PET/MR Attenuation Correction: The Bone Problem

In PET/MR, attenuation correction is the hardest quantitative problem, because the MR signal does not correspond to how 511 keV photons are attenuated — and cortical bone and lung are effectively invisible to standard MR sequences. Where PET/CT hands you an attenuation map almost for free, PET/MR forces you to estimate one, and every estimation method carries a characteristic quantitative bias.¹²

Simultaneous PET/MR combines outstanding soft-tissue contrast, functional MR, and molecular PET without added ionizing dose from a CT. But that strength creates the attenuation-correction challenge: MRI measures proton density and relaxation, not photon attenuation. Understanding how modern scanners bridge that gap — and where they still fall short — is essential for anyone interpreting or quantifying PET/MR studies.¹³

This guide explains the physics of attenuation correction, why MR-based attenuation correction (MRAC) is difficult, the SUV bias it introduces, and how zero-echo-time (ZTE), ultrashort-echo-time (UTE), and deep-learning pseudo-CT methods reduce that bias. DRPS supports PET quantification and performance testing through PET/CT and nuclear medicine physics and MRI physics testing.

Introduction

Attenuation correction compensates for annihilation photons that are absorbed or scattered before reaching the PET detectors; without it, PET images are neither quantitatively accurate nor artifact-free. In whole-body imaging, more than half of the 511 keV photon pairs can be attenuated before detection, so the correction is not a refinement — it is a prerequisite for a usable image.¹

In PET/CT, the CT scan provides a spatial map of X-ray attenuation that converts readily to 511 keV linear attenuation coefficients, a workflow we describe in CT-Based Attenuation Correction in PET/CT. PET/MR has no such luxury. The MR image must be transformed into an attenuation map by classifying tissues and assigning each a coefficient — and the tissues that matter most for accuracy, cortical bone and lung, are precisely the ones MR struggles to see.¹²

The result is a decade-plus of methodological evolution: from simple four-class segmentation, to model- and atlas-based bone insertion, to emission-based estimation, and now to deep-learning pseudo-CT. Each step has chipped away at the quantitative error, and a physicist supporting a PET/MR program needs to know which method a scanner uses and what bias to expect.

Topic Explanation

The physics: what attenuation correction actually corrects

A 511 keV photon traveling through tissue is attenuated according to the tissue's linear attenuation coefficient $μ$ . For a coincidence line of response through the body, the attenuation correction factor (ACF) is:

ACF = exp (\int_{L} μ (x) d l)

The measured coincidences along that line are multiplied by the ACF to recover the true activity. Because the integral is an exponential, small errors in $μ$ — especially over long path lengths — translate into meaningful activity errors. This is why an incorrect attenuation map does not simply add noise; it biases the reconstructed standardized uptake value (SUV).¹

Why MR cannot directly measure attenuation

CT numbers correlate closely with 511 keV attenuation because both are governed by electron density and photon interactions, so a bilinear CT-to- $μ$ conversion works well. MRI is fundamentally different: it measures the proton (hydrogen) density and the T1/T2 relaxation of tissue, which bear no direct relationship to how gamma photons attenuate. A dense cortical bone and an air-filled sinus both produce almost no MR signal, yet their attenuation coefficients could hardly be more different.¹²

Standard whole-body MRAC uses a two-point Dixon sequence that separates fat and water and classifies each voxel into a small number of tissue classes — commonly background (air), lung, fat, and soft tissue. Bone is not a class; it is assigned the attenuation of soft tissue. That single simplification is the origin of most PET/MR quantitative bias.¹²

For context on how attenuation and quantification interact more broadly, see PET SUV Quantification and PET/CT NEMA NU-2 Performance Testing.

Key Technical Principles

A worked example: why bone matters

Consider a coincidence line passing through 20 cm of soft tissue. Water at 511 keV has a linear attenuation coefficient of approximately $μ_{water} \approx 0.096 cm^{- 1}$ .¹⁰ The attenuation correction factor is:

ACF = exp (0.096 \times 20) = exp (1.92) \approx 6.8

so the true coincidence rate is nearly seven times the measured rate along that line. Now replace 4 cm of that path with cortical bone, whose 511 keV coefficient is roughly $0.17 cm^{- 1}$ . If MRAC mislabels that bone as soft tissue (0.096 cm⁻¹), the path integral is underestimated by about $(0.17 - 0.096) \times 4 \approx 0.30$ , so the ACF is too small by a factor of $e^{0.30} \approx 1.35$ . Activity in and near the bone is therefore underestimated — the mechanism behind the observed SUV bias.¹²

The methods and their biases

The table below compares the principal MRAC approaches, their handling of bone, and representative published SUV/quantification errors relative to CT-based attenuation correction. The percentages are drawn from specific studies and reflect their patient cohorts and body regions; they are illustrative, not universal constants.

