Photon non-collinearity — the two 511 keV photons are not emitted exactly 180° apart, blurring localization more in larger-diameter scanners.
Detector element size — the finite crystal width sets an intrinsic sampling limit.
Depth of interaction (parallax) — photons entering a crystal obliquely can be mispositioned, worsening resolution toward the edge of the field of view.

Because these effects are measurable, the scanner's PSF can be characterized — for example by scanning point sources at many positions — and then represented mathematically inside the reconstruction. ²⁵

What is resolution modeling?

Resolution modeling means incorporating the measured PSF into the system matrix of an iterative reconstruction so that the algorithm accounts for blur while it estimates the activity distribution. Instead of assuming each line of response maps cleanly to a voxel, the reconstruction knows that a voxel's counts are spread across neighbors according to the PSF, and it works backward toward a sharper estimate. ¹²

Two broad implementations exist: projection-space modeling, where the blur is applied along the lines of response, and image-space modeling, where the blur is represented as a convolution kernel in the reconstructed image. Vendor products differ in the details, in whether the PSF is measured or parameterized, and in how they control the resulting noise and artifacts. ²⁵ Resolution modeling is distinct from — and often combined with — the noise-control approach of Bayesian penalized-likelihood reconstruction, which regularizes the image rather than modeling blur.

Key Technical Principles

The reconstruction model

Iterative PET reconstruction estimates the activity image $λ$ that best explains the measured coincidence data $y$ . The expected counts in detector pair $i$ are modeled as:

\overset{y}{ˉ}_{i} = j \sum H_{ij} λ_{j} + \overset{s}{ˉ}_{i} + \overset{r}{ˉ}_{i}

where $H_{ij}$ is the system matrix element (the probability that a decay in voxel $j$ is detected in line of response $i$ ), and $\overset{s}{ˉ}_{i}$ and $\overset{r}{ˉ}_{i}$ are the expected scatter and random contributions. The maximum-likelihood expectation-maximization (MLEM) update, of which OSEM is the ordered-subset acceleration, is:

λ_{j}^{(n + 1)} = \frac{λ _{j}^{(n)}}{\sum _{i} H _{ij}} i \sum H_{ij} \frac{y _{i}}{\sum _{k} H _{ik} λ _{k}^{(n)} + s ˉ _{i} + r ˉ _{i}}

Resolution modeling enters through the system matrix. Without PSF, $H$ contains only geometry and attenuation. With PSF, the system matrix is factored to include a resolution kernel $R$ that describes the measured blur, conceptually $H = A R$ , so the algorithm deconvolves the scanner response as it iterates. ¹²

The physical resolution budget

The system resolution that PSF modeling tries to recover is itself the combination of the physical blurring terms. A widely used approximation for the reconstructed FWHM of a ring PET system is:

FWHM_{sys} \approx (\frac{d}{2})^{2} + b^{2} + (0.0022 D)^{2} + r^{2}

where $d$ is the detector element width, $b$ is a decoding/positioning term, $0.0022 D$ is the non-collinearity contribution for a ring diameter $D$ (mm), and $r$ is the effective positron range. Resolution modeling cannot beat physics — it recovers resolution toward these limits by compensating for the measured detector response, not below them. ¹

Contrast recovery and the recovery coefficient

The clinically meaningful metric is the recovery coefficient (RC), the ratio of measured to true activity concentration for an object of a given size:

R C = \frac{C _{measured}}{C _{true}}

For non-PSF images, RC falls smoothly toward zero as objects shrink below the resolution limit — the classic partial volume roll-off. Resolution modeling raises RC, especially for medium and small objects. Phantom and brain studies have reported RC improvements on the order of 11–40% with PSF modeling on high-resolution systems, with corresponding gains in gray-to-white contrast and receptor-binding estimates. ⁴

The catch: edge overshoot (Gibbs artifact)

Deconvolution of a band-limited system near a sharp edge produces ringing — the Gibbs phenomenon. In PSF PET this appears as edge overshoot: bright rims at high-contrast boundaries and, critically, a non-monotonic RC curve. Instead of RC always falling as objects get smaller, PSF images can show a peak RC greater than one for intermediate object sizes. One careful phantom study found that recovery coefficients for spheres did not fall monotonically, that RC could exceed one, and that maximum RC occurred for spheres around 8 mm — meaning a small lesion can read hotter than its true concentration. ³

The comparison below summarizes the main clinical reconstruction options and where each sits on the sharpness-versus-quantification spectrum.

