Parallel Imaging

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 · FOV — Field of View · 2D vs 3D Acquisition · Phase Oversampling Version 1.0 — May 2026

1. Introduction and General Purpose

Parallel imaging is the most consequential advance in MRI acquisition technique since the introduction of turbo spin echo (TSE) in the late 1980s. It exploits the spatially distinct sensitivity profiles of multiple receiver coil elements simultaneously to replace a fraction of the k-space phase-encoding gradient steps, thereby reducing acquisition time proportionally to the acceleration factor (R) — or, equivalently, enabling higher spatial resolution at fixed acquisition time.

Every modern clinical MRI scanner uses parallel imaging as a standard component of most protocols. It is not an optional “premium” feature but the foundational technology that makes contemporary MRI — sub-millimetre brain volumetry in 5 minutes, breath-hold abdominal DCE in 20 seconds, high-resolution knee cartilage at 3T — clinically feasible within acceptable examination durations.

The physical basis of parallel imaging is the spatial encoding information contained in the coil sensitivity map of a multi-element phased array. When a receiver coil element is closer to one side of the patient than another, the signals it receives are disproportionately dominated by tissue on the near side. This spatial non-uniformity — normally an unwanted complication in MRI signal reception — becomes a source of spatial information when multiple elements with different sensitivity distributions are combined. Parallel imaging reconstruction algorithms decode this spatial information to un-alias deliberately undersampled k-space.

The cost of parallel imaging is an irreducible SNR penalty: SNR ∝ 1 / (g × √R) — where g is the geometry factor (g-factor), a dimensionless noise amplification coefficient that reflects how effectively the coil array separates the aliased signals. At R=2 with a well-designed coil: SNR reduces to approximately 60–70% of the full-sampled acquisition. This SNR cost must be balanced against the acceleration benefit for every clinical application.

Historical context: the theoretical foundations of parallel imaging were laid in the late 1980s by Hutchinson and Raff [1], who proposed using multiple receivers to acquire fewer k-space lines. The breakthrough clinical implementations — SMASH (1997) [2], SENSE (1999) [3], and GRAPPA (2002) [4] — transformed parallel imaging from theoretical concept to universal clinical tool within five years. Today, essentially all clinical MRI uses parallel imaging at R=2 as the default minimum.

2. Physical Foundations

2.1 The Spatial Encoding Problem

Standard MRI encodes spatial position using magnetic field gradients applied sequentially. Each phase-encoding gradient step requires one complete TR period. For N_y phase-encoding steps, the total acquisition time is N_y × TR / ETL. Parallel imaging’s insight: if the receiver coil array provides additional spatial encoding information — beyond what the gradient can provide — then fewer gradient steps are needed, and the “missing” spatial information can be recovered from the coil sensitivity map.

2.2 Mathematical Foundations

2.2.1 The SENSE Framework (Image-Space)

In SENSE (Sensitivity Encoding) [3], the k-space is undersampled by acquiring only every Rth phase-encoding line (acceleration factor R). The missing lines are not acquired — the reconstruction algorithm must recover them from the available data.

The effect of R-fold undersampling: the reconstructed (aliased) image shows R copies of the anatomical content superimposed (wrapped or aliased) within the reduced-FOV image. For R=2: the image is compressed to FOV/2, with anatomy from two locations superimposed at each pixel.

SENSE un-aliasing: for each pixel position (x, y) in the aliased image, the signal from a coil element c is:

S_c(x, y) = Σ_k [C_c(x, y_k) × ρ(x, y_k)]

where: - S_c = signal measured by coil c at aliased position (x, y) - C_c(x, y_k) = coil sensitivity of element c at each of the R aliased positions y_k - ρ(x, y_k) = true image intensity at position y_k

This system of equations (one per coil element, one unknown per aliased position) is solved by matrix inversion using the known coil sensitivity maps C_c:

ρ = (C^H C)^{-1} C^H S

The geometry factor (g-factor) quantifies how well-conditioned this inversion is:

g(x, y) = √[ (C^H C)^{-1}{ii} × (C^H C){ii} ] for position i

where g ≥ 1; g = 1 (best case: perfect separation); g >> 1 (worst case: coils cannot separate the aliased positions → high noise amplification).

SNR_SENSE = SNR_full / (g × √R)

Physical interpretation: g-factor is a map — it varies spatially. Regions where coil elements have poor spatial encoding diversity (e.g., the centre of the body far from all coil elements) have high g-factor → high local noise amplification. For standard phased-array surface coils, g ≈ 1.1–1.4 at R=2 and g ≈ 1.5–2.5 at R=3 for typical patient geometries.

