Number of Excitations (NEX) / Number of Signal Averages (NSA)

MRIninja Knowledge Base | MRI Parameter Deep Dive Version 1.0 — May 2026

MRIninja Knowledge Base | Parameter Child Page Parent page: MRI Parameters — Overview and Classification (9501) Related pages: Acquisition Matrix · Parallel Imaging · Slice Thickness · 2D vs 3D Acquisition Version 1.0 — May 2026

1. Introduction and General Purpose

NEX (Number of Excitations, GE/Canon terminology) and NSA (Number of Signal Averages, Siemens/Philips terminology) denote the same acquisition parameter: the number of times each phase-encoding line in k-space is measured before proceeding to the next line. The averaged measurement yields an SNR improvement proportional to the square root of the number of averages — the fundamental relationship that governs all signal averaging in MRI and, by extension, in all of physics where random noise is the limiting factor.

NEX/NSA is the simplest mechanism for SNR manipulation in MRI. Unlike most other parameters (matrix, FOV, flip angle), which affect SNR indirectly through their effect on voxel volume or tissue magnetisation, NEX/NSA has a single, clean, predictable relationship with both SNR and acquisition time:

SNR ∝ √NEX T_acq ∝ NEX

These two proportionalities define the fundamental inefficiency of NSA as an SNR recovery tool: to double SNR via averaging requires quadrupling the scan time. This is why NSA is not the primary tool for SNR optimisation in modern MRI — it is the last resort when all other SNR mechanisms (coil selection, field strength, voxel size, bandwidth) have been exploited.

Nevertheless, NSA remains clinically important in specific contexts where its unique properties — simplicity, universality, and the ability to average out random motion artefacts — make it the preferred or only viable approach. High-NSA free-breathing DWI (NSA=4–8) for body applications; NSA=2 for marginal SNR situations; fractional NSA=0.5 (Partial Fourier equivalents); and NSA averaging in MR spectroscopy (NSA=64–256) — these are the domains where understanding NSA deeply matters for protocol quality.

Historical context: signal averaging was the original mechanism for noise reduction in all NMR spectroscopy from the 1950s. Every early NMR spectrometer operated by accumulating multiple free-induction decay (FID) signals and averaging them to improve SNR. When MRI was developed in the 1970s–1980s, this concept was directly imported. The first clinical MRI scanners routinely used NSA=2–4 because the hardware SNR was insufficient for NSA=1, and signal averaging was the primary quality tool. As coil technology, field strength, and parallel imaging improved through the 1990s–2020s, NSA=1 became feasible for most applications, and NSA reduction (rather than increase) became a common optimisation tool.

2. Physical Foundations

2.1 Signal Averaging and Noise Reduction

MRI signal S is deterministic: for a given TR, TE, flip angle, and tissue type, the signal from a voxel is fixed and reproducible. MRI noise η is stochastic (random): it arises from thermal motion of electrons in the receiver coil conductors (Johnson-Nyquist noise) and from losses in the patient (dielectric noise). For each measurement of a k-space line:

Measured_value = Signal + Noise

where Noise is drawn from a zero-mean Gaussian distribution with standard deviation σ_noise.

When N measurements of the same k-space line are averaged:

Average = Signal + (1/N) × Σ Noise_i

Because the noise terms are independent (uncorrelated between measurements), the standard deviation of their average is:

σ_average = σ_noise / √N

The signal is unchanged (it is the same deterministic value each time); the noise decreases as 1/√N. Therefore:

SNR_N = SNR_1 × √N

This is the foundational equation of signal averaging and explains both its benefit (SNR improves) and its cost (SNR improves only as √N — a diminishing-returns relationship).

2.2 Mathematical Foundations

SNR and NEX: SNR_NEX = SNR_1 × √NEX

where SNR_1 = SNR at NEX=1 (single acquisition).

Clinical meaning: each doubling of NEX improves SNR by √2 ≈ 41%. Each quadrupling of NEX doubles SNR.

Acquisition time and NEX: T_acq = TR × N_y × NEX / ETL

Clinical meaning: T_acq is directly proportional to NEX. Doubling NEX exactly doubles scan time.

