MRI Parallel Imaging: g-Factor and SNR

Parallel imaging accelerates MRI by acquiring fewer phase-encoding lines and using the distinct spatial sensitivities of a receiver coil array to recover the missing information, but the speed is paid for in signal-to-noise ratio (SNR) through two separate terms: a square-root-of-R sampling penalty and a spatially varying geometry factor, the g-factor, that is always at least one. A defensible MRI protocol and quality-control (QC) program treats the acceleration factor as an SNR budget decision, not a free scan-time shortcut. ¹⁶

Every modern clinical MRI system ships with parallel imaging enabled, and technologists routinely dial an acceleration factor of two, three, or more without seeing the cost that the physics imposes. That cost is real, it is quantifiable, and it varies pixel by pixel across the image. This article explains where the SNR goes, how the g-factor is defined and measured, how SENSE and GRAPPA differ, and how a medical physicist evaluates parallel-imaging performance during acceptance testing and the annual survey. ¹²⁶

Introduction

Parallel imaging is a reconstruction strategy that undersamples k-space by a reduction factor R and reconstructs an unaliased image using the independent spatial information encoded in a phased-array receiver coil. Because MRI phase encoding is the slow axis of acquisition, skipping phase-encoding lines is the most direct way to shorten a scan, reduce motion sensitivity, shorten breath-holds, and shrink echo-train blurring or distortion. ¹⁶

The catch is that undersampling folds the image onto itself — aliasing — and the reconstruction has to unfold it. Two families of methods dominate clinical practice: SENSE, which unfolds in the image domain using coil sensitivity maps, and GRAPPA, which synthesizes the missing k-space lines using autocalibration data. ¹² Both work, both are limited by the same fundamental noise physics, and both are governed by the g-factor. Understanding the g-factor is the difference between choosing acceleration deliberately and choosing it by habit.

This guide walks through the source of the SNR penalty, the two reconstruction domains, a worked SNR calculation, the clinical consequences, practical protocol and QC tips, and the standards and accreditation context that a physics program uses to keep parallel imaging defensible. DRPS provides this evaluation as part of its MRI physics testing and accreditation support services.

Topic Explanation

What is parallel imaging?

Parallel imaging exploits the fact that a receiver coil element only "sees" the part of the patient near it, so the array of elements provides spatial information that partially substitutes for gradient phase encoding. Instead of acquiring every phase-encoding line needed to satisfy the Nyquist criterion for the full field of view, the sequence acquires every R-th line. This undersampling reduces the acquired field of view and causes the reconstructed image to alias, with signal from separated locations superimposed. ¹⁶

The reconstruction's job is to separate those superimposed signals. Because each coil element weights the folded pixels differently according to its sensitivity, a system of equations can be solved to recover the true, unfolded pixel values. When the coil sensitivities at the aliased locations are well separated, the unfolding is well conditioned and adds little noise. When they are similar — because the elements overlap too much or the acceleration is too aggressive for the array — the unfolding is ill conditioned and amplifies noise. That amplification is the g-factor. ¹⁵

Key terms used throughout this guide:

Reduction / acceleration factor (R) — how many times fewer phase-encoding lines are acquired; also written as the SENSE factor or GRAPPA factor. Scan time for the accelerated portion scales roughly as 1/R.
Coil sensitivity — the complex spatial weighting each array element applies to the signal; the raw material parallel imaging uses to unfold aliasing.
g-factor (geometry factor) — the spatially varying noise-amplification term introduced by the reconstruction; g ≥ 1 everywhere. ¹
g-factor map — an image of g across the field of view, showing where noise amplification is worst (typically toward the center, away from the coil elements). ³

Where the SNR actually comes from

Before adding parallel imaging, it helps to recall the baseline. For a conventional Cartesian acquisition, SNR scales with the voxel volume and the square root of the total sampling time:

SNR \propto V_{voxel} \cdot T_{acq}

Acquiring more signal averages, more phase-encoding lines, or more readout samples increases $T_{acq}$ and raises SNR as its square root. ⁶ Parallel imaging deliberately reduces the acquired data by a factor of R, cutting $T_{acq}$ to roughly $1/ R$ of its full value. That alone reduces SNR by $R$ — a fundamental, unavoidable consequence of collecting less data. The g-factor is then an additional loss stacked on top of the sampling loss. ¹⁶