2.2.2 The GRAPPA Framework (k-Space)

GRAPPA (Generalised Autocalibrating Partially Parallel Acquisitions) [4] operates in k-space rather than image space. Every Rth k-space line is acquired, plus a set of Auto-Calibration Signal (ACS) lines at the centre of k-space. The ACS lines are acquired at full density (no undersampling) and serve two purposes: (1) calibration of the GRAPPA reconstruction kernel; (2) contribution to the final image.

The GRAPPA reconstruction kernel estimates the missing k-space lines from the acquired neighbouring lines, using the coil-to-coil relationships learned from the ACS:

S_c(k_y_missing) = Σ_{c’, Δk_y} [n(c’, c, Δk_y) × S_{c’}(k_y_missing + Δk_y)]

where n(c’, c, Δk_y) are the GRAPPA reconstruction coefficients calibrated from the ACS.

The SNR of the GRAPPA reconstruction depends similarly on the g-factor:

SNR_GRAPPA ≈ SNR_full / (g_GRAPPA × √R)

where g_GRAPPA depends on the ACS size, the kernel size, and the coil geometry.

ACS lines: the number of central k-space lines acquired at full density for calibration. Typical values: 24–48 lines for standard GRAPPA at R=2–3. More ACS lines → better kernel calibration → lower g-factor → higher image quality; but more ACS lines reduce the time savings of parallel imaging proportionally.

2.2.3 The Effective Acceleration Factor

The actual acquisition time reduction from parallel imaging is:

T_parallel = T_full × (N_ACS + N_undersampled) / N_y_full

≈ T_full / R (for large N_y where ACS overhead is small)

More precisely: Effective R = N_y_full / (N_y_acquired)

where N_y_acquired = N_ACS + (N_y_full − N_ACS) / R.

For N_y=256, R=2, N_ACS=32: N_y_acquired = 32 + (256−32)/2 = 32 + 112 = 144 → effective R = 256/144 = 1.78 (not exactly 2). The ACS overhead reduces the effective acceleration below the nominal R.

2.2.4 Simultaneous Multi-Slice (SMS / Multiband)

SMS is parallel imaging in the slice direction. Multiple slices are excited simultaneously using multi-band RF pulses (one pulse at each slice-selection frequency), and the simultaneously acquired signals are separated using the coil sensitivity maps in the z-direction — analogous to SENSE applied in z.

T_SMS = T_single_slice_2D / MB

where MB = multiband factor (number of simultaneously excited slices). SNR:

SNR_SMS = SNR_single / (g_z × √MB)

where g_z is the slice-direction g-factor (depends on coil array z-separation).

SMS is the primary acceleration technique for brain fMRI (MB=4–8, enabling 64-slice whole-brain fMRI at 1.5-second TR), DWI, and some body applications.

3. Units, Terminology and Vendor Nomenclature

Parallel imaging acceleration factor R is dimensionless (a ratio). SMS multiband factor MB is also dimensionless.

GE ARC vs ASSET: - ASSET (Array Spatial Sensitivity Encoding Technique): SENSE-based, requires separate calibration scan; original GE parallel imaging - ARC (Autocalibrating Reconstruction for Cartesian sampling): GRAPPA-based, self-calibrated from ACS lines; newer GE implementation preferred for most applications

4. Typical Value Ranges

4.1 Acceleration Factor by Application

4.2 Maximum Practical Acceleration by Field Strength and Coil

5. Parameter Interaction Ecosystem

5.1 Parameter Relationships Matrix

6. Effects on Image Appearance

6.1 Increasing R

Scan time reduces proportionally. At equivalent time budget: thinner slices, higher matrix, more phases per dynamic series, or more sequences fit in the same scan.

Noise texture changes: at higher R, the noise distribution in the image becomes spatially non-uniform (reflecting the g-factor map). High-g-factor regions (typically the centre of large body cross-sections) show more noise than low-g-factor regions (areas near the coil elements). This spatially inhomogeneous noise is the characteristic signature of high-R parallel imaging and is absent in fully sampled or low-R acquisitions.

Residual aliasing artefact (if reconstruction is imperfect or if coil geometry is suboptimal for the prescribed R): “ghosting” appears at predictable locations determined by the undersampling pattern. For GRAPPA at R=2, potential residual aliasing appears as a faint copy of the anatomy at FOV/2 offset in the phase direction. For SENSE at R=2, the same position but from incomplete un-aliasing.