SNR efficiency (SNR per unit time): SNR_efficiency = SNR_NEX / √T_acq = SNR_1 × √NEX / √(TR × N_y × NEX / ETL) = SNR_1 × √(ETL / (TR × N_y))

Key insight: the SNR efficiency is independent of NEX. Adding NSA averages does not improve SNR per unit time — it only improves absolute SNR at proportionally greater time cost. This is why NSA is the least efficient SNR improvement strategy.

Fractional NEX (0.5 NSA / Partial Fourier):

Some vendors allow NEX < 1 by acquiring only a fraction of k-space lines (partial Fourier in the NSA direction). At NEX=0.5 (Philips: “0.5 NSA” or Siemens: “Partial Fourier 4/8 in phase”): - Acquired lines: 50% of N_y - Time: halved - SNR: approximately 1/√2 of full (similar to Partial Fourier phase)

This is distinct from true NSA averaging — it is k-space reduction, not averaging.

Motion averaging vs noise averaging:

A subtle but clinically important distinction: NEX averaging reduces random noise (Gaussian noise) proportionally to √N. It does NOT reduce coherent artefacts (systematic motion ghosting, susceptibility artefacts, chemical shift artefacts) because these are correlated between averages. Motion ghosting from periodic cardiac or respiratory motion may in fact be worsened by high NSA if the motion period is commensurate with the averaging interval.

However, for incoherent (aperiodic) motion — such as involuntary patient micro-tremor, bowel peristalsis, or random patient movement — averaging does partially reduce the artefact magnitude, because the random motion is not phase-coherent across averages. This is the basis for high-NSA free-breathing body DWI: incoherent respiratory motion averages out partially across NSA=4–8.

3. Units, Terminology and Vendor Nomenclature

NEX/NSA is dimensionless (a count of measurements per k-space line). It may also be expressed as a decimal (0.5 NSA for partial k-space strategies that some vendors implement using the NSA framework).

GE NEX=0.5: GE specifically implements fractional NEX as a first-class parameter option. NEX=0.5 collects only the positive-kx half of k-space (or equivalently, every other phase-encoding line in a specific pattern) and reconstructs using conjugate symmetry. This is equivalent to half-Fourier or partial Fourier reconstruction. At NEX=0.5: T_acq = T_full/2; SNR ≈ T_full/√2. Very useful for breath-hold body acquisitions where halving scan time is critical.

Philips NSA=0.5: identical to GE NEX=0.5 in implementation; uses the same partial k-space reconstruction (homodyne). Available in Philips SE and TSE sequences.

4. Typical Value Ranges

4.1 NEX/NSA by Application and Clinical Context

4.2 NSA by Field Strength

5. Parameter Interaction Ecosystem

5.1 Parameter Relationships Matrix

5.2 NSA vs Parallel Imaging: The SNR-Time Trade-off

A common protocol question: when SNR is insufficient, should I increase NSA or decrease R?

Conclusion: R=2 + NSA=2 has approximately the same scan time as R=1 + NSA=1, but lower SNR (by the g-factor ≈ 10–30%). Reducing R is marginally more SNR-efficient than adding NSA at the same time cost, because parallel imaging avoids the √R penalty for the k-space lines that don’t need to be re-acquired.

In practice: use parallel imaging to control time; use NSA to provide incremental SNR top-up when R is already optimised.

6. Effects on Image Appearance

6.1 Increasing NSA

Noise reduction: the visual texture of the image becomes smoother. Random noise (grainy appearance) is reduced proportionally to 1/√NSA. This is the primary visual effect.

No change in spatial resolution: the voxel size is unchanged. Edge sharpness (reflecting the acquisition matrix and resolution) is identical at NSA=1 and NSA=4. Only the noise texture changes.

No change in tissue contrast: TR, TE, and other contrast parameters are unchanged. The tissue signal intensities are identical at any NSA; only the noise floor decreases.

