SPECT Filters: Butterworth and Hanning

SPECT projection data are count-starved and noisy, and the ramp filter at the heart of filtered back projection amplifies exactly the high-frequency noise you do not want—so a low-pass apodizing filter, most often a Butterworth or a Hanning, is applied to tame it. The filter type, its cutoff frequency, and (for the Butterworth) its order together decide whether a study comes out crisp and quantitatively useful or either blurred or buried in noise. ^{1, 2}

Filter selection is one of the last image-quality decisions in nuclear medicine still made largely by hand. Reconstruction algorithms have advanced, scatter and attenuation correction are routine, but the smoothing filter and its cutoff are frequently left at a vendor default that may or may not suit the radiopharmaceutical, the count level, or the clinical question. A filter that is too aggressive erases small or low-contrast lesions; one that is too permissive floods the image with noise that mimics or hides disease. ^{1, 3}

This guide explains the frequency-domain physics behind SPECT filtering, defines the Butterworth and Hanning filters precisely, works through the Nyquist limit and a cutoff calculation, compares the common filters, and gives a defensible approach to choosing and standardizing parameters. DRPS supports facilities with PET/CT and nuclear medicine physics and medical physics consulting across Florida, Maryland, Virginia, Washington DC, California, Nevada, Pennsylvania, New York, New Jersey, and Delaware.

Introduction

A SPECT study acquires a set of planar projections as the gamma camera rotates around the patient. Reconstructing a cross-sectional image from those projections classically uses filtered back projection (FBP): each projection is filtered with a ramp filter and then smeared (back-projected) across the image matrix. The ramp filter is mathematically necessary—it corrects the $1/ r$ blurring that simple back projection would otherwise produce—but it has a dangerous property: its amplitude increases linearly with spatial frequency. ^{1, 2}

That is fine for signal, but SPECT data are dominated by Poisson (counting) noise, and noise has substantial high-frequency content. The ramp filter therefore amplifies noise most strongly exactly where there is least real information. The fix is to multiply the ramp by a low-pass apodizing window that rolls the response back down at high frequencies. The Butterworth and Hanning filters are the two most common choices, and the quality of a SPECT image hinges on how that window is shaped. ^{1, 2}

Topic Explanation

Why filtering happens in the frequency domain

Image information can be described as a spectrum of spatial frequencies. Low spatial frequencies describe large, slowly varying structures—organ outlines, broad uptake patterns—and carry most of the diagnostic signal. High spatial frequencies describe fine detail and sharp edges, but in a count-limited study they are dominated by noise. A low-pass filter keeps the low frequencies and attenuates the high ones, trading a controlled loss of fine resolution for a large reduction in noise. ¹

The art is the transition: cut too low and you blur real lesions; cut too high and noise survives. The filter's job is to place and shape that transition sensibly for the data in hand. ^{1, 3}

Where the filter is applied

In FBP, the apodizing window multiplies the ramp filter, so the effective filter is "ramp × Butterworth" (or "ramp × Hanning"). In iterative reconstruction such as OSEM, there is no ramp filter; instead a post-reconstruction low-pass filter (often a Gaussian or a Butterworth) smooths the final image, and the iteration and subset counts govern how much detail and noise are present before smoothing. Either way, the same frequency-domain logic applies. ^{1, 8, 9}

For the broader comparison of the reconstruction algorithms themselves, see SPECT reconstruction: FBP versus iterative OSEM.

Key Technical Principles

The Nyquist frequency sets the ceiling

Every digital image has a maximum representable spatial frequency—the Nyquist frequency—set by the pixel size:

f_{Nyquist} = \frac{1}{2 \cdot Δ x}

where $Δ x$ is the pixel dimension. Worked example: a 64 × 64 acquisition over a 40 cm field of view with no zoom gives a pixel size of $Δ x = 40/64 \approx 0.625$ cm, so:

f_{Nyquist} = \frac{1}{2 \times 0.625 cm} = 0.80 cycles/cm

Every filter cutoff must lie at or below this value. Because the Nyquist frequency depends on pixel size, cutoffs are frequently quoted as a fraction of Nyquist (for example, 0.5 Nyquist) so a protocol transfers sensibly across matrices and zoom factors—a cutoff fixed in cycles/cm means something different on a 64 matrix than on a 128 matrix. ^{1, 2}

