Acquisition Matrix

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

1. Introduction and General Purpose

The acquisition matrix defines how many discrete data points are collected in each spatial direction during an MRI acquisition. Together with the Field of View (FOV), it is the primary determinant of spatial resolution — the ability to distinguish two adjacent structures as separate entities. The matrix is not a single number but a pair (or triplet in 3D acquisitions): the number of frequency-encoding points (N_x) and the number of phase-encoding steps (N_y), and for 3D acquisitions, the number of slice-encoding partitions (N_z).

The matrix is the critical link between two fundamental MRI concepts: the physical space of the patient (measured in millimetres) and the digital space of the image (measured in pixels). By dividing the FOV into N evenly spaced intervals, the matrix defines the voxel — the smallest individually resolved volume element — and therefore determines both how finely the image can represent anatomical detail and how much signal each voxel contains.

The practical importance of the matrix is immediate and universal: every protocol decision about diagnostic quality ultimately passes through the matrix. A sequence with ideal contrast parameters, optimal field strength, and perfect coil coupling produces a non-diagnostic image if the matrix is too coarse to resolve the target structure. Conversely, a matrix set beyond the SNR capacity of the available hardware produces a noisy, uninterpretable image despite technically correct contrast parameters.

Historical evolution: the first clinical MRI systems of 1980–1983 operated with matrices of 64×64 or 128×128 — sufficient only for gross anatomy. The 256×256 matrix emerged as the clinical standard in the mid-1980s as gradient hardware and digital computing power improved. 512×512 matrices became routine for high-resolution brain and MSK imaging in the 1990s. Modern clinical protocols regularly use 320×320 to 640×640 matrices for targeted high-resolution work, while 3D isotropic acquisitions at 1 mm or sub-millimetre resolution across 256 slices represent the current standard for volumetric imaging.

2. Physical Foundations

The matrix defines the k-space sampling strategy. k-space is the raw data space of MRI — each row of k-space corresponds to one phase-encoding step, and each column corresponds to one frequency-sampling point during the readout. The image is obtained by applying the 2D (or 3D) inverse Fourier transform to the k-space data.

Fundamental relationship: the matrix N_x × N_y defines the number of k-space samples in each dimension. The Fourier transform relationship between k-space and image space means:

• N_x samples in the frequency direction → N_x pixels in the frequency-direction image
• N_y samples in the phase direction → N_y pixels in the phase-direction image

Each pixel in the image corresponds to one spatial frequency component of the object. More pixels (larger matrix) = more spatial frequencies sampled = finer detail resolved.

2.1 Mathematical Foundations

In-plane spatial resolution

The voxel dimensions in the frequency and phase directions are:

Δx = FOV_x / N_x (frequency direction)

Δy = FOV_y / N_y (phase direction)

where FOV is in mm and N is the number of matrix points. The result Δ is in mm/pixel — the spatial resolution in that direction.

Clinical meaning: to resolve two structures separated by distance d, the voxel dimension must satisfy Δ ≤ d/2 (Nyquist criterion in image space). For a 2 mm cartilage layer: Δ ≤ 1 mm. For a plantar plate (2 mm thick): Δ ≤ 0.5–1 mm. For a brachial plexus trunk (3–5 mm diameter): Δ ≤ 1.5 mm.

Optimisation implication: for a fixed FOV, doubling N_x and N_y doubles the spatial resolution in both directions but quadruples the number of voxels — with direct consequences for SNR and acquisition time.

Voxel volume

Voxel_volume = Δx × Δy × Slice_thickness = (FOV_x / N_x) × (FOV_y / N_y) × Δz

SNR implication: SNR scales directly with voxel volume. Doubling both N_x and N_y at fixed FOV reduces Δx and Δy by half → voxel area reduces 4× → SNR reduces 4× (for the same slice thickness and NSA).

Acquisition time and matrix

Acquisition time in 2D TSE is:

T_acq = TR × N_y × NSA / ETL

Only N_y (the phase-encoding matrix) appears here — N_x has no direct time cost in standard 2D acquisitions because all N_x frequency points are collected simultaneously during one readout event. This is the foundational asymmetry of MRI matrix optimisation: increasing the frequency matrix (N_x) costs no time; increasing the phase matrix (N_y) costs time proportionally.

In 3D acquisitions, an additional phase-encoding loop in z adds N_z steps:

T_acq_3D = TR × N_y × N_z × NSA / ETL

Optimisation implication: to maximise spatial resolution within a fixed scan time budget, prioritise increasing N_x over N_y (no time cost), and use ETL/parallel imaging to recover the time cost of necessary N_y increases.