Motion artefact behaviour: - Random (incoherent) motion: ghost amplitude is partially reduced by averaging (though not eliminated — only the incoherent component averages out) - Periodic (coherent) motion: ghost amplitude may stay the same or increase at high NSA, because coherent signals add constructively while noise averages away → relative ghost amplitude increases

Practical visual change: at NSA=1 → NSA=4: the image appears “cleaner” and has a smoother background. Small structures that were at the edge of the noise floor may become more conspicuous. However, the structural detail (the highest spatial frequencies in the image) is unchanged.

6.2 Decreasing NSA (Including NSA < 1 via Partial Fourier)

Increased noise: the image appears grainier. At NSA=0.5 (half-Fourier): approximately 41% more noise than NSA=1.

Reduced scan time: the primary reason for reduced NSA. For breath-hold body MRI, NSA=0.5 (if available) enables significantly shorter acquisitions.

Gibbs ringing: for true partial Fourier (NSA=0.5 using homodyne reconstruction), mild additional Gibbs ringing may appear compared with full k-space NSA=1. The truncation artefact from half-Fourier is slightly more pronounced.

7. Effects on Acquisition Time

7.1 Direct and Linear

T_acq = TR × N_y × NSA / ETL

NSA multiplies T_acq directly. Every incremental NSA increment adds exactly one full “pass” through the N_y phase-encoding steps:

7.2 The Diminishing Return Problem

The ratio of SNR gain to time cost for NSA:

The efficiency is constant regardless of NSA — SNR always scales as √NSA = √time (for any fixed baseline). The “cost” of NSA is that to achieve N× the SNR, you must spend N² × the time. To double SNR via NSA requires 4× the scan time. This is uniquely inefficient compared with other SNR mechanisms: - Doubling field strength: 2× SNR at same scan time (but high cost, fixed parameter) - Halving slice thickness: 0.5× SNR (actually costs SNR) - Doubling voxel size: 2× SNR at same scan time (but resolution cost) - NSA=4: 2× SNR at 4× scan time

8. Effects on SNR and CNR

8.1 SNR

SNR_NSA = SNR_1 × √NSA

This relationship holds for true signal averaging (each k-space line measured NSA times, values averaged before reconstruction). It is exact for Gaussian noise.

SNR table for practical NSA values (relative to NSA=1):

8.2 CNR

CNR = (S_tissue1 − S_tissue2) / σ_noise

Since NSA averaging reduces σ_noise by 1/√NSA and leaves S unchanged, CNR improves proportionally:

CNR_NSA = CNR_1 × √NSA

Clinical interpretation: for a small lesion where the contrast difference from background is small, doubling NSA improves CNR by 41%. If the lesion was at the margin of detectability (CNR = 3–5), this improvement may be sufficient to cross the perceptual threshold and make the lesion detectable.

8.3 Field-Strength Dependency

At higher field, the baseline SNR is higher (SNR ∝ B₀ approximately). The same NSA applied at different field strengths produces proportionally different absolute SNR:

Key principle: NSA=4 at 0.55T produces equivalent absolute SNR to NSA=1 at approximately 1.5T (SNR ≈ 0.70 vs 1.0). This quantifies how many additional averages are required to compensate for lower field strength — and why 0.55T protocols often require NSA=2–4 while 3T protocols use NSA=1 for the same sequence.

9. Artefacts Associated with NEX/NSA

10. Behaviour Across Sequence Families

Spin Echo (SE)

NSA adds direct TR periods. For SE with TR=2000 ms, NSA=2: each of 256 phase-encoding lines is acquired twice consecutively → T_acq = 2000 × 256 × 2 = 1,024 s ≈ 17 min. NSA=2 in SE is rare; TSE with higher ETL is almost always preferred. NSA=1 is standard for SE applications.

Turbo Spin Echo (TSE)

NSA is ETL-divided, making higher NSA more practical. At TR=3500/ETL=16/N_y=256/NSA=2: T_acq = 3500 × 256 × 2 / 16 = 112 s ≈ 2 min. NSA=2 in TSE is common for spine protocols at 1.5T where SNR is marginal for thin slices.