The Butterworth filter

The Butterworth filter is defined in the spatial-frequency domain by two parameters—the cutoff (critical) frequency $f_{c}$ and the order $n$ :

B (f) = \frac{1}{1 + ( \frac{f}{f _{c}} ) ^{2 n}}

Two properties matter clinically. First, at $f = f_{c}$ the amplitude is always $1/ 2 \approx 0.707$ (a 3 dB point) regardless of the order—so the cutoff is a fixed reference point. Second, the order $n$ controls the steepness of the roll-off: a higher order produces a sharper transition from pass-band to stop-band, approaching an ideal "brick-wall" cut as $n$ grows. ^{1, 2}

Continuing the example above, a Butterworth with $f_{c} = 0.5$ cycles/cm and order $n = 5$ evaluated at $f = 0.6$ cycles/cm gives $B = 1/ 1 + (0.6/0.5)^{10} = 1/ 1 + 1. 2^{10} \approx 1/ 1 + 6.19 \approx 0.373$ —so a frequency just above the cutoff is already attenuated to about 37% of its original amplitude, and higher frequencies are suppressed far more. ²

The two knobs give the Butterworth its flexibility: the cutoff sets where smoothing begins and the order sets how abruptly it takes hold. That is why it is the most widely used SPECT reconstruction filter and why controlled studies repeatedly find it preserves contrast and lesion detectability better than fixed-shape windows. ^{1, 2, 3, 4}

The Hanning filter

The Hanning filter is a simple cosine-tapered low-pass window with a single parameter, its cutoff frequency $f_{c}$ :

H (f) = ⎩ ⎨ ⎧ 0.5 + 0.5 cos (\frac{π f}{f _{c}}), 0, 0 \leq f \leq f_{c} f > f_{c}

It reaches zero exactly at $f_{c}$ and has a fixed roll-off shape—there is no order parameter to adjust. The Hanning filter is effective and predictable, but because its shape is not tunable it tends to be more aggressive near the cutoff than a well-chosen Butterworth, which can cost some contrast. ¹

Comparing the common SPECT filters

Filter	Parameters	Roll-off behavior	Typical use / notes
Butterworth	Cutoff frequency and order	Tunable steepness; 0.707 at cutoff	Most widely used; strong contrast and lesion detectability ^{1, 4}
Hanning	Cutoff frequency	Fixed cosine taper to zero at cutoff	Simple, predictable; can over-smooth near cutoff ¹
Hamming	Cutoff frequency	Cosine taper with small residual at cutoff	Similar to Hanning; compared directly with Butterworth in phantoms ²
Parzen	Cutoff frequency	Very smooth, strong noise suppression	Heavy smoothing for very low-count data ¹
Shepp-Logan	Cutoff frequency	Mild smoothing, close to ramp	Retains resolution but passes more noise ¹

The central trade-off

Across brain, cardiac, bone, and other applications, the literature converges on one message: there is no universally "best" filter, and the parameters that matter most are driven by counts. Low-count studies need a lower cutoff to control noise; high-count studies tolerate a higher cutoff that preserves resolution. The order has a smaller effect than cutoff. And care is especially needed where both hot and cold regions must be seen, because an aggressive cutoff degrades cold-lesion contrast differently than hot-lesion detectability. ^{1, 2, 3, 4}

Clinical Impact

Myocardial perfusion SPECT

Cardiac SPECT is the highest-volume SPECT application and the one where filtering has been studied most. Optimization work has produced explicit parameter guidance: lower cutoff frequencies are appropriate for lower-count (for example, stress-first or weight-based reduced-activity) acquisitions, while higher-count studies support higher cutoffs. One reconstruction-parameter study of Tc-99m sestamibi myocardial perfusion SPECT with OSEM used a post-reconstruction Butterworth with a cutoff and order tuned separately for stress and rest—illustrating that even the same patient's two acquisitions can warrant different smoothing when their count statistics differ. ^{3, 5, 9}

