Reconstruction Matrix, Pixel Interpolation, and Slice Interpolation

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 · Slice Thickness · 2D vs 3D Acquisition Version 1.0 — May 2026

1. Introduction and General Purpose

The reconstruction matrix and associated interpolation techniques define the relationship between the acquired k-space data and the displayed image. While the acquisition matrix (N_x × N_y, see Acquisition Matrix child page) determines the true spatial resolution of the MRI data, the reconstruction matrix determines the number of pixels in the final displayed image — which may be larger than the acquisition matrix through a process called zero-filling or interpolation.

This distinction is clinically and technically fundamental: the reconstruction matrix changes what the radiologist sees on the workstation without changing what the scanner physically measured. Understanding this difference is essential for:

• Correctly reporting spatial resolution (using acquisition matrix, not display matrix)
• Interpreting whether apparent edge sharpness reflects true resolution or computational interpolation
• Calibrating expectations about lesion detail at different interpolation settings
• Avoiding false confidence from upsampled images that appear sharper but contain no additional information

The topic encompasses three related but distinct processes:

1. In-plane pixel interpolation: zero-filling or mathematical interpolation to display more pixels than were acquired (e.g., 256×256 acquired → 512×512 displayed)
2. Slice interpolation (z-direction): inserting synthetic slices between acquired slices to produce a smoother apparent z-coverage with reduced visible inter-slice discontinuities
3. 3D isotropic reconstruction: in 3D acquisitions, reformatting the isotropic dataset into arbitrary planes — a true resolution operation, not an interpolation artefact

Each of these processes has different physical bases, different image quality implications, and different diagnostic consequences. This page addresses all three within a unified framework.

2. Physical Foundations

2.1 The Relationship Between k-Space and Image Space

Every MRI image is produced by applying the inverse Fourier transform (IFT) to the acquired k-space data. The spatial resolution of the result is determined by the extent of k-space sampled:

Δx = FOV_x / N_x (frequency direction) Δy = FOV_y / N_y (phase direction)

The IFT maps N_x × N_y k-space points to exactly N_x × N_y image pixels. This is the true or acquired image — it contains all the information present in the k-space data, at the resolution defined by the acquisition matrix.

2.2 Mathematical Foundations

2.2.1 Zero-Filling (Zero-Padding) — In-Plane

Zero-filling is the process of appending zeros to the k-space data beyond the acquired boundary before applying the Fourier transform. If the acquired matrix is N_x × N_y and zeros are added to extend to 2N_x × 2N_y:

IFT(k-space_zero-padded_2N×2N) → image_2N×2N

The result is a 2N × 2N pixel image from N × N acquired data points.

What zero-filling does: the inverse Fourier transform of zero-padded k-space is equivalent to sinc interpolation of the original image. Sinc interpolation is the mathematically optimal interpolation method for band-limited signals (which MRI images are, by definition). The sinc kernel fills in sub-pixel detail between the original pixels using the available spectral information.

What zero-filling does NOT do: it does not add new spatial frequencies to the image. The highest spatial frequency present in the data remains k_max = N/(2 × FOV) — determined by the acquisition matrix. Zero-filling cannot recover spatial frequencies that were not acquired. The actual spatial resolution (ability to resolve two separate structures) is unchanged.

Mathematically:

Acquired image pixel value at position (i, j): f(i, j) = Σ_m Σ_n F(m, n) × e^(2πi(im/N_x + jn/N_y))

Zero-filled image (2N pixels): f_ZF(i, j) = Σ_m Σ_n F(m, n) × e^(2πi(im/(2N_x) + jn/(2N_y)))

where F(m,n) = k-space values, and the summation for ZF includes m = 0 (zero-filled entries). The result: the ZF image at 2N pixels samples the continuous underlying function at twice the density — revealing the same information at finer pixel intervals, with reduced Gibbs ringing at high-contrast boundaries.

Clinical interpretation: zero-filling produces a smoother-appearing image with less pixelated edges. The anatomical information is identical to the non-zero-filled image; only the visual presentation changes. A small lesion that was "on the edge" of being visible does not become more or less detectable; the resolution limit is set by the acquisition matrix.