MRAC method	Bone handling	Representative reported error vs CT-AC	Notes
4-class Dixon segmentation	Bone = soft tissue	~8% ± 3% SUV change in bone lesions; up to ~13% in pelvic bone ¹	Vendor default; robust and fast
Dixon + model/atlas bone	Bone added from a bone model	Bone-tissue underestimation reduced from −25.5% to −4.9%; bone lesions −7.4% → −2.9% ²	Improves bone; atlas mismatch risk
Emission-based (MLAA/TOF)	Estimated from PET data	Bone SUV error −18.4% (standard MRAC) → −10.2% (MLAA-GMM) ³	Leverages time-of-flight; cross-talk risk
Hybrid ZTE/Dixon pseudo-CT	Bone from ZTE, continuous $μ$	Bone-lesion SUVmax RMSE 11.0% → 3.3%; soft tissue 7.8% → 3.9% ⁴	Captures cortical bone signal
Deep-learning pseudo-CT (ZeDD)	Learned MR→CT mapping	Bone-lesion RMSE 10.2% → 2.7%; soft tissue 6.2% → 4.1% ⁵	State of the art; training-data dependent
Deep learning + air-pocket seg.	Learned map + gas handling	Whole-pelvis error 5.1% (model) → 2.6% (DL) ⁶	Handles bowel gas variability

Two themes stand out. First, standard Dixon methods systematically underestimate uptake in and around bone, historically on the order of 10–25%.²³ Second, adding bone information — whether by model, ZTE/UTE, or deep learning — consistently shrinks that error toward single digits, with deep-learning pseudo-CT now approaching CT-AC accuracy in validation studies.⁴⁵⁷

ZTE, UTE, and pseudo-CT

Zero-echo-time (ZTE) and ultrashort-echo-time (UTE) sequences sample the MR signal so quickly that they capture the fleeting signal from cortical bone before it decays. That signal lets bone be segmented and assigned a continuous, density-dependent attenuation coefficient rather than a single soft-tissue value. Hybrid ZTE/Dixon methods convert ZTE intensity to Hounsfield units and then to linear attenuation coefficients, replacing the bone voxels of a Dixon-based pseudo-CT.⁴

Deep-learning approaches go further, training convolutional neural networks to synthesize a full pseudo-CT directly from Dixon (and optionally ZTE) inputs. The ZeDD-CT method reduced PET quantification error in bone lesions by roughly a factor of four relative to Dixon-only correction, and augmented deep-learning models have shown similar gains in prostate PSMA PET/MR.⁵⁷ A practical caveat: deep-learning maps are only as good as their training data, and out-of-distribution anatomy, implants, or unusual gas patterns can still produce errors.⁶

Clinical Impact

The clinical consequence of attenuation-correction bias is misquantified uptake, which matters most where PET/MR is used for staging, treatment response, and quantitative follow-up. A systematic 10–25% underestimation in bone directly affects the evaluation of skeletal metastases in prostate, breast, and other cancers — a growing application for PSMA and FDG PET/MR.²⁵⁷

Bias also matters for longitudinal imaging. If a baseline scan and a follow-up scan use different attenuation-correction methods or software versions, an apparent change in SUV may reflect the method, not the disease. Harmonizing attenuation correction across time points is therefore part of trustworthy quantitative PET/MR, a concern closely related to the SUV-harmonization principles discussed in EARL PET SUV Harmonization.

Neurologic PET/MR adds its own wrinkle: the skull is a large bony structure surrounding the target organ, so ignoring bone can bias cortical uptake in amyloid and tau imaging. This is one reason dedicated bone-aware attenuation correction is emphasized for brain PET/MR — a theme relevant to quantitative brain studies such as those covered in Amyloid and Tau Brain PET.³

Practical Optimization Tips

Know your scanner's method

The single most useful step is knowing exactly which MRAC method the installed scanner uses and its software version — 4-class Dixon, Dixon plus model bone, or a ZTE/deep-learning pseudo-CT. Every method has a documented bias signature, and interpretation should account for it.¹²

Watch the classic failure modes

Truncation. The MR field of view is often narrower than the PET field of view, so arms and body edges can be truncated from the attenuation map, biasing peripheral uptake. Many scanners apply a truncation-completion algorithm; confirm it is active.¹
Metal implants. Dental work, hip prostheses, and ports create MR signal voids that can be mislabeled as air, producing large local attenuation errors and artifacts.¹
Air pockets. Bowel gas moves between the MR and PET acquisitions; misclassifying gas as tissue (or vice versa) biases pelvic and abdominal uptake. Air-pocket-aware methods help.⁶
Coils and hardware. Fixed hardware attenuation is handled by templates, but incorrectly positioned or undocumented accessories can introduce error.

