SPECT Reconstruction: FBP vs Iterative OSEM

SPECT reconstruction is the step that turns raw angular projection data into cross-sectional images, and it is done two fundamentally different ways: analytic filtered back projection (FBP), which is fast but noisy and blind to imaging physics, and iterative OSEM, which converges on a statistically consistent image while modeling attenuation, scatter, and collimator blur. The choice of algorithm and its settings shapes image noise, quantitative accuracy, and every artifact a reader has to interpret. ^{1, 2, 4}

For modern SPECT and SPECT/CT, iterative reconstruction with ordered-subsets expectation maximization (OSEM) has become the clinical standard because it enables corrections and quantitation that FBP simply cannot support. But FBP is not obsolete knowledge: it explains the streaks, the negative values, and the filter choices that still appear across nuclear medicine, and it remains the conceptual baseline against which iterative methods are judged. ^{3, 4, 5}

Introduction

Every SPECT study begins as a set of planar projections acquired as the gamma camera rotates around the patient. Reconstruction is the mathematics that recovers the three-dimensional activity distribution from those projections. It is also where much of the image quality — and nearly all of the quantitative capability — of a SPECT system is won or lost.

Two families of algorithm dominate. Analytic reconstruction, embodied by filtered back projection, treats reconstruction as a direct mathematical inversion of the projection process. Iterative reconstruction, embodied by MLEM and its accelerated form OSEM, treats reconstruction as a statistical estimation problem: find the activity distribution that, when passed through a model of the imaging system, best reproduces the measured counts.

This article walks through the physics and mathematics of both, compares their noise, artifact, and quantitative behavior, and explains how the subsets-and-iterations settings on your console actually work. DRPS supports nuclear medicine programs on exactly these questions as part of its PET/CT and nuclear medicine physics and medical physics consulting services across Florida, Maryland, Virginia, Washington DC, California, and Nevada.

Topic Explanation

From projections to images

As the camera rotates, it records a projection at each angle — essentially a line-integral of the activity along each ray reaching the detector. The complete set of projections across angles is the Radon transform of the activity distribution. Reconstruction inverts this relationship. ⁴

Filtered back projection does the inversion analytically. Each projection is filtered in the frequency domain and then "back projected" — smeared uniformly back along the direction it was acquired. Summing the filtered back projections over all angles reconstructs the image. The filtering step is essential: simple back projection without it produces a badly blurred image, because each point spreads a 1/r halo across the field of view. ⁴

Iterative reconstruction does not invert anything directly. It starts from an initial guess (often a uniform image), forward-projects that guess to predict what the camera should have measured, compares the prediction to the actual projections, and updates the image to reduce the discrepancy. Repeating this loop drives the image toward statistical consistency with the data. Crucially, the forward model — the system matrix — can incorporate the physics of the acquisition: attenuation, scatter, and the distance-dependent blur of the collimator. ^{1, 8}

Why the difference matters clinically

The practical consequences are large. FBP is fast and linear, but it amplifies noise, generates streak artifacts, and can produce negative pixel values that are physically meaningless for a radioactivity distribution. It also cannot properly model non-uniform attenuation, which is why FBP historically relied on approximate corrections such as the Chang method. ^{3, 4}

OSEM, by contrast, is inherently non-negative, produces lower noise at matched resolution, and — most importantly — lets you build attenuation, scatter, and resolution recovery directly into the reconstruction. That is what makes quantitative SPECT/CT and low-count, half-time, or half-dose protocols possible. ^{5, 6, 9} For the CT-based attenuation step that feeds this, see our SPECT/CT quality control guide, and for the attenuation-correction concept in emission tomography generally, our PET/CT attenuation correction overview.