Key Technical Principles

The g-factor and the SNR penalty

The central result of parallel imaging, established in the original SENSE formulation, is that the reconstructed SNR at each location is the full-sampled SNR reduced by both the sampling factor and the local geometry factor: ¹

SNR_{R} = \frac{SNR _{R = 1}}{g \cdot R}

Here $SNR_{R = 1}$ is the SNR of the fully sampled reference acquisition, $R$ is the acceleration factor, and $g \geq 1$ is the geometry factor at that pixel. Two facts follow immediately. First, because $g$ is at least one, the accelerated SNR is always worse than the naive $R$ expectation. Second, because $g$ varies across the image, the SNR penalty is not uniform: the center of a large body coil array, far from every element, typically suffers the most. ¹³⁵

The g-factor grows as R increases and as the unfolding becomes harder — that is, as the coil sensitivities at the aliased positions become less distinct. A well-designed array with many elements and good geometric spread keeps g near unity at low acceleration, but every array has an encoding capacity beyond which g rises steeply. Pushing R past that point produces disproportionate noise amplification and, eventually, residual aliasing. ¹⁵⁶

A comparison: SENSE versus GRAPPA

Parallel imaging is implemented in two domains. The practical differences matter for QC because the two methods fail differently and are calibrated differently.

Feature	SENSE (image domain)	GRAPPA (k-space domain)
Where unfolding happens	Reconstructed image; aliased pixels are separated using coil sensitivity maps	Raw k-space; missing lines are synthesized from acquired neighbors
Calibration data	Explicit coil sensitivity maps from a reference scan or prescan	Autocalibration signal (ACS) lines near the center of k-space
Output before combination	One unfolded image	Full k-space for each coil, then combined
Typical failure mode	Central residual aliasing if sensitivity maps are wrong or the field of view is set too small	Structured/incoherent artifacts and noise if ACS lines are too few
g-factor computation	Directly from the sensitivity encoding matrix ¹	From the reconstruction weights via a general formulation ⁴
Vendor names (examples)	SENSE, ASSET, mSENSE	GRAPPA, ARC, iPAT (GRAPPA mode)

Both approaches obey the same $g R$ SNR law; the difference is in how the coil information is estimated and where the arithmetic is done. ¹²⁴ A modern review is a useful single source that walks through both and their artifact behavior. ⁶

Worked SNR example

Consider a fully sampled acquisition with a baseline SNR of 100 (arbitrary units) at a voxel near the center of a torso array. Apply an in-plane acceleration of $R = 2$ , and suppose the g-factor map reads $g = 1.25$ at that voxel — a plausible mid-array value for a good array at modest acceleration.

SNR_{R = 2} = \frac{100}{1.25 \cdot 2} = \frac{100}{1.25 \times 1.414} = \frac{100}{1.768} \approx 56.6

The voxel retains about 57 percent of its original SNR — a 43 percent loss — even though only a factor of two acceleration was applied. Of that loss, the $2$ term accounts for a drop to about 71 percent, and the g-factor removes another chunk on top. Now push to $R = 3$ with a higher central g-factor of $g = 1.5$ :

SNR_{R = 3} = \frac{100}{1.5 \cdot 3} = \frac{100}{1.5 \times 1.732} = \frac{100}{2.598} \approx 38.5

Barely over a third of the baseline SNR survives. The table below makes the stacking explicit, using illustrative central g-factor values (actual g depends on the specific coil, anatomy, and slice):

Acceleration R	Sampling factor 1/√R	Example central g	Net retained SNR fraction 1/(g·√R)
1 (none)	1.00	1.00	100%
2	0.71	1.25	~57%
3	0.58	1.50	~38%
4	0.50	1.90	~26%

The g values here are illustrative, not pass/fail thresholds — the point is the shape of the curve: SNR falls faster than $R$ because g itself climbs with R. ¹⁵⁶ This is exactly why aggressive acceleration is reserved for high-SNR situations (high field strength, large voxels, many-element arrays, or SNR-rich contrasts) and avoided where SNR is already scarce.