Support quantitative QC

Confirm attenuation and scatter correction are validated as part of routine PET performance QC.
Keep the attenuation-correction method fixed for longitudinal patients whenever possible.
Document software/method changes so quantitative trends remain interpretable.
Include phantom-based quantitative checks consistent with NEMA NU 2 methodology.⁸

Avoid common pitfalls

Treating PET/MR SUV as identical to PET/CT SUV. Method-dependent bias, especially near bone, means the two are not automatically interchangeable.²
Ignoring the map. Always review the attenuation map for truncation, metal voids, and gas misclassification before trusting a quantitative value.
Mixing methods over time. Longitudinal comparisons require consistent attenuation correction.

Regulatory Considerations

PET/MR sits at the intersection of radioactive-material regulation for the PET tracer and MR-safety governance for the magnet, with quantitative performance judged against nuclear medicine physics standards. The radiopharmaceutical — most commonly F-18 FDG, but increasingly Ga-68 and other tracers — is byproduct material regulated under NRC or Agreement State rules.⁹

Key frameworks:

NEMA NU 2-2024, Performance Measurements of Positron Emission Tomographs, the current standard (superseding NU 2-2018), which defines how PET performance — including corrections that underpin quantification — is measured.⁸
ACR–AAPM technical standards for PET performance monitoring, which frame the medical physicist's role in verifying quantitative accuracy and image quality.⁹
10 CFR Part 35, Medical Use of Byproduct Material, governing the PET radiopharmaceutical, with occupational and public dose limits under 10 CFR Part 20.¹¹
MR-safety governance for the magnet, static field, gradients, and RF, which operates independently of the radioactive-material program; for a facility-level view, see MRI Safety Program.

Because PET/MR combines two regulatory worlds, a facility should verify both its radioactive-material license conditions and its MR-safety program, and confirm which authority — NRC or Agreement State — issues the materials license. DRPS coordinates PET quantification support with PET/CT and nuclear medicine physics, MRI physics testing, and medical physics consulting.

Frequently Asked Questions (FAQs)

What is attenuation correction in PET/MR?

Attenuation correction compensates for the 511 keV annihilation photons absorbed or scattered before they reach the PET detectors. In PET/CT the CT image directly provides a map of attenuation coefficients, but in PET/MR the MR signal does not correspond to photon attenuation, so an attenuation map must be estimated from MR images using segmentation, atlas, or deep-learning methods.

Why can't MRI directly measure attenuation like CT?

CT measures X-ray attenuation, which is closely related to 511 keV attenuation, so CT numbers convert readily to linear attenuation coefficients. MRI measures proton density and relaxation properties of tissue, which are unrelated to photon attenuation. Cortical bone and lung produce almost no MR signal with conventional sequences, so they cannot be distinguished or assigned correct attenuation values without special techniques.

How large is the SUV error from ignoring bone in PET/MR?

Standard 4-class Dixon attenuation correction assigns bone the attenuation of soft tissue, which underestimates uptake in and near bone. Published studies report bone-tissue underestimations on the order of roughly 10 to 25 percent, with smaller errors in soft-tissue lesions, compared with CT-based attenuation correction. The exact bias depends on the lesion location, method, and body region.

What is a pseudo-CT in PET/MR attenuation correction?

A pseudo-CT (or synthetic CT) is a CT-like attenuation map estimated from MR images. It can be generated by atlas registration, by adding bone from zero-echo-time or ultrashort-echo-time sequences, or by deep-learning models trained to map MR images to CT Hounsfield units, which are then converted to 511 keV linear attenuation coefficients.

Do ZTE and UTE sequences improve PET/MR quantification?

Yes. Zero-echo-time (ZTE) and ultrashort-echo-time (UTE) sequences capture signal from cortical bone that standard Dixon sequences miss, allowing bone to be included in the attenuation map. Hybrid ZTE/Dixon and deep-learning pseudo-CT methods have been shown to substantially reduce SUV error in bone and nearby soft-tissue lesions relative to Dixon-only attenuation correction.

Is PET/MR attenuation correction a physics QC concern?

Absolutely. Attenuation correction directly determines quantitative accuracy, so the medical physicist should understand which method the scanner uses, its known biases, and how truncation, metal, and air pockets affect the attenuation map. Quantitative performance is verified within a broader QC framework aligned with NEMA NU 2 performance testing.