No change in tissue contrast: parallel imaging does not alter the T1, T2, or proton density weighting of the acquisition. The TE, TR, flip angle, and echo train structure are unchanged.

6.2 Decreasing R (More Conservative Acceleration)

Scan time increases proportionally. The image has: lower noise (higher SNR per voxel); more uniform noise distribution (lower g-factor contribution); less residual aliasing risk. At R=1 (no acceleration): conventional fully sampled acquisition with maximum SNR.

For most clinical sequences at R=2, decreasing to R=1 provides only a modest image quality improvement (SNR ×1.3–1.4) at double the scan time — rarely justified except for quantitative applications where every dB of SNR is critical.

7. Effects on Acquisition Time

7.1 Direct Time Reduction

T_parallel = T_full × (N_ACS + N_undersampled/R) / N_y

For practical approximation with large N_y (ACS overhead negligible):

T_parallel ≈ T_full / R

7.2 Time Savings — Multiple Uses

The time saved by parallel imaging can be applied in different ways:

1. Direct scan time reduction: the most common use. Shortening a 4-minute sequence to 2 minutes (R=2).

2. Higher spatial resolution: at fixed scan time, R=2 enables doubling of N_y (phase matrix) → halving the voxel size in the phase direction.

3. More slices per TR: the freed time allows additional slices to be packed into the TR interval.

4. Higher temporal resolution (DCE): in dynamic sequences, each 3D volume is acquired faster → more time points per contrast passage.

5. Longer coverage (3D): at fixed acquisition time, R=2 enables doubling of the 3D slab thickness at fixed resolution.

8. Effects on SNR and CNR

8.1 The Fundamental SNR Equation

SNR_R = SNR_full / (g × √R)

This equation determines the practical limits of parallel imaging. The two factors:

√R factor: the intrinsic SNR penalty from acquiring fewer k-space lines. Halving the acquired lines → √2 SNR penalty → unavoidable physics.

g-factor (g ≥ 1): additional noise amplification from the reconstruction algorithm. g = 1 (theoretical best) means the reconstruction introduces no additional noise. g > 1 means the coil geometry is not perfectly suited to the required un-aliasing → extra noise.

8.2 g-Factor — The Critical Quality Metric

g depends on: - Coil element count: more elements → lower g at given R - Coil geometry: elements must have distinct sensitivity profiles for the anatomical region - Phase encoding direction: coil separation must be in the phase-encoding direction for effective spatial encoding - FOV relative to anatomy: smaller FOV relative to the coil separation → higher g - R value: g increases non-linearly with R

g-factor maps: the g-factor is spatially variable. For any parallel imaging acquisition, the scanner can compute and display the g-factor map. The maximum g-factor in the imaging FOV determines the worst-case local SNR. Clinical imaging should keep g_max < 1.5 for routine applications; research applications may tolerate g_max < 2.5.

8.3 SNR at Practical R Values

8.4 3T Advantage for Parallel Imaging

At 3T, the baseline SNR is approximately 1.7–2× that at 1.5T. This extra SNR can be “invested” in higher acceleration: - 3T at R=3 has approximately (1.7 × 0.5) = 0.85× the SNR of 1.5T at R=1 — slightly below, but achievable - 3T enables meaningful R=3–4 for applications that are limited to R=2 at 1.5T

This is one of the primary clinical advantages of 3T: not just higher intrinsic SNR, but the ability to accelerate more aggressively to reduce scan time or improve resolution.

9. Artefacts Associated with Parallel Imaging

10. Behaviour Across Sequence Families

Spin Echo (SE)

Standard SE with parallel imaging at R=2: T_acq halved. The ACS lines in GRAPPA SE must use the same TR as the imaging lines to maintain T1 steady state. SE with R=2 is rarely used (TSE replaces SE for most applications), but in cardiac SE or diffusion SE, R=2 is standard.

Turbo Spin Echo (TSE)

The combination of parallel imaging (R) and long ETL (E) provides multiplicative time reduction: T_acq = T_SE / (R × E/E_2D). At R=2, ETL=16: time = 1/32 of standard SE — the entire clinical brain T2 TSE in ~2 minutes. 3D TSE (SPACE/CUBE) uses dual acceleration R_y × R_z for very efficient acquisition.