Key Technical Principles

Filtered back projection and the ramp filter

For a projection $p_{\theta }(t)$ acquired at angle $\theta$, where $t$ is the position across the detector, the filtered-back-projection reconstruction of the image $f(x,y)$ is:

$$f(x,y)={\int }_0^{\pi }{\left[{\int }_{-\infty }^{\infty }{P}_{\theta }(\omega )|\omega |e^{i2\pi \omega t}d\omega \right]}_{t=x\cos \theta +y\sin \theta }d\theta$$

Here $P_{\theta }(\omega )$ is the one-dimensional Fourier transform of the projection, and $|\omega |$ is the ramp filter. The ramp compensates for the oversampling of low spatial frequencies inherent in back projection, but because its amplitude grows without bound at high frequency, it also amplifies statistical noise. ⁴

In practice the ramp is never used alone. It is multiplied by a smoothing apodization window — most commonly a Butterworth filter:

$$W(\omega )=\frac{1}{\sqrt{1+{\left(\omega /{\omega }_c\right)}^{2n}}}$$

with cutoff frequency ${\omega }_c$ and order $n$, or a Hanning window. The effective filter is then $H(\omega )=|\omega |W(\omega )$. Lowering the cutoff suppresses more noise at the cost of spatial resolution; raising it preserves resolution but lets noise through. Selecting this filter is one of the few reconstruction decisions a technologist still makes routinely in FBP workflows, and the appropriate value depends on the tracer, count level, and camera. ^{4, 5}

MLEM and OSEM: the iterative update

Iterative reconstruction models the mean detected counts in projection bin $i$ as a forward projection of the voxel activities ${\lambda }_j$:

$${\bar{y}}_i=\sum _j{a}_{ij}{\lambda }_j$$

where the system matrix element $a_{ij}$ is the probability that an emission from voxel $j$ is detected in bin $i$. This matrix is where attenuation, scatter, and collimator-detector response (resolution recovery) are encoded. The maximum-likelihood expectation-maximization (MLEM) algorithm, derived for Poisson-distributed emission data, updates the image multiplicatively: ¹

$${\lambda }_j^{(k+1)}=\frac{{\lambda }_j^{(k)}}{\sum _i{a}_{ij}}\sum _i{a}_{ij}\frac{y_i}{\sum _b{a}_{ib}{\lambda }_b^{(k)}}$$

Each iteration compares the measured counts $y_i$ to the current forward projection ${\sum }_b{a}_{ib}{\lambda }_b^{(k)}$ and scales each voxel accordingly. Because every term is non-negative, the image can never go negative — a structural advantage over FBP. ^{1, 3}

MLEM converges slowly, often needing dozens of iterations. OSEM accelerates it by dividing the projections into ordered subsets $S_n$ and performing one update per subset, cycling through all subsets to complete a single iteration: ²

$${\lambda }_j^{(k+1)}=\frac{{\lambda }_j^{(k)}}{\sum _{i\in S_n}{a}_{ij}}\sum _{i\in S_n}{a}_{ij}\frac{y_i}{\sum _b{a}_{ib}{\lambda }_b^{(k)}}$$

Hudson and Larkin showed this yields roughly an order-of-magnitude acceleration over MLEM while preserving the non-negativity and restoration quality. ² The practical convergence bookkeeping is captured by:

$$\text{effective (MLEM-equivalent) iterations}=\text{iterations}\times \text{subsets}$$

A worked example: subsets, iterations, and convergence

Consider a study acquired over 120 projection angles. Choosing $N=10$ subsets places 12 projections in each subset. Running 4 iterations performs:

$$4\text{ iterations}\times 10\text{ subsets}=40\text{ effective MLEM-equivalent iterations}$$

but only four full passes through the data. To reach comparable convergence, plain MLEM might need on the order of 40 iterations — hence the roughly tenfold speed-up. ²

Concrete timing and recovery figures come from a published Lu-177 SPECT/CT study: with 10 subsets the reconstruction converged in about 6 iterations at roughly 20 seconds per iteration, so a full reconstruction ran in about two to three minutes. In that same work, the recovered activity for the largest sphere rose from 66% with attenuation-corrected OSEM, to 79% when resolution recovery was added, to 88% with a Monte-Carlo–based system model — a direct demonstration of how modeling more physics in $a_{ij}$ improves quantitative accuracy. ^{7, 8}

FBP vs OSEM at a glance

Clinical Impact

Half-time and half-dose imaging

The single biggest clinical payoff of iterative reconstruction with resolution recovery (IRR) is the ability to trade recovered image quality for shorter scans or lower administered activity. Because IRR restores resolution and controls noise better than FBP, studies have shown that myocardial perfusion SPECT can be acquired at substantially reduced time or dose while maintaining diagnostic image quality — reductions of at least 50%, and in some implementations up to roughly 89%, have been reported. ^{6, 9} That directly supports the dose-optimization goals covered in our cardiac SPECT MPI quality control guide.