Gradient Echo (GRE)

At very short TR (3–5 ms), each NSA average adds N_y × TR milliseconds. For GRE with TR=5 ms, N_y=200, NSA=2: T_acq = 5 × 200 × 2 = 2000 ms = 2 s. NSA=2 is commonly used in 2D GRE for brain T2* (SWI equivalent 2D) and rapid surveys.

Inversion Recovery (STIR, FLAIR)

Same as TSE. For STIR at 1.5T: NSA=2 is standard for whole-body STIR (lower SNR per station due to large FOV and body coil). For brain FLAIR: NSA=1 is standard.

EPI (DWI, fMRI, DSC)

In EPI, NSA averages are implemented slightly differently: in single-shot EPI (whole brain acquired in one shot), each “NSA” average is a complete brain volume acquisition. The averaging is performed in image space (not k-space line by line). The effect is identical to k-space averaging for Gaussian noise.

For brain DWI: NSA=1 at 3T is standard (adequate SNR). NSA=2 at 1.5T is common. NSA=4 is used for high b-value DWI (b=2000) where SNR is very low.

For body DWI (free-breathing): NSA=4–8 serves a dual purpose: (1) SNR improvement (×2–2.83); (2) partial motion averaging — respiratory motion is incoherent between averages and partially cancels.

Dixon

NSA applies to the underlying GRE or TSE readout. Dixon fat-water separation operates on the reconstructed images; it is independent of NSA. Higher NSA → better SNR → more robust Dixon reconstruction for fat quantification.

DCE

For DCE, NSA is effectively fixed at 1 (each dynamic phase is a single volume acquisition). Temporal resolution requirements prevent NSA > 1 in any standard DCE protocol. The exception: if DCE is performed with a slow-injection kinetic model where temporal resolution is relaxed (> 60 s per phase), NSA=2 per phase is occasionally used in research.

ASL

ASL uses many “averages” (label-control pairs), but these are not conventional NSA in the k-space sense. Each pair contributes one measurement of the labelled vs control brain. Typically 30–80 pairs are acquired and averaged to produce one perfusion map — this is the functional equivalent of NSA=30–80 for the perfusion signal (which is only ~1% of the baseline MRI signal).

The perfusion SNR: SNR_perfusion ≈ M₀ × label_efficiency × √(N_pairs) / σ_noise

Adding more pairs improves SNR as √N_pairs — the same √NSA relationship.

bSSFP

bSSFP sequences are steady-state acquisitions. Repeating the acquisition (NSA=2) would require two complete bSSFP steady-state transitions (dummy TRs + imaging), effectively doubling scan time. For 2D cardiac cine, NSA=1 is invariant (temporal averaging is achieved by the multi-cardiac-cycle k-space filling). For 3D bSSFP (CISS/inner ear), NSA=1 is standard.

Spectroscopy (SVS)

MR spectroscopy is one of the few MRI applications where very high NSA is not only appropriate but necessary. The absolute sensitivity of metabolite signals (NAA, Cho, Cr at millimolar concentrations) is 10,000–100,000× lower than water. Standard clinical 1H SVS: NSA=64–256.

For a single-voxel Cho peak to be detectable at SNR=5 from a 2 cm³ voxel at 3T: NSA ≈ 64 averages (8× SNR improvement). Acquisition time: TR=2000 ms × 64 = 128 s ≈ 2 min — acceptable.

For 31P spectroscopy (intrinsically lower γ and lower tissue concentration): NSA=128–512; acquisition times of 15–30 min are not uncommon.

11. Field Strength Behaviour

At 0.55T: the standard NSA for many body and MSK sequences is 2–4 to maintain diagnostic SNR. DLR (deep learning reconstruction) at 0.55T effectively provides an SNR boost equivalent to NSA=4–16 without the time cost — fundamentally changing the NSA strategy at low field.

At 3T: NSA=1 is the standard for virtually all brain, MSK, and body sequences. NSA=2 is occasionally needed for thin-slice acquisitions (< 2 mm) or small structures at the edge of the coil’s sensitivity. NSA > 2 at 3T is rarely justified outside spectroscopy and body DWI.