Gradient Echo (GRE)

DCE body and breast use R=3 as the standard for adequate temporal resolution. Cardiac GRE: R=2 as minimum. For single-shot GRE (FLASH), parallel imaging is applied in the segmented k-space acquisition.

Inversion Recovery (STIR, FLAIR, MPRAGE)

3D FLAIR (brain): R_y × R_z dual acceleration (total R=3–4); essential for clinically feasible 3D FLAIR. MPRAGE: typically R=2 (single direction); the ACS lines for GRAPPA must be placed at the same temporal position in the MP (magnetisation preparation) recovery curve as the surrounding lines — otherwise, ACS lines are T1-weighted differently → ACS calibration mismatch. Vendors handle this automatically; the technologist should not place ACS acquisition at a different delay time from the imaging acquisition.

EPI (DWI, fMRI, DSC)

EPI is the sequence most severely affected by parallel imaging, and where parallel imaging has the greatest impact:

In EPI, the phase-encoding dimension is traversed in a single shot. Each phase-encoding step takes one echo spacing (ESP ≈ 0.5–1 ms). For N_y=96 phase steps at ESP=0.8ms: total EPI readout = 76.8 ms. At R=2: N_y_acquired=48, readout = 38.4 ms.

The EPI readout duration determines: - Geometric distortion (scales with readout duration) - T2* signal decay (signal decays during the long readout)

R=2 in EPI halves both the distortion and T2* blurring — a major quality improvement beyond simple time savings. This is why parallel imaging R=2 is mandatory for clinical EPI DWI at 3T.

SMS for EPI: multiband EPI enables simultaneous acquisition of 4–8 slices per shot. For brain fMRI at TR=1 second with 40 slices at 2 mm: previously only achievable at TR=4 seconds (10 s/slice × 40 = 40 s → TR=40 s). With MB=4: TR=1 second for 40 slices. This revolutionised resting-state and task-based fMRI protocols.

Dixon

Dixon fat-water separation is applied to the Fourier-reconstructed data after parallel imaging reconstruction. Dixon requires specific TE values (IP/OP); these TE values are independent of parallel imaging R. The combination Dixon + parallel imaging is standard for body 3D GRE (mDixon VIBE at R=2–3).

DCE

The most time-constrained MRI application. The temporal resolution of DCE (duration of each 3D volume) determines the kinetic information extractable. For liver DCE arterial phase (≤ 20 s breath-hold):

At R=3, N_y=200, TR=4 ms, N_z=60, ETL=1: T_acq = 4 × 200 × 60 / (3 × 1) / 1000 = 16 s → fits in breath-hold.

Without R=3 (R=1): T_acq = 48 s → impractical for breath-hold.

Parallel imaging is the enabling technology for clinical liver DCE. Without it, the arterial phase would require a 4-second temporal resolution at most — inadequate for kinetic analysis.

ASL

ASL at 3T uses R=2 (in-plane) + background suppression. The low SNR of ASL (1% label fraction) means that parallel imaging SNR penalties are felt most acutely. R=2 is the practical maximum for most ASL applications at 3T; R=3 requires either very high NSA or DLR post-processing.

bSSFP (TrueFISP/FIESTA)

Cardiac bSSFP: R=2 standard. The bSSFP steady state is maintained despite k-space undersampling because the phase-encoding order (centric or linear) is consistent with the GRAPPA reconstruction. For 3D CISS/FIESTA-C (inner ear, brachial plexus root sleeves): R=2 × R=1.7 (dual acceleration) to achieve sub-millimetre resolution.

11. Field Strength Behaviour

At 3T, parallel imaging has three distinct benefits: 1. Time reduction (same as at 1.5T) 2. EPI distortion reduction (more important at 3T where distortion is 2× worse) 3. SAR management (critical at 3T; shorter TR due to parallel imaging reduces SAR accumulation per sequence)

At 7T: the B1+ inhomogeneity at 7T interacts with the coil sensitivity maps used for parallel imaging. In regions where B1+ is low (temporal lobes, cerebellum at 7T), the coil sensitivity is also reduced → higher g-factor in these regions. Dense 32+ channel coils and parallel transmit (pTX) are required for reliable parallel imaging at 7T.

12. Vendor-Specific Implementation

Siemens

iPAT (integrated Parallel Acquisition Technique) covers both GRAPPA and SENSE implementations. The technologist selects the acceleration factor and the algorithm (GRAPPA or SENSE). Siemens recommends GRAPPA as the default for most applications.

iPAT in 3D (Siemens): dual acceleration is set as “iPAT in phase direction” and “iPAT in partition direction” independently. The product of the two gives total acceleration.