SNR and matrix: the complete relationship

SNR ∝ (FOV_x × FOV_y / (N_x × N_y)) × Δz × √(NSA) × B₀ × Coil_factor / √BW

This can be rewritten as:

SNR ∝ Voxel_volume × √(NSA) × B₀ × Coil / √BW

where Voxel_volume = (FOV_x/N_x) × (FOV_y/N_y) × Δz.

Key insight: a 512×512 matrix acquisition at a given FOV has 4× smaller voxels than a 256×256 acquisition → 4× lower SNR (if all other parameters are equal). To recover this SNR deficit requires 16× more NSA (since SNR ∝ √NSA → 4× SNR recovery requires NSA = 16) — a completely impractical solution. In practice, the SNR deficit is accepted as the necessary cost of higher resolution, compensated partially by field strength, coil optimisation, reduced bandwidth, and increasingly by DLR.

k-space density and image contrast

The matrix also determines the spatial frequency content of the image. In TSE/FSE sequences, different echoes in the echo train fill different rows of k-space:

• Central k-space (near k_y = 0): low spatial frequencies; determines overall image contrast and bulk signal levels. The echo filling k₀ defines the effective TE and therefore the primary image contrast.
• Peripheral k-space (high k_y values): high spatial frequencies; determines fine structural detail, edges, small lesion margins.

A larger matrix extends further into peripheral k-space → more fine-detail information. But the peripheral k-space echoes are acquired at later echo times in the train → more T2 weighting → T2-dependent blurring of fine detail (see ETL interaction in Section 5.1).

3. Units, Terminology and Vendor Nomenclature

The matrix is dimensionless (a count of pixels), but is always expressed as the product of two integers representing the number of data points in each encoding direction.

Standard notation: N_x × N_y (e.g., 320 × 256), where N_x = frequency matrix and N_y = phase matrix. In 3D: N_x × N_y × N_z.

Critical distinction — acquired vs reconstructed (display) matrix:

• Acquired matrix: the number of k-space lines actually measured. This is the true matrix for SNR and resolution calculations.
• Reconstructed (display) matrix: the matrix of the image shown on the console, which may be zero-filled (interpolated) to a larger value. A GE acquisition at 256×192 with ZIP2 produces a 512×384 display image — the spatial resolution is still 256×192; only the display is interpolated.

Always confirm the acquired matrix when documenting spatial resolution. The displayed matrix is misleading if zero-filling has been applied.

4. Typical Value Ranges

4.1 Matrix by Anatomical Region and Field Strength

4.2 Matrix vs Field Strength Context

5. Parameter Interaction Ecosystem

5.1 Parameter Relationships Matrix

The N_x vs N_y asymmetry is the single most important operational fact about matrix in clinical MRI optimisation:

6. Effects on Image Appearance

6.1 Increasing Matrix (N_y and/or N_x)

• Spatial resolution: directly improved. Fine structures — cartilage tears, nerve fibres, small vessel anatomy — become separately resolved rather than averaged within a voxel.
• Voxel volume: decreases → less partial volume averaging → more accurate signal representation of small structures; apparent contrast between adjacent small structures improves.
• SNR: decreases (inversely proportional to voxel volume). At excessively high matrix for the available SNR, image appears grainy/noisy — loss of image texture that the visual system uses for tissue discrimination.
• T2 blurring in TSE: increases slightly as peripheral k-space lines (filled by later echoes) extend to higher spatial frequencies. Short-T2 structures (cartilage, tendons) appear progressively blurred at very high matrix with long ETL.
• Gibbs ringing: all Fourier-reconstructed images show truncation artefact (Gibbs ringing) at sharp interfaces (brain–CSF, bone–soft tissue). A larger matrix (more k-space samples → less truncation) reduces Gibbs ringing amplitude. Low matrix → prominent Gibbs ringing.
• Edge sharpness: improved. Lesion margins appear sharper; cortico-medullary distinction is more precise.

6.2 Decreasing Matrix

• Spatial resolution: decreases. Structures smaller than the voxel are not resolved individually → partial volume averaging → apparent signal reduction and margin blurring.
• SNR: increases (larger voxel). Image appears smoother and less noisy.
• Acquisition time: decreases (for N_y reduction).
• Gibbs ringing: increases. The brain–CSF interface at low matrix commonly shows prominent bright/dark parallel bands adjacent to the interface — a diagnostic pitfall if the bands are interpreted as cortical signal change.
• Partial volume artefact: worsens. A 3 mm lesion at 5 mm voxel resolution appears as a faint signal change averaged with surrounding tissue; the same lesion at 1 mm resolution appears as a clearly defined focus.