SAR at 3T: high NSA (NSA=4+) at 3T with TSE may trigger the SAR limit, causing automatic TR extension. This extends the scan time further and changes the T1 contrast. Always check TR after setting high NSA at 3T.

12. Vendor-Specific Implementation

Siemens

“Number of averages” (NSA) parameter in the contrast tab. Standard range: 0.5 (via partial Fourier option) to 16 (typical maximum displayed). Siemens provides a direct readout of the resulting scan time after any NSA change. The “PACE” (Prospective Acquisition Correction) respiratory gating tool provides a functional alternative to high NSA for motion management in body acquisitions.

GE

“NEX” parameter. GE specifically supports NEX=0.5 as a first-class clinical option (fractional excitation via partial Fourier). NEX values can be non-integer (0.5, 0.75, 1.5) on some GE sequences. For DWI, GE implements the effective NSA averaging within the DW direction framework (separate control from phase NEX).

Philips

“NSA” parameter with 0.5 option. Philips also uses the “Half scan” parameter (separate from NSA=0.5) as the partial Fourier equivalent. The Philips system gives explicit time penalty feedback when NSA is changed.

Canon

“NSA” or “Number of averages.” Standard integer values. Canon’s AiCE DLR is particularly effective at recovering SNR lost from NSA=1 acquisitions at 1.5T — reducing the need for NSA=2 in Canon 1.5T protocols.

United Imaging

Standard “NSA” parameter. UIH systems provide explicit SNR simulation based on current protocol parameters including NSA, allowing the technologist to estimate the expected image SNR before acquisition.

Hidden coupling — all vendors: when NSA is increased and the total scan time would exceed the patient slot time, some vendor auto-planners automatically reduce ETL or matrix to maintain scan time. The technologist must verify that these auto-adjustments have not unexpectedly changed the resolution or contrast.

13. Practical Optimisation Strategies

13.1 Clinical Optimisation Recipes

13.2 NSA vs Alternative SNR Strategies — Priority Order

When SNR is insufficient, the strategies should be applied in this order of efficiency:

1. Coil optimisation (use dedicated coil, not body coil) — most efficient; zero scan time cost
2. Field strength increase — highest SNR gain; not always available
3. Voxel size increase (increase slice thickness or reduce matrix) — efficient but resolution cost
4. Reduce bandwidth — modest SNR gain; affects minimum TE and chemical shift
5. Increase parallel imaging R reduction (reduce R) — recovers g-factor penalty; modest time cost
6. DLR (enable deep learning reconstruction) — effective at maintained scan time; texture change
7. Increase NSA — least efficient (diminishing returns); last resort; most time-expensive

14. Parameter Extremes

14.1 NSA < 1 (Fractional NEX: 0.5)

Implemented as partial Fourier (homodyne reconstruction). Collects 50–60% of k-space; reconstructs using conjugate symmetry. At NEX=0.5: - T_acq = 0.5 × T_full (approximately; additional Nyquist lines are acquired for phase correction) - SNR ≈ 0.71× T_full SNR - Gibbs ringing slightly more pronounced - Available on GE (NEX=0.5) and Philips (NSA=0.5)

Primary application: breath-hold body acquisitions where 20-second breath-hold is too long; NEX=0.5 brings it to ~10 seconds.

14.2 Very High NSA (NSA > 8)

Beyond NSA=8 (2.83× SNR), the time cost becomes extreme for standard clinical sequences: - NSA=16: 4× SNR at 16× time — rarely justified for imaging - NSA=64 (spectroscopy standard): 8× SNR at 64× time — appropriate because the baseline MRS SNR is so low that no other strategy is viable - NSA=256 (research spectroscopy): 16× SNR at 256× time — used for 31P or 13C spectroscopy where metabolite concentration is extremely low

For imaging sequences, NSA > 4 is almost never the optimal strategy. The same scan time invested in higher ETL, lower R, or 3D acquisition almost always produces better results.