Key Siemens feature: the g-factor map can be computed and displayed before the actual scan (using a pre-scan sensitivity map). This allows prospective verification that g_max is acceptable for the planned protocol.

GRAPPA ACS: Siemens uses “Reference Lines PE” to specify the ACS count. Default: 24–32 lines. More lines → better quality; more lines also reduce effective R.

GE

ARC (Autocalibrating Reconstruction for Cartesian) is the current standard GE parallel imaging implementation. ASSET (SENSE-based) requires a separate calibration scan; ARC is self-calibrating (uses ACS integrated into the acquisition). GE recommends ARC for most applications due to its motion robustness (no separate calibration scan that could become misregistered).

GE HyperBand (SMS): simultaneous multi-slice for GE. Supported on specific platforms. Used primarily for brain fMRI and DWI.

Philips

SENSE and GRAPPA are both available. Philips recommends SENSE for most applications. The “SENSE factor” is displayed prominently in the Philips protocol card. Philips provides explicit g-factor warnings in the system messages when the prescribed SENSE factor may produce unacceptably high g.

mSENSE (Philips): a modified SENSE implementation with coil sensitivity estimation from the centre of k-space rather than a separate pre-scan.

Canon

SPEEDER is the Canon parallel imaging implementation (SENSE-based). Acceleration factor is set as “SPEEDER factor” (typically 1.5, 2, 3).

United Imaging

UIH uses GRAPPA-equivalent (labeled as uSENSE or equivalent) with ACS calibration integrated into the acquisition. UIH scanners support both in-plane and 3D dual acceleration.

Hidden coupling — all vendors: when R is increased, the scanner may automatically extend TR or reduce ETL to maintain image quality within SAR limits. At 3T, enabling R=2 often automatically reduces the TR extension that was previously required for SAR management → the TR may actually decrease when R is increased. Technologists must verify TR after changing R.

13. Practical Optimisation Strategies

13.1 Clinical Optimisation Recipes

13.2 The R=2 Standard

For routine clinical MRI, R=2 with GRAPPA is the de facto standard across all body regions and field strengths ≥ 1.5T. This represents the most favourable SNR-efficiency trade-off: - √2 penalty × g ≈ 1.1–1.3 = net SNR ≈ 60–70% of full → adequate for all routine clinical tasks - Time halved → practical protocol duration - Distortion halved in EPI → important at 3T

Increasing to R=3 is justified when temporal resolution (DCE) or scan time (breath-hold) demands it, and the anatomy/coil combination supports adequate g-factor.

14. Parameter Extremes

14.1 R=1 (No Parallel Imaging)

Full k-space sampling; maximum SNR; no g-factor penalty. Used for: - Quantitative T1/T2 mapping where SNR must be maximised - ADNI-equivalent MPRAGE (some institutions use R=1 for maximum volumetric accuracy) - High-resolution 2D SE in targeted small-structure protocols (wrist, inner ear) - Single-channel coil acquisitions (parallel imaging impossible)

At R=1, the scan is the conventional fully-sampled acquisition with the maximum achievable SNR for the protocol parameters.

14.2 Very High R (R=5–8, SMS MB=6–8)

Extremely high in-plane R (R=5–8) or SMS (MB=6–8) is approaching or at the limit of coil encoding capacity: - In-plane R=5–8: requires 32–64 channel coils; g-factor typically > 2 in body regions → SNR per voxel < 25% of full → only acceptable at high field (3T, 7T) with excellent coil arrays and high-SNR acquisitions - SMS MB=6–8: standard for HCP (Human Connectome Project) fMRI at 3T with 64-channel coil; requires careful CAIPIRINHA slice shift to improve z-direction g-factor (see Section 20); standard in research neuroimaging

The practical upper limit for routine clinical imaging: R=3–4 in-plane at 3T; MB=4–6 at 3T for brain EPI. Beyond these values, residual aliasing and g-factor noise become clinically unacceptable for diagnostic imaging.

15. Common Optimisation Errors

16. MRI Technologist Pearls

R=2 is the clinical default — know why: R=2 with GRAPPA represents the optimal practical trade-off for all routine clinical MRI: SNR penalty ≈ 30–35% (g×√2 ≈ 1.41); acquisition time halved; g-factor well within acceptable range for all standard clinical coil/anatomy combinations. Every departure from R=2 (to R=1 for maximum SNR or to R=3 for time pressure) requires justification.