Reconstruction	Primary benefit	Main limitation	Quantitative behavior
OSEM (baseline)	Robust, well-understood noise	Lower resolution; partial volume roll-off	RC falls monotonically with size; most reproducible for small ROIs
OSEM + TOF	Better SNR and convergence, esp. large patients	Modest resolution change	Improved contrast; stable quantification
PSF (+ TOF)	Higher resolution and contrast recovery	Edge overshoot (Gibbs); noise amplification	RC can exceed 1 and peak at intermediate sizes; small-lesion SUV inflated and less reproducible ³
Bayesian penalized likelihood (BPL)	Full convergence with controlled noise	Vendor-specific β tuning	Improved CR/SNR trade-off; behavior depends on β ⁶

Noise and reproducibility

Resolution modeling increases correlations between neighboring voxels. That improves visual appearance but can reduce the precision (reproducibility) of quantitative measurements in small regions of interest, because a small ROI now samples a more spatially correlated, sharper, and noisier estimate. Comprehensive reviews of resolution modeling explicitly flag this loss of precision for small ROIs as a central pitfall of the technique. ¹ Comparative work has shown that penalized-likelihood methods can offer a more favorable contrast-recovery-to-noise trade-off than PSF+TOF for some detector configurations, underscoring that "sharper" is not synonymous with "more accurate." ⁶

Clinical Impact

The clinical value of PSF reconstruction is real but task-dependent. For lesion detection — the radiologist's yes/no question of whether a focus of uptake is present — resolution modeling generally helps. Sharper, higher-contrast images make small nodes and nodules more conspicuous, and observer studies have supported detectability gains. ¹

For lesion quantification — the oncologist's question of exactly how avid a lesion is, and whether it changed — the picture is more nuanced:

Staging SUVs may read higher on PSF images than on the OSEM images a facility historically used. A treatment-naïve lesion that would have been SUV max 4.5 on OSEM might read 5.5–6.5 on a sharp PSF preset, purely from reconstruction.
Small-lesion values are the most affected: the closer a lesion is to the size where the RC curve peaks, the more its SUV can be inflated by overshoot. ³
Response monitoring across time points, or across scanners in a multicenter trial, is corrupted if the reconstruction changes between scans. A "response" or "progression" can be a reconstruction artifact.

This is why quantitative PET frameworks — PERCIST for response, and multicenter oncology trials — insist on consistent, harmonized acquisition and reconstruction. See our discussion of SUV quantification and the partial volume effect for the quantitative context that PSF modeling sits inside.

Practical Optimization Tips

1. Run two reconstructions

The most robust operational answer is to reconstruct both a sharp clinical PSF (often PSF+TOF) image for reading and a harmonized quantitative image for SUV reporting. The clinical read benefits from resolution modeling; the quantitative values come from a reproducible, comparable reconstruction.

2. Generate an EARL-compliant preset for quantitative reporting

The EANM Research Ltd. (EARL) FDG PET/CT accreditation program defines contrast-recovery limits, measured on the NEMA image quality phantom, that harmonize SUVs across scanners and sites. Because PSF reconstruction raises recovery coefficients out of the EARL envelope, sites typically apply a Gaussian post-smoothing filter or a dedicated harmonized preset to bring recovery coefficients back within EARL limits for quantitative work. ⁶ For the harmonization background, see EARL PET SUV harmonization.

3. Fix the reconstruction for longitudinal studies

For any patient being followed over time, lock the reconstruction algorithm, iterations, subsets, filters, and matrix size, and keep them identical at every time point. Document the preset in the report so a downstream reader knows what generated the SUV.

4. Characterize your presets on the NEMA phantom

Measure contrast recovery coefficients and background variability for each reconstruction preset across the full range of NEMA sphere sizes. This is the single most useful physics deliverable: it shows exactly where a given PSF preset overshoots, where RC peaks, and whether a "quantitative" preset meets harmonization limits. Track these values as part of the NEMA NU 2 performance program.

5. Beware the small, bright rim

Educate readers that a thin hyperintense rim around a photopenic or high-contrast structure on a PSF image may be an edge artifact, not disease. Correlate with CT and with a non-PSF or smoothed reconstruction when in doubt.