7. Effects on Acquisition Time

The matrix has the most direct and operationally important time relationship of all MRI acquisition parameters. The relationship is explicit:

T_acq = TR × N_y × NSA / ETL (2D TSE)

T_acq_3D = TR × N_y × N_z × NSA / ETL (3D TSE)

7.1 Direct Effects

Every additional phase-encoding step (N_y increment) adds exactly one TR period to the acquisition time. This is the most predictable time cost in MRI:

ΔT = TR × ΔNSN_y × NSA / ETL

Example: T2 TSE brain, TR=5000 ms, ETL=16, NSA=1:

• N_y = 256 → T_acq = 5000 × 256 / 16 = 80,000 ms ≈ 1.33 min
• N_y = 384 → T_acq = 5000 × 384 / 16 = 120,000 ms = 2.0 min (+50%)
• N_y = 512 → T_acq = 5000 × 512 / 16 = 160,000 ms = 2.67 min (+100%)

7.2 Indirect Effects

The time cost of increasing N_y can be partially or fully recovered by:

• Increasing ETL: doubling ETL halves the time for the same N_y. Cost: T2 blurring.
• Increasing parallel imaging R: doubling R halves the acquired k-space lines (but not the reconstructed N_y). Cost: SNR × 1/(g×√R).
• Reducing TR: only for GRE sequences (where TR is the primary time variable). Cost: less T1 recovery → T1 contrast change.
• Partial Fourier: acquiring 5/8 or 6/8 of N_y lines. Cost: Gibbs ringing, minor SNR reduction.

7.3 N_x Has No Direct Time Cost

The frequency matrix N_x is collected simultaneously during each readout gradient — all N_x points are sampled in a single readout period. Doubling N_x from 256 to 512 adds zero TR periods. The only time-related effect of increasing N_x is that it may require a wider readout bandwidth to collect more points in the same time → minor SNR reduction.

Practical rule: always maximise N_x first (to the scanner maximum, typically 512 or 640 for high-resolution work) before considering N_y increases. The N_x resolution improvement is essentially free.

8. Effects on SNR and CNR

8.1 SNR and Matrix

SNR ∝ Voxel_volume = (FOV_x/N_x) × (FOV_y/N_y) × Δz

The SNR penalty from high matrix is the largest single image quality cost in high-resolution MRI:

Critical observation: at 3T with a dedicated surface coil, the SNR headroom may be sufficient to support 512×512 at 3 mm slice thickness for brain or knee imaging with acceptable residual SNR. At 1.5T with a body coil for abdominal imaging, 256×192 with 5 mm slices may be the maximum before SNR becomes inadequate.

8.2 CNR and Matrix

CNR = (S_tissue1 − S_tissue2) / σ_noise = (S_tissue1 − S_tissue2) / (SNR × S_tissue_ref)

At high matrix (low SNR), CNR decreases even if the tissue contrast (S_tissue1 − S_tissue2) is unchanged. For small lesions where the lesion signal difference from background is small, this CNR reduction may take the lesion below the perceptual threshold despite technically adequate spatial resolution.

The resolution-CNR trade-off for small lesions: a 3 mm lesion requires at least 1.5 mm voxels to be resolved. But if 1.5 mm voxels produce SNR too low for adequate CNR (noisy image), the lesion may still be missed despite being technically resolved. The optimal matrix for lesion detection balances spatial resolution against CNR — not simply the maximum achievable resolution.

8.3 Field-Strength Dependency

Higher field strength directly relaxes the matrix-SNR trade-off:

9. Artefacts Associated with the Matrix

10. Behaviour Across Sequence Families

Spin Echo (SE)

Standard SE: T_acq = TR × N_y × NSA. Since there is no ETL (one echo per TR), the time cost of increasing N_y is proportionally the highest of any sequence. SE is therefore the sequence family most severely constrained by matrix size — at TR=2000 ms, N_y=512, NSA=1: scan time = 17 min. Standard SE is rarely used above N_y=256 for this reason.

Turbo Spin Echo (TSE/FSE)

TSE is the workhorse of clinical MRI precisely because ETL divides the N_y time cost. At ETL=16, N_y=512 requires only 512/16 = 32 TR intervals. TSE enables high-matrix acquisitions (512–640) that would be impractical with standard SE. The T2 blurring at the high matrix end is the primary quality cost.