Check the g-factor before high-R protocols: on Siemens scanners, a g-factor pre-scan can be run in < 30 seconds before the diagnostic acquisition. For any novel protocol with R > 2, or any body region with uncertain coil geometry, run this pre-scan. If g_max > 1.5 in the diagnostic region, reduce R.

Phase encoding direction determines which coil elements provide encoding: parallel imaging exploits coil element sensitivity differences in the phase-encoding direction. If the phase direction is A-P and the coil elements are separated A-P → good parallel imaging. If the phase direction is R-L but the coil elements are only separated A-P → poor parallel imaging despite high element count. This is why phase direction choice matters for parallel imaging quality, not just for artefact management.

SMS for brain EPI — verify slice leakage before the study: for SMS-accelerated DWI or fMRI, run a test acquisition and check for inter-slice signal leakage by reviewing the slice quality in the acquisition software. Many platforms provide an SMS quality metric; use it.

DWI at 3T: R=2 is not optional: at 3T, single-shot EPI DWI without parallel imaging produces geometric distortion of 20–40 mm at common EPI readout durations — the image is diagnostically unreliable. R=2 is mandatory for clinical DWI at 3T, reducing distortion by 50%.

Dual acceleration for 3D sequences at 3T: for 3D TSE (SPACE/CUBE), FLAIR, or SPACE at 3T, use R_y=2, R_z=1.5–2 for total effective R=3–4. The dual acceleration is more efficient than increasing R in one direction alone, because the g-factors in two independent directions combine more favourably than a single high-R in one direction.

17. Real Clinical Examples

Example 1: Liver DCE at 3T — R=3 Enabling Temporal Resolution

Clinical scenario: HCC screening with liver DCE-MRI at 3T. Required: three-phase hepatic imaging (arterial, portal venous, delayed) with breath-hold per phase. Arterial phase must be ≤ 18 seconds.

Protocol requirements: 3D VIBE; N_y = 192; N_z = 60 partitions; TR = 4 ms; ETL=1.

At R=1: T_acq = 4 × 192 × 60 / 1000 = 46.1 s → cannot fit in breath-hold; arterial phase not achievable.

At R=2: T_acq = 46.1 / 2 = 23 s → marginal; just barely possible with hyperventilation pre-breath.

At R=3: T_acq = 46.1 / 3 = 15.4 s → fits comfortably in 18-second breath-hold.

SNR impact at R=3 (3T): SNR_R3 = SNR_full / (g × √3) ≈ SNR_full / (1.25 × 1.73) ≈ 46% of full. At 3T, this is diagnostically adequate — the arterial phase HCC detection depends more on temporal adequacy than on absolute SNR.

Diagnosis: the 2.2 cm HCC in segment VII shows arterial enhancement on the correctly timed arterial phase (acquired at 15 s post-injection). This lesion would have been missed on the late arterial phase from R=2 or R=1 (both exceed the 18-s breath-hold limit without patient performance issues).

Example 2: Brain MPRAGE Parallel Imaging for AD Volumetry

Clinical scenario: MPRAGE for hippocampal volumetry; ADNI protocol specifies R=2 (Siemens iPAT=2).

R=1 option: T_acq ≈ 9–12 min; clinically impractical for routine use; some quantitative tool validation datasets used R=1.

R=2 (ADNI-standard): T_acq ≈ 5–6 min; SNR_R2 ≈ 65% of R=1; adequate for FreeSurfer volumetry; validated across ADNI centres.

R=3 option (non-standard): T_acq ≈ 3–4 min; SNR_R3 ≈ 45% of R=1; g-factor increases in temporal lobes; FreeSurfer hippocampal segmentation at R=3 has not been validated against ADNI norms.

Decision: use ADNI-standard R=2. Do not deviate — the hippocampal volume normative database is based on R=2 acquisitions.

Lesson: parallel imaging acceleration factor is part of the quantitative protocol specification, not a free variable. For longitudinal studies, maintain consistent R across all timepoints.

Example 3: DWI at 3T — Distortion Reduction by Parallel Imaging

Clinical scenario: prostate DWI at 3T; PI-RADS v2.1 protocol.

EPI without parallel imaging (R=1): N_y=96, ESP=0.8 ms → EPI readout = 76.8 ms. Geometric distortion in the posterior peripheral zone (adjacent to rectal air-tissue interface): up to 6–8 mm apparent displacement. Posterior peripheral zone lesions appear at incorrect position relative to the T2 anatomy.