Common pitfalls to avoid

Treating PSF SUVs as interchangeable with legacy OSEM SUVs. They are not; a step-change in reported SUV can be pure reconstruction.
Comparing scans reconstructed differently. Response assessment requires identical processing.
Using a single sharp preset for everything. Detection and quantification are different tasks with different optimal settings.
Skipping phantom characterization. Without contrast-recovery curves, a facility cannot say how its PSF preset behaves.
Ignoring lesion size. The quantitative bias from overshoot is largest exactly where small-lesion decisions are made.

Regulatory Considerations

PET reconstruction settings sit at the intersection of accreditation, quantitative harmonization, and the facility's radioactive material program. While reconstruction algorithm choice is not directly dictated by federal rule, several frameworks govern the environment in which it operates:

NEMA NU 2-2018 defines the standardized performance measurements — spatial resolution, sensitivity, count-rate performance, image quality, and, in the current edition, time-of-flight resolution and PET/CT co-registration — that let physicists characterize a scanner and its reconstructions consistently. ⁷
ACR PET/CT Accreditation and equivalent programs require phantom-based image quality and SUV accuracy testing by a qualified medical physicist, which is where reconstruction presets are validated.
EANM/EARL FDG PET/CT accreditation provides the international harmonization standard for quantitative SUVs used in multicenter oncology and therapy monitoring. ⁶
10 CFR Part 35 (or the equivalent Agreement State program) governs the medical use of the F-18 and other positron-emitting radiopharmaceuticals that generate the data, along with the dose calibrator and instrument QC that underpins any quantitative claim. ⁸

Of the states DRPS serves, Florida, Maryland, Virginia, California, and Nevada are NRC Agreement States that license medical use under their own radiation-control rules, while Washington, DC is regulated directly by the NRC. A defensible quantitative PET program documents its reconstruction presets, their phantom-measured recovery behavior, and their harmonization status so the numbers in a report can be defended during accreditation review or a clinical-trial audit.

Frequently Asked Questions (FAQs)

What is PSF reconstruction in PET?

PSF reconstruction, also called resolution modeling, is an iterative PET reconstruction technique that incorporates a measured or modeled description of the scanner's spatially varying blur (the point spread function) into the system matrix. By accounting for how the scanner blurs a point source, the algorithm partially recovers spatial resolution, improving contrast and edge definition compared with standard OSEM.

Does PSF reconstruction improve PET image quality?

For most clinical reading tasks it does. Resolution modeling improves apparent spatial resolution, contrast recovery, and small-lesion detectability, and it can reduce the partial volume effect. The trade-off is that PSF reconstruction can introduce edge overshoot (Gibbs) artifacts and can reduce the reproducibility of SUV measurements in very small structures, so it is not automatically better for every task.

What is the Gibbs or edge artifact in PSF PET?

Resolution modeling can produce ringing or overshoot near sharp activity boundaries, an effect related to the Gibbs phenomenon. This can create bright rims around lesions and can push the measured concentration of small objects above the true value, so a small lesion may show a recovery coefficient greater than one and an artificially elevated SUV max.

Why can PSF reconstruction change SUV values?

PSF reconstruction increases contrast recovery, which raises SUV in small and medium lesions relative to non-PSF images. Because the amount of increase depends on lesion size, count statistics, and reconstruction parameters, SUVs are not directly comparable between PSF and non-PSF reconstructions or between differently tuned PSF settings, which is why harmonization matters.

Is PSF reconstruction compatible with EARL harmonization?

Yes, but it requires care. Because PSF changes recovery coefficients, sites that report quantitative SUVs for multicenter studies or therapy monitoring often generate an EARL-compliant reconstruction (frequently by applying a Gaussian post-filter or a dedicated harmonized preset) alongside the sharper clinical PSF image, so quantitative values remain comparable across scanners.

Should PSF be used for treatment-response monitoring?

Use it deliberately. Because PSF can reduce SUV reproducibility for small regions and can bias small-lesion values upward, longitudinal response assessment should use the same reconstruction and analysis method at every time point, and ideally a harmonized quantitative reconstruction, so that changes reflect biology rather than reconstruction settings.

How do we validate PSF reconstruction during PET QC?

Validation is done with phantom measurements, typically the NEMA NU 2 image quality phantom, measuring contrast recovery coefficients and background variability across sphere sizes for each reconstruction preset. Tracking these values, along with SUV recovery, lets the medical physicist document how a PSF preset behaves and confirm that quantitative presets meet harmonization limits.