Why a single ROI cannot measure accelerated SNR

Conventional SNR measurement takes the mean signal in a region of interest (ROI) over the standard deviation of a background ROI. That method assumes the noise is spatially stationary. Parallel-imaging reconstruction breaks that assumption: the g-factor makes noise vary from pixel to pixel, so a background ROI far from the signal no longer represents the noise where the signal is measured. ³

Two rigorous options exist. The NEMA image-subtraction method acquires two identical images and uses their difference to isolate noise, which remains valid when the noise field is stationary within the measured region. ¹⁰ For fully spatially resolved SNR and g-factor maps under parallel imaging, the pseudo multiple replica method synthesizes many noise realizations from the measured noise covariance and a single accelerated acquisition, producing pixel-wise SNR and g maps for both SENSE-type and GRAPPA-type reconstructions. ³ These are the tools a physicist reaches for when a site reports a parallel-imaging problem that a single-number SNR cannot explain.

Clinical Impact

The g-factor turns an abstract acceleration setting into a concrete, location-dependent image-quality decision. In body imaging, the center of the abdomen — precisely where the liver, pancreas, and retroperitoneum live — is where g is largest and SNR loss is worst. In neuro imaging, deep central structures pay more than the cortex. Choosing an acceleration factor without knowing the array's g-factor behavior risks quietly degrading the exact region the study is meant to evaluate. ¹³⁵

Parallel imaging also interacts with the rest of the protocol. Accelerating a diffusion or single-shot sequence reduces echo-train length, which reduces geometric distortion and blurring — a genuine image-quality gain that can offset the SNR loss for the right task. Accelerating a routine high-SNR T2 sequence may simply throw away SNR the study did not need to spend. The clinical value of acceleration is therefore task-specific: the same R can be excellent for one sequence and wasteful for another. ⁶

Finally, residual aliasing from over-acceleration or poor calibration is a different failure than noise. A faint central "unfolding" band in a SENSE image, or structured noise in a GRAPPA image, can mimic or obscure pathology. Recognizing these as parallel-imaging artifacts — and separating them from coil failures, motion, or hardware problems — is part of a competent artifact workup, and it connects directly to routine MRI image-artifact QC. ⁶

Practical Optimization Tips

Choose acceleration as an SNR budget

Start from the SNR you have, not the time you want. High-field, large-voxel, many-channel acquisitions have SNR to spend; low-SNR sequences do not. Set R against the SNR reserve for the specific contrast and anatomy.
Know your array's encoding capacity. Every coil has an R beyond which g climbs steeply. Ask the physicist for representative g-factor maps at the accelerations your protocols use, so the "cliff" is documented, not discovered clinically. ¹⁵
Prefer acceleration that buys physics, not just time. On echo-planar and single-shot sequences, acceleration shortens the echo train and reduces distortion and blurring — a real quality gain. On already-fast, high-SNR sequences, the same acceleration may only cost SNR. ⁶

Protect the calibration

Guard the reference/ACS data. SENSE sensitivity maps and GRAPPA autocalibration lines are the foundation of the unfolding. Motion between the reference and the accelerated scan, or too few ACS lines, degrades both SNR and artifact behavior. ²⁶
Keep the field of view honest. In SENSE, setting the field of view smaller than the object invites central residual aliasing. Confirm coverage matches the anatomy.
Coil selection and positioning matter. The g-factor is a property of the array geometry relative to the patient. Correct coil choice, seating, and centering keep the sensitivities well separated. This is the same coil discipline that underpins routine MRI SNR and RF-coil QC. ⁵

Verify, then trend

Measure SNR with a valid method. Use NEMA subtraction or a replica-based method, never a single background ROI, whenever parallel imaging is on. ³¹⁰
Baseline and trend coil performance. Acceptance testing establishes coil SNR and uniformity; the annual survey and routine QC trend them, so a failing element or a solder-joint problem shows up before it corrupts accelerated images. ⁸⁹¹¹
Add g-factor mapping for problem solving. When a site reports central noise or unfolding artifacts that ordinary QC cannot explain, pixel-wise g maps localize the cause. ³⁴

Regulatory Considerations

Parallel imaging is not separately "regulated," but the SNR and image-quality performance it affects sits squarely inside the MRI physics testing and accreditation framework. MRI is non-ionizing, so it falls outside the NRC's byproduct-material rules and, because it produces no ionizing radiation, outside state radiation-machine registration programs such as Florida's Chapter 64E-5. The binding quality requirements come instead from accreditation and from the equipment-evaluation standards that accreditation and hospital accreditors adopt. ¹¹

The relevant technical anchors are:

ACR MRI Quality Control Manual (2015 edition). Defines the roles of the MRI technologist, the QC program, and the qualified medical physicist or MR scientist, and specifies the routine SNR and image-quality tests that support ACR accreditation. ¹¹
ACR MRI Accreditation Program. Requires an annual MRI equipment performance evaluation by a qualified medical physicist or MR scientist and sets image-quality expectations that accelerated protocols must still meet. ¹¹
AAPM Report No. 100. The AAPM's acceptance-testing and QA procedures for MRI facilities, including phantom-based SNR, uniformity, and ghosting evaluation used at commissioning and annually. ⁸
NEMA MS 1-2008 (R2014). The standardized single-channel/volume-coil SNR measurement method, including the image-subtraction approach that remains valid when noise is spatially stationary. ¹⁰
NEMA MS 9-2008 (R2014, R2020). The characterization standard for phased-array coils — the array class on which all parallel imaging depends — covering array SNR and image uniformity. ⁹

Hospital accreditors that adopt an annual MRI equipment-performance evaluation (for example, through standards requiring a qualified physicist's assessment) rely on these same measurement methods. A parallel-imaging protocol that quietly falls below the SNR needed for the diagnostic task is an accreditation and quality problem even though no radiation rule is implicated. Tying acceleration choices to documented coil SNR and g-factor behavior is what makes the protocol defensible during an ACR accreditation physics review. ⁸⁹¹¹

Frequently Asked Questions (FAQs)

What is the g-factor in MRI parallel imaging?

The g-factor, or geometry factor, is a spatially varying number that describes how much the parallel-imaging reconstruction amplifies noise when it unfolds aliased signal. It is always greater than or equal to one, depends on the coil array geometry and the acceleration factor, and is displayed as a g-factor map rather than a single value. ¹³

How much SNR does parallel imaging cost?

Signal-to-noise ratio drops by the square root of the acceleration factor R from acquiring less data, and then by the local g-factor from the reconstruction. The retained SNR is the full-sampled SNR divided by g times the square root of R, so even a modest acceleration of R equal to 2 with a g-factor of 1.25 keeps only about 57 percent of the original SNR. ¹

What is the difference between SENSE and GRAPPA?

SENSE unfolds aliased pixels in the image domain using measured coil sensitivity maps, while GRAPPA fills in missing k-space lines using autocalibration data acquired near the center of k-space. Both are parallel-imaging methods governed by the same g-factor and square-root-of-R SNR penalty, but they fail differently when calibration is poor. ¹²⁶

Why can't you measure SNR with a single ROI after parallel imaging?

Parallel-imaging reconstruction makes the noise spatially varying, so a single background region no longer represents the noise where signal is measured. Valid SNR and g-factor evaluation uses methods such as NEMA image subtraction or the pseudo multiple replica technique that account for the spatially varying noise field. ³¹⁰

Does higher acceleration always mean worse image quality?

Higher acceleration reduces SNR and, above the array's encoding capacity, raises the g-factor sharply and can leave residual aliasing artifacts. The right acceleration balances scan time, motion, and geometric coverage against the SNR budget for the specific coil, field strength, and clinical task. ⁵⁶

How is parallel imaging performance checked in MRI QC?

Acceptance testing and annual physics surveys evaluate coil SNR and uniformity using NEMA MS 1 and MS 9 methods and AAPM Report No. 100 procedures, and the ACR MRI Quality Control Manual defines the routine SNR and image-quality checks that support accreditation. A physicist can add g-factor mapping when a parallel-imaging problem is suspected. ³⁸⁹¹¹

Who should evaluate parallel imaging on a clinical scanner?

A qualified or board-certified medical physicist performs MRI acceptance testing and the annual equipment evaluation, including coil SNR, uniformity, and artifact assessment. The physicist ties NEMA and AAPM measurement methods to ACR accreditation requirements and to the site's clinical protocols. ⁸¹¹

Key Takeaways

Parallel imaging trades SNR for speed through two terms. The $R$ sampling penalty comes from acquiring less data; the g-factor comes from the reconstruction unfolding aliased signal. ¹⁶
The g-factor is always at least one and varies across the image. It is worst where coil sensitivities overlap most — typically the center of the array — so central anatomy pays the biggest SNR price. ¹³⁵
SNR falls faster than $R$ . Because g climbs with acceleration, doubling R costs more than a factor of $2$ ; a worked R = 2 case with g = 1.25 retains only about 57 percent of baseline SNR. ¹
SENSE and GRAPPA obey the same law but calibrate differently. Sensitivity maps versus autocalibration lines change how each fails, not the underlying SNR penalty. ²⁴⁶
A single ROI cannot measure accelerated SNR. Use NEMA subtraction or a replica-based method, and add g-factor maps for problem solving. ³¹⁰
It is an accreditation and QC matter, not a radiation one. NEMA MS 1/MS 9, AAPM Report 100, and the ACR MRI QC Manual anchor the measurements that keep accelerated protocols defensible. ⁸⁹¹¹