Brain SPECT

Brain perfusion SPECT demands good gray–white differentiation and symmetric, low-noise images for side-to-side comparison. Dedicated optimization of the Butterworth cutoff and order for brain SPECT showed that parameter choice materially changes apparent cortical uptake and image noise, and that a standardized filter improves reproducibility across patients and over time. ⁶

Quantitative SPECT/CT

As SPECT/CT moves toward absolute quantitation—standardized uptake values, dosimetry for radiopharmaceutical therapy—the filter interacts directly with measured activity concentrations. Over-smoothing lowers apparent peak uptake through partial-volume blurring; under-smoothing inflates noise-driven variability. A reproducible filter is part of a defensible quantitative pipeline. See quantitative SPECT/CT calibration for how calibration and reconstruction settings combine. ^{7, 10}

Practical Optimization Tips

1. Anchor the cutoff to counts, not habit

Tie the cutoff frequency to the study's count statistics. Reduced-activity or fast acquisitions need lower cutoffs; high-count studies can use higher cutoffs to keep resolution.

2. Express cutoffs as a fraction of Nyquist when protocols move between matrices

A cutoff fixed in cycles/cm shifts meaning if the matrix or zoom changes. Quoting it as a fraction of Nyquist keeps the intent intact.

3. Tune order second, cutoff first

Because cutoff dominates image quality, set it first. Use the order to fine-tune roll-off steepness; a moderate order is usually sufficient.

4. Standardize per study type and document it

Lock a filter and parameters for each study type (cardiac stress, cardiac rest, brain, bone, and so on) and record them. Standardization reduces interpretation variability and makes results comparable within and across departments. ¹

5. Do not assume the vendor default fits

Manufacturer defaults are starting points, not validated choices for your collimator, activity, and matrix. Verify against phantom data and clinical images. ¹

6. Treat OSEM smoothing deliberately

In OSEM, the iteration/subset product controls effective resolution and noise before the post-filter. Choose iterations/subsets and the post-reconstruction filter together, not in isolation. ^{8, 9}

Common pitfalls to avoid

Copying a cutoff across matrices in cycles/cm. The Nyquist ceiling moves with pixel size.
Chasing a "sharp" image with a high cutoff on a low-count study. You are amplifying noise, not recovering detail.
Leaving stress and rest on identical filters when their counts differ. Match smoothing to each acquisition. ⁹
Ignoring cold lesions. A cutoff optimized for hot-lesion detectability may degrade cold-region contrast. ²
Never revisiting defaults. Filters set once and forgotten drift out of alignment with changing protocols and hardware.

Regulatory Considerations

SPECT image processing is governed by professional standards and the facility's radioactive-material license rather than by a prescriptive federal filter rule—but the medical physicist is expected to document reconstruction and filter parameters as part of the quality program. The radiopharmaceuticals imaged are byproduct material regulated under NRC or Agreement State rules, and image-quality QC is addressed through professional guidance.

Key frameworks:

IAEA Nuclear Medicine Physics handbook. The IAEA's teaching handbook details the mathematics of FBP, the ramp filter, apodizing windows, the Nyquist limit, and iterative reconstruction—the canonical reference for the physics in this article. ¹
SNMMI guidance. Society of Nuclear Medicine and Molecular Imaging procedure standards and performance guidance address acquisition, processing, and QC expectations for SPECT studies, within which filter standardization sits.
10 CFR Part 35 — Medical Use of Byproduct Material. Governs authorized use of the radiopharmaceuticals and the radiation safety 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. A facility should confirm which authority licenses its program. ¹¹

Documenting reconstruction and filter parameters—and keeping them standardized—supports both image-quality QC and the reproducibility expected during accreditation and inspection. DRPS helps facilities build and verify these protocols through PET/CT and nuclear medicine physics and accreditation support.

Frequently Asked Questions (FAQs)

What does a reconstruction filter do in SPECT?