Key Takeaways

PSF reconstruction models the scanner's measured blur and partially recovers spatial resolution, improving contrast recovery and lesion detectability.
It cannot beat physics. Resolution modeling recovers toward the physical resolution limits set by positron range, non-collinearity, and detector size — not below them.
Edge overshoot (Gibbs) is the signature artifact. Recovery coefficients can exceed one and peak for intermediate object sizes, inflating small-lesion SUVs. ³
Quantification and detection are different tasks. PSF often helps detection while complicating small-ROI SUV reproducibility. ¹
Harmonize for quantitative reporting. Use an EARL-compliant or otherwise harmonized reconstruction, and keep it fixed across longitudinal and multicenter studies. ⁶
Characterize presets on the NEMA phantom. Contrast-recovery curves are the physics evidence that documents how a preset behaves.

Conclusion

PSF reconstruction is one of the most useful and most misunderstood tools in modern PET. It genuinely sharpens images and improves detection, and for that reason it belongs in the clinical workflow. But resolution modeling is a form of deconvolution, and deconvolution near sharp edges rings. The result is a technique that can make a small lesion look brighter than it truly is, and that can silently break the comparability of the SUVs oncology depends on.

The right posture is neither to avoid PSF nor to trust it blindly, but to govern it: read on a sharp preset, quantify on a harmonized one, fix the reconstruction across time, and characterize every preset on the phantom bench. Handled that way, resolution modeling delivers its benefits while keeping quantitative PET honest.

How DRPS Can Help

Diagnostic Radiation Physics Services helps PET/CT and nuclear medicine facilities turn reconstruction settings into defensible imaging. Our PET/CT and nuclear medicine physics support includes NEMA NU 2 performance testing, contrast-recovery characterization of PSF and harmonized presets, EARL and ACR quantitative validation, SUV accuracy checks, and reconstruction-protocol review — delivered by board-certified medical physicists through our medical physics consulting service.

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

A sharper PET image is only an improvement if the numbers on it can still be trusted. We help make sure they can.

Related Resources

References

Rahmim A, Qi J, Sossi V. Resolution modeling in PET imaging: theory, practice, benefits, and pitfalls. Medical Physics. 2013;40(6):064301. doi:10.1118/1.4800806. doi.org
Rapisarda E, Bettinardi V, Thielemans K, Gilardi MC. Image-based point spread function implementation in a fully 3D OSEM reconstruction algorithm for PET. Physics in Medicine and Biology. 2010;55(14):4131-4151. doi:10.1088/0031-9155/55/14/012. doi.org
Munk OL, Tolbod LP, Hansen SB, Bogsrud TV. Point-spread function reconstructed PET images of sub-centimeter lesions are not quantitative. EJNMMI Physics. 2017;4(1):5. doi:10.1186/s40658-016-0169-9. doi.org
Varrone A, Sjöholm N, Eriksson L, Gulyás B, Halldin C, Farde L. Advancement in PET quantification using 3D-OP-OSEM point spread function reconstruction with the HRRT. European Journal of Nuclear Medicine and Molecular Imaging. 2009;36(10):1639-1650. doi:10.1007/s00259-009-1156-3. doi.org
Schmitt D, Karuta B, Carrier C, Lecomte R. Fast point spread function computation from aperture functions in high-resolution positron emission tomography. IEEE Transactions on Medical Imaging. 1988;7(1):2-12. doi:10.1109/42.3923. doi.org
Rogasch JM, Suleiman S, Hofheinz F, et al. Reconstructed spatial resolution and contrast recovery with Bayesian penalized likelihood reconstruction (Q.Clear) for FDG-PET compared to time-of-flight (TOF) with point spread function (PSF). EJNMMI Physics. 2020;7(1):2. doi:10.1186/s40658-020-0270-y. doi.org
National Electrical Manufacturers Association. NEMA NU 2-2018: Performance Measurements of Positron Emission Tomographs (PET). Rosslyn, VA: NEMA; 2018. nema.org
U.S. Nuclear Regulatory Commission. 10 CFR Part 35: Medical Use of Byproduct Material. ecfr.gov
Aide N, Lasnon C, Veit-Haibach P, Sera T, Sattler B, Boellaard R. EANM/EARL harmonization strategies in PET quantification: from daily practice to multicentre oncological studies. European Journal of Nuclear Medicine and Molecular Imaging. 2017;44(Suppl 1):17-31. doi:10.1007/s00259-017-3740-2. doi.org
American College of Radiology. ACR Accreditation (Nuclear Medicine and PET). acr.org