Gradient Echo (GRE/FLASH/SPGR)

At short TR (3–10 ms, typical for DCE VIBE), N_y directly multiplies the very short TR. For N_y=224 at TR=5 ms: 1120 ms = ~1 sec per volume. For breath-hold DCE, increasing N_y significantly extends the breath-hold duration. In GRE, increasing N_x (frequency matrix) is universally used to improve resolution without time cost — 512 frequency matrix is routine in GRE protocols.

Inversion Recovery (STIR, FLAIR, MPRAGE)

STIR and FLAIR: same time equation as TSE (same underlying readout). In MPRAGE (3D GRE): T_acq = TR_GRE × N_y × N_z — the matrix is 3D and all three dimensions contribute to acquisition time. ADNI MPRAGE at 1 mm isotropic: N_y=256, N_z=192, TR=2300 ms → effective: each partition group adds ~5.5 min; full 3D: 4–6 min.

EPI (DWI, DSC, fMRI)

EPI acquires all N_y phase-encoding steps in a single shot using a rapidly switching readout gradient (oscillating gradient). The entire k-space is filled in one TR. This makes EPI independent of N_y for TR-based time — but introduces a critical constraint: the readout duration for all N_y steps is fixed (no additional TRs). Longer N_y in EPI means longer readout → more geometric distortion (phase BW ∝ 1/(N_y × ESP)) → more EPI-specific artefact. Clinical EPI is constrained to N_y ≈ 64–160. Standard DWI: 96–128 phase steps; fMRI: 64–80 per slice at 2–3 mm resolution.

Dixon

The underlying sequence family (GRE or TSE) determines the matrix time relationships. The Dixon fat-water separation adds a specific TE constraint (IP/OP TE values) but does not independently constrain the matrix.

DCE

As for GRE above; the key DCE matrix constraint is temporal resolution (the scan must complete in ≤ 60–90 seconds per phase for standard protocols). The matrix × TR × N_z product must fit within this window with the available parallel imaging and CS acceleration.

ASL (pCASL)

3D pCASL uses a 3D readout (GRASE or spiral EPI). The matrix determines the spatial resolution of the perfusion map. Typical brain ASL: 64×64 in-plane (3–4 mm voxels) — much coarser than anatomical images, reflecting the intrinsically low SNR of ASL (signal difference ≈ 1% of total signal). At small matrix (32×32), perfusion maps become unreliably pixellated; at large matrix (128×128), SNR drops below the perfusion signal threshold.

bSSFP (TrueFISP/FIESTA/b-FFE)

Very short TR (3–5 ms per TR). Matrix requirements are similar to GRE. For cardiac cine bSSFP: N_y per cardiac phase × number of phases × TR must fit within the RR interval. Typical cardiac cine: 160–192 phase-encoding lines per cardiac cycle (with segmented k-space acquisition over multiple heartbeats).

Spectroscopy (SVS/MRSI)

In MRSI (chemical shift imaging), the spatial encoding is performed by phase-encoding gradients, exactly as in imaging. The MRSI matrix defines the voxel size of the spectral map: an 8×8 MRSI grid at 240 mm FOV → 30 mm voxels — much coarser than imaging. SNR per spectral voxel scales as voxel_volume, and at 30 mm each spectral voxel is 27,000 mm³ — still only barely adequate for clinical 1H spectroscopy. High-resolution MRSI (16×16 or 32×32) is feasible only at high field (3T+) or with long NSA.

11. Field Strength Behaviour

3T matrix advantage: the higher SNR at 3T is most productively spent on matrix (resolution) rather than NSA. The standard optimisation strategy at 3T is: use the ~2× SNR gain to double the matrix compared with the 1.5T protocol (voxel area halved → SNR halved → net SNR approximately equal to 1.5T at the lower matrix). This produces higher-resolution images at equivalent SNR — the primary clinical benefit of 3T for MSK, brain, and prostate MRI.

12. Vendor-Specific Implementation

Siemens

Siemens separates "Base resolution" (N_x) and "Phase resolution" (expressed as a % of Base resolution, e.g., "Phase resolution 75%" = N_y = 0.75 × N_x). The "Phase oversampling" adds extra steps beyond the phase FOV. Siemens also provides the "Interpolation" toggle (zero-filling to double matrix in each direction for display). The Inline reconstruction at the scanner always generates the interpolated display matrix; the acquired matrix is documented in the DICOM header under the "Acquisition Matrix" field.