Conclusion

Parallel imaging is one of the most valuable tools in modern MRI, but it is not free, and treating the acceleration factor as a simple speed knob hides a real, quantifiable, spatially varying SNR cost. The physics is unambiguous: SNR falls by $R$ from undersampling and again by the g-factor from the reconstruction, and the g-factor grows as acceleration approaches the coil array's encoding limit. A strong MRI program chooses acceleration deliberately against an SNR budget, protects the calibration data the reconstruction depends on, measures SNR with methods valid under parallel imaging, and trends coil performance over time. Done that way, acceleration delivers faster, more robust scans without quietly eroding the diagnostic quality of the exact anatomy the study is meant to show. ¹⁶¹¹

How DRPS Can Help

Diagnostic Radiation Physics Services (DRPS) supports MRI facilities with MRI physics testing, acceptance testing and commissioning, annual equipment-performance evaluations, coil SNR and uniformity assessment, artifact and parallel-imaging problem solving, protocol review, and accreditation support prepared by board-certified medical physicists. DRPS serves facilities across Florida, Maryland, Virginia, Washington DC, California, Nevada, New York, Pennsylvania, New Jersey, and Delaware — see our service locations.

A strong MRI QC program does more than pass accreditation. It makes sure the acceleration settings running on the scanner every day protect the SNR the diagnosis actually depends on.

Related Resources

References

Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: sensitivity encoding for fast MRI. Magnetic Resonance in Medicine. 1999;42(5):952-962. doi:10.1002/(SICI)1522-2594(199911)42:5<952::AID-MRM16>3.0.CO;2-S. PubMed
Griswold MA, Jakob PM, Heidemann RM, et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magnetic Resonance in Medicine. 2002;47(6):1202-1210. doi:10.1002/mrm.10171. PubMed
Robson PM, Grant AK, Madhuranthakam AJ, Lattanzi R, Sodickson DK, McKenzie CA. Comprehensive quantification of signal-to-noise ratio and g-factor for image-based and k-space-based parallel imaging reconstructions. Magnetic Resonance in Medicine. 2008;60(4):895-907. doi:10.1002/mrm.21728. PubMed
Breuer FA, Kannengiesser SAR, Blaimer M, Seiberlich N, Jakob PM, Griswold MA. General formulation for quantitative G-factor calculation in GRAPPA reconstructions. Magnetic Resonance in Medicine. 2009;62(3):739-746. doi:10.1002/mrm.22066. PubMed
Ohliger MA, Sodickson DK. An introduction to coil array design for parallel MRI. NMR in Biomedicine. 2006;19(3):300-315. doi:10.1002/nbm.1046. PubMed
Deshmane A, Gulani V, Griswold MA, Seiberlich N. Parallel MR imaging. Journal of Magnetic Resonance Imaging. 2012;36(1):55-72. doi:10.1002/jmri.23639. PubMed
Larkman DJ, Hajnal JV, Herlihy AH, Coutts GA, Young IR, Ehnholm G. Use of multicoil arrays for separation of signal from multiple slices simultaneously excited. Journal of Magnetic Resonance Imaging. 2001;13(2):313-317. doi:10.1002/1522-2586(200102)13:2<313::AID-JMRI1045>3.0.CO;2-W. PubMed
American Association of Physicists in Medicine. Report No. 100: Acceptance Testing and Quality Assurance Procedures for Magnetic Resonance Imaging Facilities. College Park, MD: AAPM; 2010. aapm.org
National Electrical Manufacturers Association. NEMA MS 9-2008 (R2014, R2020): Characterization of Phased Array Coils for Diagnostic Magnetic Resonance Images. Rosslyn, VA: NEMA. nema.org
National Electrical Manufacturers Association. NEMA MS 1-2008 (R2014): Determination of Signal-to-Noise Ratio (SNR) in Diagnostic Magnetic Resonance Imaging. Rosslyn, VA: NEMA. nema.org
American College of Radiology. ACR MRI Quality Control Manual. Reston, VA: ACR; 2015. acr.org