Key Takeaways

MR does not measure attenuation. Unlike CT, MR signal is unrelated to 511 keV photon attenuation, so PET/MR must estimate an attenuation map.¹²
Bone is the core problem. Standard 4-class Dixon MRAC labels bone as soft tissue, systematically underestimating uptake in and near bone by roughly 10–25% in reported studies.²³
Adding bone shrinks the error. Model/atlas bone, emission-based methods, ZTE/UTE, and deep-learning pseudo-CT all reduce bone bias toward single digits.²⁴⁵
Deep-learning pseudo-CT is state of the art but depends on training data and can fail on out-of-distribution anatomy, implants, or gas.⁵⁶⁷
Method awareness is essential. Know your scanner's MRAC method and version, watch for truncation, metal, and air pockets, and keep the method consistent for longitudinal patients.¹
Quantitative QC matters. Verify corrections within a NEMA NU 2-aligned performance framework.⁸

Conclusion

Attenuation correction is where PET/MR's elegance meets its hardest physics. Because MR signal carries no information about photon attenuation, the attenuation map must be inferred — and the tissues that matter most, bone and lung, are the ones MR sees least. Standard Dixon methods pay for that with a systematic underestimation of uptake in and around bone.¹²

The field has responded impressively. Model- and atlas-based bone insertion, emission-based estimation, ZTE/UTE bone imaging, and deep-learning pseudo-CT have each narrowed the gap to CT-based accuracy, with the newest deep-learning methods approaching it in validation studies.⁴⁵⁷ For the medical physicist and the interpreting physician, the takeaway is practical: know the method, respect its biases, review the map, and keep quantification consistent. Done well, PET/MR delivers molecular imaging with superb soft-tissue context and no CT dose — provided the attenuation correction is understood, not assumed.

How DRPS Can Help

Diagnostic Radiation Physics Services supports PET and PET/MR programs with quantitative performance evaluation, PET/CT and nuclear medicine physics, MRI physics testing, SUV-quantification review, and medical physics consulting provided by board-certified medical physicists.

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

Trustworthy PET/MR quantification is not automatic. It comes from understanding the attenuation-correction method behind every SUV — and from a QC program that keeps that number honest over time.

Related Resources

References

Martinez-Möller A, Souvatzoglou M, Delso G, et al. Tissue classification as a potential approach for attenuation correction in whole-body PET/MRI: evaluation with PET/CT data. Journal of Nuclear Medicine. 2009;50(4):520-526. doi:10.2967/jnumed.108.054726. doi.org
Paulus DH, Quick HH, Geppert C, et al. Whole-Body PET/MR Imaging: Quantitative Evaluation of a Novel Model-Based MR Attenuation Correction Method Including Bone. Journal of Nuclear Medicine. 2015;56(7):1061-1066. doi:10.2967/jnumed.115.156000. doi.org
Mehranian A, Zaidi H. Clinical Assessment of Emission- and Segmentation-Based MR-Guided Attenuation Correction in Whole-Body Time-of-Flight PET/MR Imaging. Journal of Nuclear Medicine. 2015;56(6):877-883. doi:10.2967/jnumed.115.154807. doi.org
Leynes AP, Yang J, Shanbhag DD, et al. Hybrid ZTE/Dixon MR-based attenuation correction for quantitative uptake estimation of pelvic lesions in PET/MRI. Medical Physics. 2017;44(3):902-913. doi:10.1002/mp.12122. doi.org
Leynes AP, Yang J, Wiesinger F, et al. Zero-Echo-Time and Dixon Deep Pseudo-CT (ZeDD CT): Direct Generation of Pseudo-CT Images for Pelvic PET/MRI Attenuation Correction Using Deep Convolutional Neural Networks with Multiparametric MRI. Journal of Nuclear Medicine. 2018;59(5):852-858. doi:10.2967/jnumed.117.198051. doi.org
Sari H, Reaungamornrat J, Catalano OA, et al. Evaluation of Deep Learning-Based Approaches to Segment Bowel Air Pockets and Generate Pelvic Attenuation Maps from CAIPIRINHA-Accelerated Dixon MR Images. Journal of Nuclear Medicine. 2022;63(3):468-475. doi:10.2967/jnumed.120.261032. doi.org
Pozaruk A, Pawar K, Li S, et al. Augmented deep learning model for improved quantitative accuracy of MR-based PET attenuation correction in PSMA PET-MRI prostate imaging. European Journal of Nuclear Medicine and Molecular Imaging. 2021;48(1):9-20. doi:10.1007/s00259-020-04816-9. doi.org
National Electrical Manufacturers Association. Performance Measurements of Positron Emission Tomographs (PET). NEMA Standards Publication NU 2-2024 (supersedes NU 2-2018). Rosslyn, VA: NEMA; 2024. nema.org
American College of Radiology and American Association of Physicists in Medicine. ACR–AAPM Technical Standard for Nuclear Medical Physics Performance Monitoring of PET Imaging Equipment (ACR–AAPM Practice Parameters and Technical Standards). aapm.org
National Institute of Standards and Technology. XCOM: Photon Cross Sections Database (mass attenuation coefficients for water and tissue at 511 keV). nist.gov
U.S. Nuclear Regulatory Commission. 10 CFR Part 35: Medical Use of Byproduct Material. ecfr.gov