A critical Siemens-specific behaviour: increasing N_x ("Base resolution") beyond 512 may automatically increase the minimum bandwidth to accommodate the higher readout sampling rate → the technologist may see unexpected BW changes when pushing N_x. This is the frequency matrix → TE_min coupling described in Section 5.

GE

GE expresses frequency and phase matrix separately. ZIP2 (zero-filling × 2 in both directions) and ZIP512 (zero-filling to 512 in the frequency direction) are post-acquisition interpolation options that increase display matrix without increasing acquired resolution. These are applied by default in many GE protocols — a common source of confusion between displayed and acquired resolution. The "Freq. Matrix" and "Phase Matrix" fields in the GE protocol record the acquired (not display) values.

Philips

Philips uses "Scan matrix" for the acquired matrix and "Reconstruction matrix" for the display matrix (which may be zero-filled). The Philips system is notable for displaying the actual voxel size in mm directly in the protocol design interface — making resolution verification straightforward. "Half scan" (Philips equivalent of partial Fourier) reduces the acquired N_y.

Canon

Canon uses "Matrix" for the frequency direction and "Phase matrix" for the phase direction. The "Recon matrix" is the interpolated display matrix. Canon's AiCE deep learning reconstruction is applied post-reconstruction and can be configured for different aggressiveness levels (mild, moderate, strong denoising) — the higher levels enable more aggressive matrix reduction at acquisition while maintaining apparent image quality.

United Imaging

UIH uses standard international terminology consistent with Siemens conventions. Their uMR scanner displays the effective voxel size on the protocol card, facilitating real-time resolution verification. uAI reconstruction includes deep learning denoising that supports higher matrix acquisitions at 0.55T than would otherwise be feasible.

Hidden parameter coupling — all vendors: on all platforms, the phase matrix (N_y) is coupled to the minimum scan time via TR and ETL. When the technologist increases N_y, the scanner recalculates scan time immediately. However, some vendors (particularly GE and Philips) automatically adjust the ETL or parallel imaging factor when N_y is increased beyond a threshold, to maintain the scan time within a pre-set limit. The technologist must verify that the auto-adjusted ETL has not introduced unacceptable T2 blurring.

13. Practical Optimisation Strategies

13.1 Clinical Optimisation Recipes

14. Parameter Extremes

14.1 Extremely Low Matrix (N_y < 96)

At very low matrix, the voxel volume is large → excellent SNR but severe spatial resolution loss. Used for:

• EPI-based acquisitions (DWI, DSC, fMRI) where the EPI readout limits N_y
• ASL perfusion (intrinsically low-SNR technique requires large voxels)
• Rapid survey scans (localisers, whole-body HASTE)
• Patient who cannot cooperate with longer acquisitions

Pitfalls: Gibbs ringing dominates at all interfaces; small lesions invisible; structure boundaries grossly misrepresented. A 64×64 matrix at 250 mm FOV produces 3.9 mm voxels — acceptable for functional imaging but inadequate for any structural diagnostic task.

14.2 Extremely High Matrix (N_y > 512)

At very high matrix, SNR becomes the limiting factor. The voxel is smaller than the intrinsic noise floor can support. Used for:

• Research 7T neuroimaging (cortical layer imaging at 0.5 mm isotropic)
• High-resolution inner ear MRI (0.5–0.7 mm) with dedicated coils
• Dental MRI
• Histological-equivalent MRI in excised specimens (outside the bore)

Clinical limit: at 3T with a 32-channel head coil, the practical upper N_y limit for routine brain T2 TSE is approximately 512 at acceptable SNR (assuming 3 mm slice, NSA=1, ETL=16). Beyond this, DLR is required to maintain diagnostic SNR.

15. Common Optimisation Errors

16. MRI Technologist Pearls

The N_x/N_y asymmetry is your most powerful optimisation tool: before accepting a long scan time, check whether N_x is already at its maximum. If N_x = 256 and N_y = 256, doubling N_x to 512 (a one-click change on most consoles) improves frequency-direction resolution by 2× at zero time cost. Most protocols run with N_x = N_y (symmetric matrix) when asymmetric optimisation (high N_x, moderate N_y) would be superior and faster.

Verify the actual voxel size before starting: all consoles show the resulting voxel dimensions when you set FOV and matrix. Read these numbers and verify they meet the protocol specification. This habit prevents the most common resolution errors in clinical practice.