Quantitative SPECT and theranostic dosimetry

Iterative reconstruction is the enabling technology for quantitative SPECT/CT. By modeling CT-based attenuation correction, scatter correction, and resolution recovery, and calibrating the reconstruction to absolute activity concentration, modern systems can report becquerels per milliliter — the foundation of post-therapy dosimetry for radiopharmaceutical therapies such as Lu-177. The EANM practice guideline for quantitative SPECT/CT lays out the reconstruction, correction, and calibration requirements for this. ^{7, 10} FBP, which cannot properly incorporate these corrections, is unsuitable for quantitative work.

Reading artifacts correctly

Understanding the reconstruction still matters at the reading station. A reader who knows FBP will recognize streak artifacts radiating from a hot bladder or injection site, and will not mistake ramp-filter negatives for true photopenia. A reader who knows OSEM will recognize that an under-converged reconstruction can blur small lesions, while an over-iterated one can look noisy and "grainy." The IAEA SPECT/CT artifact atlas catalogs many of these reconstruction-related appearances. ¹³

Practical Optimization Tips

1. Treat subsets × iterations as one convergence knob

Do not tune subsets and iterations independently as if they were unrelated. Their product sets convergence. When a vendor changes the default subset count, the number of iterations must change to keep effective iterations — and therefore image appearance — consistent. ²

2. Standardize reconstruction across cameras

For multi-camera departments and for any quantitative or serial-imaging program, harmonize reconstruction settings (algorithm, subsets, iterations, corrections, and post-filter) so that a patient's follow-up study is comparable to the baseline. Inconsistent reconstruction is a common, avoidable source of apparent change.

3. Match corrections to the clinical task

Attenuation correction, scatter correction, and resolution recovery each change quantitative values and image appearance. Turn them on deliberately and document them. Resolution recovery in particular improves small-structure recovery but can introduce edge artifacts (Gibbs-type overshoot) if pushed too far. ^{7, 8}

4. Do not over-iterate

More iterations increase noise as the solution approaches the noisy maximum-likelihood estimate. Use the vendor-validated iteration/subset combination for each tracer and task, and control residual noise with an appropriate post-filter rather than by simply reducing iterations, which also sacrifices resolution.

5. Verify the system with standardized QC

Reconstruction performance should be checked with standardized measurements — reconstructed spatial resolution, uniformity, and, for quantitative programs, a calibration and recovery-coefficient assessment. NEMA NU 1-2023 defines the standardized performance metrics, and the IAEA quality-assurance guidance describes the acceptance and routine QC procedures. ^{11, 12}

Regulatory Considerations

Reconstruction sits inside the broader quality-assurance and accreditation framework for nuclear medicine, even though the algorithm itself is not "regulated" the way dose limits are. What matters for compliance is that the imaging system's performance is validated, documented, and reproducible.

Relevant standards and guidance:

• NEMA NU 1-2023, Performance Measurements of Gamma Cameras, is the current standardized performance standard, superseding NU 1-2018. It defines how reconstructed spatial resolution, system sensitivity, and count-rate performance are measured, giving physicists a common language for specifying and verifying SPECT systems. ¹¹
• IAEA Human Health Series No. 6, Quality Assurance for SPECT Systems (2009), provides acceptance, reference, and routine QC test procedures — uniformity, center-of-rotation, and resolution — that underpin reliable reconstruction. ¹² The related center-of-rotation check is covered in our SPECT center-of-rotation QC guide.
• IAEA Human Health Series No. 36, SPECT/CT Atlas of Quality Control and Image Artefacts (2019), catalogs reconstruction- and correction-related artifacts for visual QC training. ¹³
• The EANM practice guideline for quantitative SPECT/CT (2023) defines the reconstruction and correction requirements for absolute quantitation and cross-system harmonization. ¹⁰

Accreditation bodies expect a documented QC program and physicist involvement in system performance evaluation. DRPS provides this through its PET/CT and nuclear medicine physics testing service and supports the underlying camera QC covered in our gamma-camera NEMA NU-1 performance testing guide.