15. Common Optimisation Errors

16. MRI Technologist Pearls

NSA is the most time-expensive SNR tool — exhaust all other options first: before increasing NSA, ask: is the coil optimal? Is the bandwidth as narrow as possible? Is the slice thickness acceptable? Is DLR available? Is the current R value the most conservative reasonable? Only after all these options are considered should NSA be increased.

The √NSA rule in one sentence: doubling the SNR requires quadrupling the scan time. If a 2-minute sequence is noisy at NSA=1, achieving ×2 SNR requires 8 minutes (NSA=4). This simple calculation is the most important decision tool for NSA optimisation.

For body DWI, high NSA is different — it’s not just about noise: in free-breathing body DWI, NSA=4–6 serves a dual purpose. The noise averaging (×2–2.45 SNR) is one component; the incoherent respiratory motion averaging is the other. For this specific application, high NSA is the appropriate strategy even though its SNR efficiency is low, because there is no substitute for motion averaging in free-breathing protocols.

Monitor TR at 3T after NSA changes: at 3T with TSE sequences, adding NSA=2 doubles the RF pulse count per unit time. If the protocol was at the SAR limit, the scanner will automatically extend TR. This TR change is not always obvious and can significantly alter T1 contrast. Always verify TR after NSA changes at 3T.

NSA and spectroscopy are inseparable: for MR spectroscopy (SVS), NSA=64–128 is not optional — it is the fundamental mechanism for achieving diagnostic SNR. Unlike imaging where alternatives exist, MRS has no equivalent substitute for signal averaging. Understanding that MRS acquisitions at NSA=128 are not “inefficient” but physically necessary is important for protocol justification.

Document NSA for serial studies: NSA is part of the acquisition parameter set for serial comparison studies (AD volumetry, tumour response, demyelination monitoring). Changing NSA between examinations changes the noise texture and apparent lesion conspicuity without any biological change. Include NSA in the technique documentation.

17. Real Clinical Examples

Example 1: Spine Protocol at 1.5T — NSA=2 for Marginal SNR

Clinical scenario: cervical spine T2 TSE sagittal at 1.5T; 3 mm slices; 256×192 matrix; the initial NSA=1 acquisition shows adequate quality for the disc spaces but the cord appears marginally noisy (SNR ≈ 15; borderline for cord signal characterisation).

NSA=1: T_acq = 3000 × 192 / 16 = 36,000 ms = 36 s. SNR cord = 15.

NSA=2: T_acq = 36 × 2 = 72 s. SNR cord = 15 × √2 = 21. The cord is now clearly characterised with adequate CNR for detecting intramedullary signal change.

Alternative considered: reducing bandwidth from 200 Hz/px to 100 Hz/px → SNR improves by √2 → same SNR gain without time cost, but chemical shift displacement doubles (from 1.1 pixel to 2.2 pixels at the fat-water interfaces adjacent to the cord) — acceptable.

Decision: use reduced BW (100 Hz/px) instead of NSA=2 → same SNR, no time cost, slightly more chemical shift artefact but clinically acceptable. NSA=2 reserved for cases where BW reduction is already maximised.

Example 2: Body DWI — High NSA for Free-Breathing Liver DWI

Clinical scenario: liver DWI at 3T for HCC detection; 5 mm slices; FOV 360 mm; b=0, 50, 800 s/mm²; free-breathing acquisition.

NSA=2: T_acq per b-value = 6000 × 96 × 2 / 1 = 1,152 s — far too long. Corrected: EPI single-shot; TR per slice = 4000 ms; N_slices = 32; T_acq = 4000 × 32 × 2 / 1 = 256 s ≈ 4.3 min per b-value. At NSA=2: SNR_DWI = SNR_1 × 1.41; partial motion averaging → some ghosting reduction.

NSA=6: T_acq = 4000 × 32 × 6 / 1 = 768 s ≈ 12.8 min for all b-values. SNR_DWI = SNR_1 × 2.45; substantially improved motion averaging → DWI lesion conspicuity markedly improved vs NSA=2.

Clinical result: a 1.4 cm HCC in segment VI (low contrast on morphological sequences) shows restricted diffusion (ADC = 1.0 × 10⁻³ mm²/s) on the NSA=6 acquisition. On NSA=2, the same lesion was barely above the noise floor.