Check scan time immediately after any N_y change: the time impact of N_y changes is instantaneous and displayed on the console. A 20-second check after changing N_y prevents the discovery of a 12-minute scan that was supposed to be 4 minutes.

For breath-hold body MRI, work backwards from the time budget: decide the maximum acceptable breath-hold (e.g., 18 seconds), calculate the maximum N_y given TR and ETL, then check whether this N_y provides adequate resolution at the chosen FOV. If not, increase ETL or parallel imaging R before asking the patient to hold longer.

Gibbs ringing at spine: if the coronal or sagittal T2 TSE of the spine shows prominent bright/dark bands parallel to the spinal cord surface, the N_y is insufficient (Gibbs ringing). Increasing N_y by 50–100 steps (with ETL increase to maintain time) resolves this — the radiologist will stop getting calls about "possible cord signal change."

Protocol rescue when scan is too long: if a protocol has been acquired with insufficient matrix and the radiologist requests higher resolution: (a) increase N_x first (no time cost) and re-run; (b) if phase resolution must also be increased, apply higher ETL or higher R; (c) if the patient has already left, note the matrix limitation in the report and request a follow-up targeted acquisition.

Document the acquired matrix in the report for quantitative protocols: for AD volumetry (MPRAGE), tumour response assessment, and cartilage grading, the acquired matrix must match the prior examination. Document it explicitly in the technique section.

17. Real Clinical Examples

Example 1: Knee Cartilage Assessment at 1.5T vs 3T

Clinical scenario: early osteoarthritis; suspected focal cartilage defect at the medial femoral condyle. Cartilage is 2–2.5 mm thick at this location.

1.5T protocol: FOV 160 × 130 mm; matrix 320 × 256; Δx × Δy = 0.5 × 0.5 mm; slice 3 mm; TR=3500/TE=35/ETL=12; NSA=1 → T_acq = 3.1 min.

3T protocol (same time budget): FOV 160 × 130 mm; matrix 512 × 384; Δx × Δy = 0.31 × 0.34 mm; slice 2 mm; TR=3500/TE=35/ETL=20; R=2; NSA=1 → T_acq ≈ 3.3 min.

Impact on diagnosis: at 0.5 mm, the cartilage layer is resolved in 5 pixels (2.5 mm / 0.5 mm). A 0.5 mm focal defect occupies 1 pixel — at the resolution limit. At 0.31 mm, the same layer is resolved in 8 pixels; a 0.5 mm defect occupies 1.6 pixels — approaching reliable detection threshold.

Trade-off: 3T protocol requires higher ETL (20 vs 12) → more T2 blurring of cartilage fine structure. Net resolution gain from higher matrix partially offset by T2 blurring. Optimised VFA ETL (SPACE/CUBE) at 3T can address this at the cost of longer scan time.

Example 2: Brain T2 TSE — Gibbs Ringing at Insufficient Phase Matrix

Clinical scenario: patient referred for "possible cervical syrinx" seen on an outside MRI. The report describes a bright linear structure in the cervical cord.

Outside MRI: T2 TSE sagittal; N_y = 160; FOV 240 mm; Δy = 1.5 mm/pixel; slice 3 mm. The cord–CSF interface shows prominent Gibbs ringing bands (CSF signal oscillating into the cord → apparent bright intramedullary signal).

Re-examination: increase N_y to 320 (same ETL → 2× time) or maintain time with ETL increase. Δy = 0.75 mm; Gibbs ringing amplitude reduced 2×.

Result: the apparent "syrinx" disappears on the higher-matrix acquisition — confirmed as Gibbs truncation artefact from the low-matrix outside study.

Lesson: N_y < 192 for sagittal spine T2 routinely produces Gibbs ringing bands at the cord surface that simulate intramedullary pathology. Minimum N_y for reliable cord assessment: 256–320 in the phase direction (S-I for sagittal spine).

Example 3: Prostate T2 TSE — PI-RADS Matrix Requirements

Clinical scenario: elevated PSA; prostate MRI for suspicious nodule detection.

PI-RADS v2.1 requirement: T2 in-plane resolution ≤ 0.7 × 0.7 mm; slice thickness ≤ 3 mm [1].

Protocol design: FOV 200 × 180 mm; matrix 320 × 288 → Δx × Δy = 0.625 × 0.625 mm; slice 3 mm; TR=4500/TE=90/ETL=14; R=2; NSA=1 → T_acq ≈ 4.6 min.