Frequently Asked Questions (FAQs)

What is the difference between FBP and OSEM in SPECT?

Filtered back projection (FBP) is an analytic, single-pass algorithm that filters each projection with a ramp filter and smears it back across the image; it is fast but amplifies noise, can produce streak artifacts and negative pixel values, and cannot model imaging physics. OSEM (ordered-subsets expectation maximization) is an iterative algorithm that repeatedly compares a forward model of the data to the measured projections and updates the image, enforcing non-negativity and allowing attenuation, scatter, and resolution recovery to be modeled directly.

Why has OSEM largely replaced FBP for clinical SPECT?

OSEM produces lower noise at matched resolution, guarantees non-negative pixel values, avoids the streak artifacts of FBP, and — most importantly — lets the reconstruction model CT-based attenuation correction, scatter correction, and collimator-detector resolution recovery inside the system matrix. This supports quantitative SPECT and enables half-time or half-dose imaging protocols. FBP is now used mostly as a fast comparator or in legacy workflows.

What are subsets and iterations in OSEM?

OSEM divides the projection data into groups called subsets and updates the image once per subset, cycling through all subsets to complete one iteration. Using more subsets accelerates convergence roughly in proportion to the number of subsets. The practical rule is that the product of iterations and subsets — the effective or MLEM-equivalent iterations — describes how far the reconstruction has converged.

Does running more iterations always give a better image?

No. More iterations recover resolution and contrast but also amplify noise, because OSEM converges toward the noisy maximum-likelihood solution. Clinical protocols stop at a limited number of iterations, or apply a post-reconstruction filter or regularization, to balance resolution against noise. The optimal setting depends on the radiotracer, count level, camera, and clinical task.

Can SPECT be quantitative?

Yes. With iterative reconstruction that models CT-based attenuation correction, scatter correction, resolution recovery, and a proper calibration to becquerels per milliliter, modern SPECT/CT can produce absolute activity concentrations. This quantitative capability is central to theranostic dosimetry, for example in Lu-177 therapy, and is the subject of the EANM practice guideline for quantitative SPECT/CT.

Why does FBP produce streak artifacts and negative pixel values?

FBP is a linear, unconstrained analytic inversion. The ramp filter boosts high spatial frequencies, so statistical noise and incomplete angular sampling produce streaks radiating from high-count structures. Because the algorithm has no non-negativity constraint, the filtered back projection can drive pixels below zero around hot regions, which is physically impossible for a radioactivity distribution.

What filter is used with FBP, and how is it chosen?

FBP uses a ramp filter to compensate for the blurring inherent in back projection, but the ramp alone amplifies noise, so it is combined with a smoothing apodization window such as a Butterworth or Hanning filter. The cutoff frequency and, for a Butterworth filter, the order control the trade-off between spatial resolution and noise. Lower cutoffs give smoother images with suppressed noise at the cost of resolution.

Key Takeaways

• Two paradigms. FBP inverts the projection data analytically; OSEM estimates the image iteratively as a statistical best fit to the data.
• OSEM models physics. Attenuation, scatter, and collimator-detector resolution recovery live in the system matrix — the reason iterative methods enable quantitation and dose reduction.
• FBP explains the artifacts. Ramp-filter noise amplification, streaks, and negative pixel values are FBP behaviors every reader should recognize.
• Subsets × iterations is one knob. Their product sets convergence; change one and you must adjust the other to keep image appearance constant.
• More iterations is not always better. Resolution recovers but noise grows; stop at a validated setting and control noise with a post-filter.
• Standards anchor the program. NEMA NU 1-2023, IAEA QA guidance, and the EANM quantitative guideline define how to measure and verify reconstruction performance.

Conclusion

The move from FBP to iterative OSEM was not just a software update — it changed what SPECT can do. By moving reconstruction from a blind analytic inversion to a physics-aware statistical estimation, OSEM made attenuation correction, scatter correction, resolution recovery, and absolute quantitation routine, and opened the door to half-time and half-dose protocols.