Protocol decision: NSA=6 is the standard for free-breathing liver DWI at most centres (as per BHS and EPOS guidelines) — the high NSA is accepted as a mandatory component of the protocol for adequate diagnostic quality.

Example 3: MR Spectroscopy — NSA=128 for Hippocampal Metabolites

Clinical scenario: suspected early AD; hippocampal single-voxel MRS for NAA/Cr ratio assessment. Voxel size: 2 cm³; TR=2000 ms; TE=30 ms.

NSA=1: single FID acquisition; hippocampal metabolite SNR ≈ 0.3 (completely uninterpretable; metabolite peaks invisible).

NSA=16: T = 2000 × 16 = 32 s; SNR = 0.3 × 4 = 1.2 (still inadequate for quantification).

NSA=128: T = 2000 × 128 = 256 s = 4.3 min; SNR = 0.3 × 11.3 = 3.4 (marginal but interpretable for NAA/Cr ratio with peak fitting).

NSA=256: T = 512 s = 8.5 min; SNR = 0.3 × 16 = 4.8 (adequate for robust NAA/Cr and Cho/Cr quantification).

Protocol: NSA=128 as the clinical minimum; NSA=256 for research-grade quantification. These high NSA values are not optional — they are physically required for the MRS signal to rise above the noise floor.

Example 4: 0.55T Protocol — NSA=4 vs DLR

Clinical scenario: knee MRI at 0.55T (low-field open scanner for claustrophobic patient); T2 TSE coronal; 3 mm slices. Standard 3T protocol has NSA=1. At 0.55T with same protocol: SNR ≈ 0.35× 3T SNR → visually inadequate.

Option A — NSA=4: SNR = 0.35 × 2.0 = 0.70× 3T. Scan time: 4× longer than 3T protocol. At TR=3500, ETL=12, N_y=256, NSA=4: T_acq = 3500 × 256 × 4 / 12 = 298 s ≈ 5 min. Marginal quality; acceptable for meniscus survey.

Option B — NSA=1 + DLR (Canon AiCE or UIH equivalent): T_acq = 74 s (NSA=1). DLR applied post-acquisition provides SNR equivalent to approximately NSA=4–8 of standard reconstruction, without additional scan time. Image quality approaches diagnostic standard for meniscus/cruciate assessment.

Decision at this centre: NSA=1 + DLR at 0.55T. The DLR strategy is 4× faster than NSA=4 with equivalent or superior SNR. This exemplifies the future trend: DLR replacing NSA as the primary SNR rescue tool at low and standard field.

18. Visual Educational Material

18.1 SNR vs NSA — The Diminishing Returns Curve

SNR multiple
  16× |                                                     ● NSA=256
      |
   8× |                                         ● NSA=64
      |
   4× |                             ● NSA=16
      |
   2× |               ● NSA=4
      |
1.41× |        ● NSA=2
      |
   1× | ● NSA=1
      |───────────────────────────────────────────────────
         1    4    16    64    256   NSA (time multiple)

OBSERVATION: Each doubling of SNR requires quadrupling of time
             The curve flattens — diminishing returns with every increase

PRACTICAL RANGE:
  NSA=1:   Standard clinical imaging (1.5T–3T)
  NSA=2–4: Low SNR applications; 0.55T; thin slices
  NSA=4–8: Free-breathing body DWI; specific body applications
  NSA=64+: MR spectroscopy ONLY

18.2 NSA vs Other SNR Strategies — Efficiency Comparison

To achieve 2× SNR improvement:

Strategy          | Time cost | Resolution cost | Other cost
──────────────────|──────────|─────────────────|──────────
NSA × 4           | × 4      | None            | Time only
2× slice thickness| × 0.5    | Through-plane ↑ | Resolution loss
1/2 bandwidth     | × 0 (≈)  | None            | ↑ Chemical shift
Field strength ×2 | × 0      | None            | Equipment cost (fixed)
DLR enable        | × 0      | None(ish)       | Texture change
R=2 → R=1        | × 2      | None            | Time × 2

LESSON: NSA is the most time-expensive SNR tool available.
        It should be the last resort, not the first response.