Alternative (inadequate): FOV 200 × 180 mm; matrix 256 × 192 → Δx × Δy = 0.78 × 0.94 mm; exceeds PI-RADS v2.1 specification → non-compliant; potentially non-diagnostic for small peripheral zone lesions.

Impact: at 0.625 mm, a 5 mm anterior fibromuscular stroma lesion occupies 8 pixels in both dimensions — adequate for PI-RADS morphological assessment. At 0.94 mm phase resolution, the same lesion occupies 5 pixels — marginal for shape and margin assessment.

Lesson: PI-RADS specifies minimum matrix requirements that are not arbitrary. These requirements represent the minimum spatial resolution for the diagnostic task defined in the guidelines.

Example 4: Whole-Body MRI Myeloma — Matrix-SNR-Time Constraint

Clinical scenario: multiple myeloma staging; WB-STIR coronal + WB-DWI required per IMWG protocol.

STIR coronal per station: FOV 420 × 350 mm; matrix 320 × 260 → Δx × Δy = 1.3 × 1.3 mm; slice 4 mm; TI=160 ms (1.5T); TR=3500/ETL=20; NSA=1 → T_acq per station ≈ 3.3 min × 4 stations = 13.2 min.

DWI per station: FOV 420 × 380 mm; matrix 128 × 96 → Δx × Δy = 3.3 × 4.0 mm; b=50/900; free-breathing NSA=6; TR=6000/ETL=1 (EPI) → per station ≈ 4.8 min × 4 stations = 19.2 min.

Total WB-MRI time: ≈ 45–55 min.

Matrix consideration: the WB-DWI matrix (128 × 96) reflects the fundamental EPI constraint — EPI is SNR/distortion-limited to ≤ 128 phase steps. Focal myeloma lesions ≥ 10 mm are reliably detected at this resolution. Smaller lesions (< 5 mm) are at or below the detection threshold — an accepted limitation of WB-DWI documented in IMWG guidelines. Increasing the DWI matrix would require either longer echo train (→ more distortion) or higher parallel imaging R (→ SNR reduction below the already-marginal DWI SNR).

18. Visual Educational Material

18.1 Matrix, FOV, and Voxel Size Relationship

SAME FOV (200 mm × 200 mm), DIFFERENT MATRIX:

Matrix 128 × 128:
  Δx = Δy = 200/128 = 1.56 mm/pixel
  Voxel area = 2.44 mm²  [large; high SNR; coarse detail]

Matrix 256 × 256:
  Δx = Δy = 200/256 = 0.78 mm/pixel
  Voxel area = 0.61 mm²  [moderate; standard clinical]

Matrix 512 × 512:
  Δx = Δy = 200/512 = 0.39 mm/pixel
  Voxel area = 0.15 mm²  [small; low SNR; high detail]

SNR ratio: 128 : 256 : 512 = 4 : 1 : 0.25
   (SNR ∝ voxel area for same slice thickness and NSA)

18.2 N_x vs N_y: The Asymmetric Time-Cost Decision Tree

RESOLUTION IMPROVEMENT NEEDED?
│
├── IN FREQUENCY DIRECTION (N_x)?
│   └── Increase N_x
│       ├── Time cost:  ZERO
│       ├── SNR cost:   minor (wider BW may be needed)
│       └── DO THIS FIRST
│
└── IN PHASE DIRECTION (N_y)?
    └── Increase N_y
        ├── Time cost:  PROPORTIONAL (ΔT = TR × ΔN_y / ETL)
        ├── SNR cost:   proportional (smaller voxel)
        └── RECOVERY OPTIONS:
            ├── Increase ETL (time recovery; T2 blurring cost)
            ├── Increase R (time recovery; g-factor SNR cost)
            ├── Reduce TR (contrast change; only for GRE)
            ├── Partial Fourier (minor time/SNR recovery)
            └── Enable DLR (SNR recovery; texture change)

18.3 Matrix Optimisation for Specific Diagnostic Targets

TARGET STRUCTURE SIZE → REQUIRED VOXEL ≤ size/2 → REQUIRED MATRIX
                                      
Plantar plate (2 mm)   → Δ ≤ 1.0 mm  → N = FOV/1.0 mm
                                         e.g. FOV 80mm → N ≥ 80
Cartilage (2.5 mm)     → Δ ≤ 1.2 mm  → N = FOV/1.2 mm
                                         e.g. FOV 160mm → N ≥ 133
Nerve trunk (4 mm)     → Δ ≤ 2.0 mm  → N = FOV/2.0 mm
                                         e.g. FOV 300mm → N ≥ 150
Prostate lesion (5mm)  → Δ ≤ 2.5 mm  → N = FOV/2.5 mm
                                         e.g. FOV 200mm → N ≥ 80
Focal myeloma (10mm)   → Δ ≤ 5.0 mm  → N = FOV/5.0 mm
                                         e.g. FOV 420mm → N ≥ 84