A reconstruction filter in SPECT suppresses the high-frequency statistical noise that the ramp filter amplifies during filtered back projection, while trying to preserve the spatial resolution and contrast of real structures. It is applied in the spatial-frequency domain as a low-pass (apodizing) window that keeps low frequencies, which carry most anatomical information, and rolls off high frequencies, where noise dominates.

What is the Butterworth filter cutoff frequency?

The cutoff (or critical) frequency of a Butterworth filter is the spatial frequency at which the filter's amplitude falls to 1 over the square root of 2, about 0.707, of its maximum—equivalent to a 3 dB reduction. Frequencies well below the cutoff pass nearly unchanged; frequencies well above it are strongly attenuated. Cutoff is usually expressed in cycles per centimeter or as a fraction of the Nyquist frequency.

How do cutoff frequency and filter order affect the image?

A lower cutoff frequency produces a smoother, less noisy but lower-resolution image, which suits low-count studies; a higher cutoff preserves resolution but passes more noise. The filter order controls how sharply the filter rolls off beyond the cutoff: a higher order gives a steeper transition. In practice, cutoff frequency has a larger effect on image quality than order.

What is the difference between the Butterworth and Hanning filters?

The Hanning filter is a simple cosine-shaped low-pass window defined by a single cutoff frequency, with a fixed roll-off shape and no order parameter. The Butterworth filter has two parameters—cutoff frequency and order—so its roll-off steepness can be tuned independently of the cutoff. The added flexibility is why Butterworth is the most widely used SPECT reconstruction filter, often outperforming fixed-shape windows for contrast and lesion detectability.

What is the Nyquist frequency and why does it matter for filtering?

The Nyquist frequency is the highest spatial frequency an image can represent without aliasing; it equals 1 divided by twice the pixel size. All reconstruction-filter cutoffs are bounded by it, and cutoffs are often quoted as a fraction of Nyquist so they transfer across matrices. Because the Nyquist frequency depends on pixel size, the same cutoff in cycles per centimeter corresponds to different fractions of Nyquist on different matrices and zoom settings.

Do iterative reconstructions like OSEM still need filters?

Yes. Ordered-subset expectation maximization (OSEM) does not use the ramp filter, so it does not need a reconstruction filter in the filtered-back-projection sense, but OSEM images are still noisy and are almost always smoothed with a post-reconstruction low-pass filter, commonly a Gaussian or a Butterworth. The number of iterations and subsets sets how much high-frequency detail and noise the image contains, and the post-filter controls the final noise texture.

Is there one correct filter for every SPECT study?

No. The optimal filter and parameters depend on the radiopharmaceutical, administered activity, counts acquired, matrix and zoom, collimator, and the clinical task. Low-count studies need lower cutoffs; high-count studies tolerate higher cutoffs. The practical goal is a documented, standardized set of filter parameters for each study type so results are consistent within a department and comparable over time.

Key Takeaways

The ramp filter amplifies noise; the apodizing filter tames it. FBP needs the ramp, but it boosts high-frequency noise, so a low-pass Butterworth or Hanning window is applied on top.
The Nyquist frequency is the ceiling. It equals $1/ (2Δ x)$ ; express cutoffs as a fraction of Nyquist so protocols survive matrix and zoom changes.
Cutoff first, order second. The cutoff frequency dominates image quality; the Butterworth order tunes roll-off steepness, with the amplitude fixed at 0.707 at the cutoff.
Butterworth's flexibility wins. Two parameters let it preserve contrast and lesion detectability better than fixed-shape windows in many studies.
Counts drive the choice. Low-count studies need lower cutoffs; high-count studies tolerate higher ones—even stress and rest can warrant different smoothing.
OSEM still needs smoothing. Choose iterations, subsets, and the post-reconstruction filter together.
Standardize and document. A fixed, recorded filter per study type is the foundation of reproducible, defensible SPECT.

Conclusion

Reconstruction filtering is where the statistics of radioactive decay meet the physics of image formation. SPECT counts are precious and noisy, the ramp filter makes the noise worse before back projection can make an image, and a well-chosen low-pass window is what turns raw projections into a readable, quantifiable study. The Butterworth filter earns its dominance through two tunable parameters—cutoff and order—while the Hanning and other fixed-shape windows remain useful, predictable alternatives.