EPI with R=2: N_y_acquired=48, readout = 38.4 ms. Distortion: ≈ 3–4 mm → within acceptable tolerance for clinical DWI-T2 co-registration.

EPI with R=2 + reduced-FOV (ZOOMit): FOV_p=80mm, N_y=60, readout = 12 ms. Distortion < 1 mm → essentially distortion-free prostate DWI.

Diagnostic impact: with R=1, a PI-RADS 4 lesion in the posterior peripheral zone at 6 o’clock appears displaced toward the 7 o’clock position on DWI but correctly appears at 6 o’clock on T2 → discordance raises diagnostic uncertainty. With R=2, the DWI and T2 are co-registered within 3 mm → confident co-localisation for PI-RADS assessment.

Example 4: Brain fMRI — SMS Enabling Temporal Resolution

Clinical scenario: resting-state fMRI at 3T; research protocol for default mode network connectivity. Required: whole-brain 60 slices at 2 mm isotropic; TR ≤ 1.5 s.

Without SMS (conventional EPI): 60 slices × 60 ms per slice (TR_per_slice) = 3600 ms minimum TR → TR=3.6 s → temporal resolution insufficient for capturing neural oscillations above ~0.14 Hz.

With MB=4 SMS (GRAPPA R=2 in-plane + MB=4): 60/4 = 15 shot groups × 60 ms = 900 ms per volume → TR=1 s achievable.

g-factor: z-direction SMS g-factor ≈ 1.15–1.25 with 64-ch head coil at MB=4; in-plane R=2 g ≈ 1.15 → total g_effective ≈ 1.35; total SNR = SNR_full / (1.35 × √(2×4)) = SNR_full / (1.35 × 2.83) ≈ 26% of full. At 3T with 64-ch coil: adequate for fMRI (SNR per voxel ~50 at 2 mm → sufficient for robust BOLD detection).

Research impact: TR=1 s captures temporal dynamics up to 0.5 Hz → detects high-frequency BOLD fluctuations → enables more powerful connectivity analysis. This was the enabling technique of the Human Connectome Project [5].

18. Visual Educational Material

18.1 SENSE Un-aliasing Mechanism

FULLY SAMPLED (R=1):                SENSE UNDERSAMPLED (R=2):
k-space: all N_y lines              k-space: every other line (N_y/2)
    │                                   │
    IFT                                 IFT
    │                                   │
Image: clean, full FOV              Image: ALIASED (FOV/2)
    [Anatomy A]                     [A + B superimposed]
    [Anatomy B]                     (two positions folded together)
                                        │
                                    SENSE RECONSTRUCTION
                                    (using coil sensitivity maps)
                                        │
                                    Image: unaliased [A], [B] separated
                                    COST: SNR / (g × √2) per voxel

18.2 g-Factor Map Concept

BODY AXIAL SECTION (12-ch body coil, R=2, A-P phase):

g-factor map (schematic):
    
  LOW g (near coil):    1.1 – 1.2  [adequate SNR]
  MODERATE g (sides):   1.2 – 1.5  [acceptable SNR]
  HIGH g (centre):      1.5 – 2.5  [SNR reduced; may limit diagnosis]

Coil elements:
  ████████████████████████████  ← anterior body coil elements
  
  Patient body:
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░
  ░░░░░░░░  HIGH g  ░░░░░░░░░░  ← centre of body, far from all coil elements
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░
  
  ████████████████████████████  ← posterior body coil elements

→ Hepatic dome (near anterior coil): g ≈ 1.15 → SNR ≈ 60% of full
→ Spinal canal (near posterior coil): g ≈ 1.20 → SNR ≈ 58% of full
→ Central liver (equidistant): g ≈ 1.30 → SNR ≈ 54% of full (adequate)

18.3 SNR vs Acceleration Factor Decision Tree

WHAT IS THE CLINICAL PRIORITY?