Yet FBP remains the conceptual foundation. Its ramp filter, its noise behavior, and its artifacts explain why iterative reconstruction is built the way it is, and they still appear in day-to-day nuclear medicine. A physicist or technologist who understands both algorithms — and, critically, how the subsets-and-iterations settings translate into convergence and noise — is equipped to optimize protocols, standardize across cameras, and defend the quality and quantitative accuracy of the department's SPECT program.

How DRPS Can Help

Diagnostic Radiation Physics Services helps nuclear medicine departments get the most from their reconstruction pipeline. That can include PET/CT and nuclear medicine physics testing, reconstruction protocol review and harmonization across cameras, quantitative SPECT/CT calibration support, gamma-camera and SPECT QC program development, artifact troubleshooting, and medical physics consulting aligned with NEMA, IAEA, and accreditation expectations.

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

The goal is reconstruction that is consistent, artifact-aware, and — where it matters — genuinely quantitative.

Related Resources

• SPECT/CT quality control
• SPECT center-of-rotation QC
• Gamma-camera NEMA NU-1 performance testing
• Cardiac SPECT MPI quality control
• PET/CT attenuation correction
• PET/CT and nuclear medicine physics
• Medical physics consulting

References

1. Shepp LA, Vardi Y. Maximum Likelihood Reconstruction for Emission Tomography. IEEE Transactions on Medical Imaging. 1982;1(2):113-122. doi:10.1109/TMI.1982.4307558. PubMed
2. Hudson HM, Larkin RS. Accelerated Image Reconstruction Using Ordered Subsets of Projection Data. IEEE Transactions on Medical Imaging. 1994;13(4):601-609. doi:10.1109/42.363108. PubMed
3. O'Sullivan F, Pawitan Y, Haynor D. Reducing negativity artifacts in emission tomography: post-processing filtered backprojection solutions. IEEE Transactions on Medical Imaging. 1993;12(4):653-663. doi:10.1109/42.251115. PubMed
4. Bruyant PP. Analytic and iterative reconstruction algorithms in SPECT. Journal of Nuclear Medicine. 2002;43(10):1343-1358. PubMed
5. Bouchareb Y, AlSaadi A, Zabah J, et al. Technological Advances in SPECT and SPECT/CT Imaging. Diagnostics. 2024;14(13):1431. doi:10.3390/diagnostics14131431. PubMed
6. Brambilla M, Lecchi M, Matheoud R, et al. Comparative analysis of iterative reconstruction algorithms with resolution recovery and new solid state cameras dedicated to myocardial perfusion imaging. Physica Medica. 2017;41:109-116. doi:10.1016/j.ejmp.2017.03.008. PubMed
7. Rydén T, Heydorn Lagerlöf J, Hemmingsson J, et al. Fast GPU-based Monte Carlo code for SPECT/CT reconstructions generates improved 177Lu images. EJNMMI Physics. 2018;5(1):1. doi:10.1186/s40658-017-0201-8. PubMed
8. Lazaro D, El Bitar Z, Breton V, Hill D, Buvat I. Fully 3D Monte Carlo reconstruction in SPECT: a feasibility study. Physics in Medicine and Biology. 2005;50(16):3739-3754. doi:10.1088/0031-9155/50/16/006. PubMed
9. Lecchi M, Malaspina S, Scabbio C, Gaudieri V, Del Sole A. Myocardial perfusion scintigraphy dosimetry: optimal use of SPECT and SPECT/CT technologies in stress-first imaging protocol. Clinical and Translational Imaging. 2016;4(6):491-498. doi:10.1007/s40336-016-0212-9. PubMed
10. Dickson JC, Armstrong IS, Gabiña PM, et al. EANM practice guideline for quantitative SPECT-CT. European Journal of Nuclear Medicine and Molecular Imaging. 2023;50(4):980-995. doi:10.1007/s00259-022-06028-9. PubMed
11. National Electrical Manufacturers Association. NEMA Standards Publication NU 1-2023: Performance Measurements of Gamma Cameras. Rosslyn, VA: NEMA; 2023. nema.org
12. International Atomic Energy Agency. Quality Assurance for SPECT Systems. IAEA Human Health Series No. 6. Vienna: IAEA; 2009. iaea.org
13. International Atomic Energy Agency. SPECT/CT Atlas of Quality Control and Image Artefacts. IAEA Human Health Series No. 36. Vienna: IAEA; 2019. iaea.org