19. Evidence Gaps and Ongoing Debate

Minimum matrix for PI-RADS compliance: the PI-RADS v2.1 guidelines specify T2 in-plane resolution ≤ 0.7 mm and DWI in-plane ≤ 2.5 mm [1]. These thresholds are based on expert consensus in the guideline development process, not on formal prospective trials comparing diagnostic accuracy at multiple matrix values. Whether the 0.7 mm threshold is the true inflection point for PI-RADS lesion characterisation accuracy, or whether slightly coarser resolution (e.g., 0.9 mm) is clinically equivalent, has not been formally tested.

DLR-enabled matrix expansion: deep learning reconstruction enables acquisition at lower matrix (lower SNR) with effective image quality approaching higher-matrix acquisitions. The degree to which DLR-recovered "virtual high matrix" images are diagnostically equivalent to truly acquired high-matrix images has been demonstrated for some applications (brain, body) but not comprehensively across all diagnostic tasks or all DLR aggressiveness levels.

Optimal N_x:N_y ratio (asymmetric matrix): the optimal asymmetric matrix for any given anatomy (how much to prioritise frequency vs phase resolution) has not been formally studied. The common clinical practice of using square matrices (N_x = N_y) may be suboptimal for most applications — for example, for sagittal spine MRI where the SI extent is much greater than the AP extent, a larger N_y (SI direction) and smaller N_x (AP direction, if AP is the phase direction) might be more efficient.

Matrix and AI-assisted lesion detection: AI-based lesion detection algorithms have been validated predominantly on images acquired at the matrix and resolution standards prevailing at the time of training-set acquisition. The sensitivity of AI detection tools for lesions at the spatial resolution limit (e.g., prostate lesions at N_y=160 vs N_y=320) has not been systematically characterised. As AI tools are deployed clinically, the matrix requirements to maintain AI performance may differ from the requirements for human reader performance.

20. Miscellaneous and Future Directions

Historical context: the progression of clinical MRI matrix capability tracks directly with gradient hardware development (higher slew rates → shorter gradient switching → shorter echo spacing → shorter readout for given N_x) and digital computing advances (larger memory for k-space storage, faster Fourier transform computation). The first 512×512 clinical brain T2 images appeared around 1989–1991 as dedicated neuro scanners with high-performance gradients became available. The 1024×1024 matrix (0.2–0.3 mm in-plane at standard FOV) remains largely a research tool today but is increasingly feasible at 3T with 32+ channel coils.

Beyond Cartesian matrix — non-Cartesian k-space: all of the above assumes Cartesian k-space filling (sequential phase-encoding steps on a regular grid). Non-Cartesian trajectories (radial, spiral, PROPELLER/BLADE) fill k-space differently and their resolution behaviour does not follow the simple N × Δk = FOV equation. Radial acquisitions oversample central k-space (better low-frequency/contrast behaviour) and undersample peripheral k-space (lower effective resolution per acquisition). The "matrix" concept for these sequences requires more nuanced interpretation.

Simultaneous multi-slice (SMS/multiband) and effective matrix: SMS techniques increase temporal efficiency (more slices per TR) without changing the in-plane matrix. For fMRI, the combination of SMS (factor 6–8) and in-plane acceleration (R=2–3) has enabled whole-brain acquisitions at 2 mm isotropic resolution in 1 second — a 10× increase in temporal efficiency over 2010-era standards, enabling high-resolution resting-state fMRI that was previously impractical.

Sub-voxel resolution via super-resolution reconstruction: multiple low-resolution images acquired with deliberate slight positional offsets can be combined using super-resolution algorithms to produce an effective resolution higher than any single acquisition. This approach (used in fetal MRI to achieve 0.5 mm effective resolution from multiple 2 mm acquisitions) is being explored for adult body imaging and may eventually relax the fundamental matrix-SNR trade-off at the protocol design level.

Related parameter deep dive: Parallel Imaging explains acceleration factor, g-factor, SENSE/GRAPPA reconstruction and SMS/Multiband trade-offs relevant to this parameter.

End of document — Acquisition Matrix — MRIninja v1.0 — May 2026

Parent page: MRI Parameters — Overview and Classification (9501)