The enduring lesson is that filtering is a deliberate, task-specific choice, not a default to inherit. Anchor the cutoff to counts, express it against the Nyquist limit, match smoothing to each acquisition, and standardize and document the result. Done well, filtering is invisible—the image simply looks right and measures consistently. Done carelessly, it is one of the quietest ways to lose a lesion or invent one.

How DRPS Can Help

Diagnostic Radiation Physics Services helps nuclear medicine and SPECT/CT facilities translate reconstruction physics into standardized, documented protocols. Our support includes reconstruction and filter-parameter review, phantom-based image-quality evaluation, quantitative SPECT/CT calibration, QC program development, and accreditation readiness through PET/CT and nuclear medicine physics, accreditation support, and medical physics consulting.

DRPS supports facilities across our service locations, including Florida, Maryland, Virginia, Washington DC, California, Nevada, New York, Pennsylvania, New Jersey, and Delaware. Contact us to review your SPECT reconstruction and filter protocols.

Related Resources

References

Lyra M, Ploussi A. Filtering in SPECT image reconstruction. Int J Biomed Imaging. 2011;2011:693795. doi:10.1155/2011/693795. PubMed
Sayed IS, Ismail SS. Comparison of low-pass filters for SPECT imaging. Int J Biomed Imaging. 2020;2020:9239753. doi:10.1155/2020/9239753. PubMed
Shibutani T, Onoguchi M, Yamada T, et al. Optimization of the filter parameters in 99mTc myocardial perfusion SPECT studies: the formulation of flowchart. Australas Phys Eng Sci Med. 2016;39(2):571-581. doi:10.1007/s13246-016-0433-2. PubMed
Lyra M, Ploussi A, Rouchota M, Synefia S. Filters in 2D and 3D cardiac SPECT image processing. Cardiol Res Pract. 2014;2014:963264. doi:10.1155/2014/963264. PubMed
Honda N, Machida K, Tsukada J, et al. Optimal preprocessing Butterworth-Wiener filter for Tl-201 myocardial SPECT. Eur J Nucl Med. 1987;13(8):404-407. doi:10.1007/BF00292492. PubMed
Minoshima S, Maruno H, Yui N, et al. Optimization of Butterworth filter for brain SPECT imaging. Ann Nucl Med. 1993;7(2):71-77. doi:10.1007/BF03164571. PubMed
Ekjeen T, Tocharoenchai C, Pusuwan P, et al. Optimization and evaluation of reconstruction-based compensation methods and reconstruction parameters for Tc-99m MIBI parathyroid SPECT. Phys Med. 2015;31(2):159-166. doi:10.1016/j.ejmp.2014.12.005. PubMed
Frey EC, Gilland KL, Tsui BMW. Application of task-based measures of image quality to optimization and evaluation of three-dimensional reconstruction-based compensation methods in myocardial perfusion SPECT. IEEE Trans Med Imaging. 2002;21(9):1040-1050. doi:10.1109/TMI.2002.804437. PubMed
Dheer P, Gupta P, Taywade SK, et al. Optimization of ordered subset expectation maximization parameters for image reconstruction in Tc-99m MIBI myocardial perfusion SPECT and comparison with corresponding filtered back projection-reconstructed images. Indian J Nucl Med. 2021;36(1):14-20. doi:10.4103/ijnm.IJNM_140_20. PubMed
Winz OH, Hellwig S, Mix M, et al. Image quality and data quantification in dopamine transporter SPECT: advantage of 3-dimensional OSEM reconstruction? Clin Nucl Med. 2012;37(9):866-871. doi:10.1097/RLU.0b013e318251e1b3. PubMed
International Atomic Energy Agency. Nuclear Medicine Physics: A Handbook for Teachers and Students. Vienna: IAEA; 2014. iaea.org
U.S. Nuclear Regulatory Commission. 10 CFR Part 35: Medical Use of Byproduct Material. ecfr.gov