├── Maximum image quality (quantitative, research)
│   → R=1 (no acceleration); maximum SNR; maximum scan time
│
├── Standard clinical imaging (routine diagnosis)
│   → R=2; standard SNR (~65%); halved scan time
│   → Default for all routine clinical protocols
│
├── Time-critical (breath-hold DCE, high-resolution 3D)
│   → R=3 (body); requires adequate SNR headroom at field strength
│   → Check: SNR_R3 = SNR_full / (g × √3) ≥ minimum acceptable SNR
│
└── Ultra-high temporal resolution (fMRI, ultrafast DCE)
    → R=2 in-plane + MB=4–6 SMS (brain EPI)
    → Total acceleration 8–12×; requires 32+ channel coil; verify g-factor
    
ALWAYS VERIFY: g_max ≤ 1.5 for routine; g_max ≤ 2.5 for research

19. Evidence Gaps and Ongoing Debate

Optimal ACS line count for GRAPPA at different R values: the number of ACS lines determines the quality of the GRAPPA reconstruction kernel calibration. More ACS lines → better reconstruction but reduced effective acceleration. The optimal ACS count for each R value and coil configuration has not been systematically validated across clinical applications. Vendor defaults (24–32 lines) represent expert consensus without formal comparative evidence.

GRAPPA vs SENSE — clinical equivalence: both GRAPPA and SENSE are in widespread clinical use with equivalent theoretical performance for ideal conditions. In practice, GRAPPA (self-calibrated, motion-robust) has largely displaced SENSE (requires separate calibration scan, sensitive to motion between calibration and acquisition) for routine clinical imaging. Formal comparative clinical studies demonstrating equivalent or superior diagnostic performance have been published for specific applications (brain, body, cardiac) but not comprehensively across all indications.

Combined acceleration (R × MB) optimal limits: the combination of in-plane parallel imaging (R) and SMS multiband (MB) provides total acceleration R × MB. The optimal combination of R and MB for specific applications (fMRI, DWI) — particularly in terms of g-factor distribution and slice leakage — has not been established by formal clinical studies. HCP-style protocols (MB=6–8, R=2) are based on empirical optimisation in research settings.

DLR as a replacement for parallel imaging: DLR can in principle reconstruct images from aggressive k-space undersampling without explicit coil sensitivity calibration, using the trained network’s implicit understanding of image statistics. Whether DLR-based acceleration (without GRAPPA/SENSE) provides equivalent or superior image quality to conventional parallel imaging at equivalent acceleration factors has been studied in limited research settings but is not standardised for clinical use.

Parallel transmit (pTX) and parallel imaging at 7T: at 7T, parallel transmit addresses B1+ inhomogeneity while parallel receive (parallel imaging) addresses encoding efficiency. The combination of pTX and parallel receive places complex demands on both the hardware and reconstruction pipeline. The optimal strategy for 7T clinical imaging — integrating parallel transmit, parallel receive, and SMS — remains an active research topic.

20. Miscellaneous and Future Directions

Historical milestone — SMASH (1997): Sodickson and Manning published the SMASH (SiMultaneous Acquisition of Spatial Harmonics) algorithm in 1997 [2] — the first clinically implemented parallel imaging technique. SMASH demonstrated that 2× acceleration was achievable in cardiac MRI with a phased-array receiver coil, establishing the clinical feasibility of parallel imaging.

CAIPIRINHA (Controlled Aliasing In Parallel Imaging Results IN Higher Acceleration): published by Breuer et al. in 2005 [6], CAIPIRINHA applies additional phase shifts to the simultaneously excited SMS slices, such that the aliasing pattern of the slice images is shifted relative to each other — reducing the g-factor in the slice direction and enabling higher MB factors at acceptable image quality. CAIPIRINHA is now the standard implementation for all high-MB SMS protocols. The key insight: by controlling the aliasing pattern (making it less coherent), the coil elements can more easily separate the simultaneously acquired slices.

Compressed Sensing + Parallel Imaging (PICS): compressed sensing (CS) and parallel imaging exploit different redundancies in MRI data. CS exploits sparsity in a transform domain; parallel imaging exploits coil sensitivity diversity. Combining both (PICS — Parallel Imaging and Compressed Sensing) enables acceleration factors beyond what either method alone can support. Implementations include: 3D CS-SENSE for liver; XD-GRASP for free-breathing body; wave-CAIPI for brain. Total acceleration factors of 10–20× have been demonstrated with acceptable image quality in research settings.

Deep learning parallel imaging reconstruction: neural networks trained on reference fully-sampled MRI datasets have shown promising ability to reconstruct images from aggressively undersampled k-space data — potentially without explicit coil sensitivity calibration, with lower g-factor noise, and with higher acceleration factors than classical GRAPPA/SENSE. E2E-VarNet and similar architectures have been demonstrated at R=4–8 with competitive image quality at 3T.

End of document — Parallel Imaging — MRIninja v1.0 — May 2026 Parent page: MRI Parameters — Overview and Classification (9501)