18.3 NSA vs Free-Breathing Motion — Decision Tree

IS THE MOTION COHERENT (cardiac, respiratory rhythmic) OR INCOHERENT (random)?

Coherent motion (cardiac ghost, aortic pulsation, respiratory ghost):
  → HIGH NSA does NOT help (coherent signal adds; noise averages → relative ghost worsens)
  → Use: cardiac gating / respiratory triggering / saturation bands / phase direction change

Incoherent motion (random patient movement, bowel peristalsis, involuntary tremor):
  → HIGH NSA PARTIALLY helps (incoherent positions partially average)
  → Use: NSA=4–8 for body DWI (free-breathing); accept residual incoherent motion contribution

Pure noise (no motion, standard imaging):
  → HIGH NSA helps (noise averages as √NSA)
  → Use: when absolutely needed; prefer other SNR strategies first

19. Evidence Gaps and Ongoing Debate

DLR as NSA replacement at 0.55T: deep learning reconstruction at low-field (0.55T) systems (GE SIGNA Open, UIH uMR 550W) has been proposed as a replacement for high NSA protocols, providing equivalent SNR at NSA=1 compared with standard NSA=4. Prospective clinical studies comparing DLR-NSA=1 vs standard-NSA=4 for specific diagnostic tasks (knee meniscus, spine cord signal) at 0.55T have not been published at the time of writing. Vendor claims require independent validation.

Free-breathing DWI NSA optimisation: the optimal NSA for body DWI (liver, prostate, kidney) in terms of the trade-off between scan time and image quality has not been formally optimised across patient populations. Values of NSA=4–8 are based on expert consensus and empirical evaluation rather than formal prospective SNR-outcome studies.

NSA and AI reconstruction synergy: the combination of high NSA (for robust signal averaging) followed by DLR (for noise reduction without blurring) may provide image quality superior to either approach alone. Whether this combination is more effective than maximising NSA alone or using DLR at lower NSA has not been formally compared for clinical applications.

NSA fractional values and Gibbs ringing: the clinical impact of NEX=0.5 / NSA=0.5 on Gibbs ringing severity for specific anatomical contexts (spine cord, brain cortex) has been described anecdotally but not quantified in large clinical studies. The threshold at which homodyne ringing becomes diagnostically significant for specific structures is institution-dependent.

20. Miscellaneous and Future Directions

Historical context: signal averaging was the original NMR quality tool before gradient-based MRI even existed. Richard Ernst’s 1966 paper [1] introducing Fourier transform NMR used signal averaging as the primary SNR mechanism. The concept was directly imported to MRI in the 1970s. The very first MRI images (Lauterbur 1973) benefited from signal averaging.

Fractional NEX and conjugate symmetry: the concept of using conjugate symmetry of k-space to reduce acquisition by ~50% was introduced in the late 1980s (MacFall et al., 1986 [2]) as “partial Fourier” or “half-Fourier” imaging. GE implemented this as NEX=0.5, providing a practical fractional averaging option that has remained a standard feature.

DLR as functional NSA replacement: deep learning reconstruction — when applied to NSA=1 acquisitions — effectively provides the SNR benefit of higher NSA without the time cost. The trained neural network acts as a non-linear, structure-aware noise filter. The key difference: DLR’s noise reduction is informed by the image content (it preserves edges and structures while reducing noise), whereas NSA averaging is agnostic (reduces all noise including that at high-signal regions). DLR is therefore superior in SNR efficiency to NSA, but has a different artefact profile (smoothing, hallucination risk at extreme settings).

Compressed sensing and NSA: compressed sensing (CS) acceleration exploits sparsity rather than SNR averaging. A CS-accelerated acquisition at 1/4 the k-space lines achieves similar SNR to NSA=4 of a fully-sampled acquisition — but in 1/4 the time. The CS reconstruction essentially performs structured “averaging” in the transform domain rather than in the data domain. This positions CS as a fundamentally more efficient substitute for NSA when the image has suitable sparsity properties.