SNR effect of zero-filling: the zero-filled image has more pixels but the same total signal. Each zero-filled pixel has lower signal than a non-interpolated pixel would have (the signal is distributed across more pixels). In practice, the SNR per pixel in a zero-filled image is lower than in the original N×N image by √(ZF_factor). For ×2 zero-filling (N→2N in both directions): SNR per pixel decreases by 1/√4 = 50%. However, the SNR per unit area (the diagnostically relevant quantity) is unchanged — the same total signal is present; it is just distributed over more pixels.

2.2.2 Gibbs Ringing Reduction by Zero-Filling

One genuine benefit of zero-filling beyond the cosmetic improvement is partial reduction of Gibbs ringing (truncation artefact). At sharp signal boundaries (brain–CSF interface, cord–CSF interface), the abrupt transition in signal creates high-frequency spectral content that, when truncated at k_max, produces oscillating overshoot and undershoot (the Gibbs phenomenon). By extending the k-space representation (zero-filled beyond k_max), the Fourier synthesis produces a smoother representation of the sharp boundary with reduced ringing amplitude.

Quantitatively: the Gibbs ringing amplitude with N-point DFT is approximately 9% of the step height (the Gibbs constant ≈ 1.089). With 2N zero-filling, the ringing amplitude is approximately 4.5% — reduced but not eliminated. With infinite zero-filling, the ringing would approach the limiting Gibbs constant of 8.9%. Zero-filling is thus a genuine (though incomplete) Gibbs ringing reduction strategy.

2.2.3 Slice Interpolation (Slice Profile Interpolation)

In 2D multi-slice acquisitions, the acquired slices are discrete measurements separated by the nominal slice thickness (plus any gap). Slice interpolation creates synthetic "in-between" slices by mathematical interpolation between adjacent acquired slices.

Unlike in-plane zero-filling (which has a solid physical basis in Fourier interpolation), slice interpolation in 2D acquisitions is a purely spatial interpolation with no physical basis in the slice-selection direction:

The slice direction in 2D is not Fourier-encoded — there is no k-space in z. The RF pulse selects the slice; the signal within each slice is a through-plane average over the slice thickness. Interpolating between slices therefore does not recover any z-frequency information. It simply creates smooth transitions between adjacent slices in MPR reconstructions.

Methods:

• Linear interpolation: inserted slice = weighted average of adjacent slices. Fast; smooth appearance; no ringing. Blurry.
• Cubic spline interpolation: uses four adjacent slices; smoother; less blurry; may introduce mild overshoot.
• Sinc interpolation: theoretically optimal for band-limited data; in practice, the slice direction of 2D acquisitions is NOT band-limited (it is discretely sampled with a non-Fourier profile) → sinc interpolation in z for 2D is less physically justified than in-plane.

Clinical relevance: slice interpolation improves the visual quality of MPR reconstructions from 2D datasets (curved reformats along a vessel, oblique sections through a joint) at the cost of true z-resolution. For a 5 mm slice with 1 mm gap, interpolating to display 1 mm slices creates 5 synthetic intermediate slices — the appearance improves but anatomical detail between slices is invented, not measured.

2.2.4 3D Isotropic Reformatting — True Resolution

In 3D acquisitions with isotropic voxels (Δx = Δy = Δz = a), the full 3D Fourier dataset is reconstructed with equal resolution in all directions. Reformatting this dataset into a non-primary plane (e.g., reformatting a sagittally-acquired brain MPRAGE into coronal and axial planes) is not interpolation — it is a full-resolution reconstruction in the new plane. The reformatted image has the same voxel dimension in all three directions and does not introduce interpolation artefacts beyond the standard voxel-size limitations.

This is the fundamental advantage of 3D isotropic acquisition over 2D slice interpolation: in 3D, all planes have equal true resolution; in 2D with interpolation, the z-direction resolution is fixed by the slice thickness and cannot be recovered by post-processing.

3. Units, Terminology and Vendor Nomenclature

Reconstruction matrix and interpolation parameters are expressed in pixel counts (dimensionless) or as multiplication factors (×2, ×4).

GE ZIP terminology deserves specific explanation:

• ZIP2: ×2 zero-filling in both phase and frequency directions (e.g., 256→512 in both)
• ZIP4: ×4 zero-filling (256→1024)
• ZIP512: frequency-direction zero-filling to 512 pixels only (useful when the frequency matrix is 256 but a 512-pixel frequency display is desired)

Siemens "Interpolation": a simple ON/OFF toggle that applies ×2 zero-filling in both directions when enabled. The resulting reconstruction matrix is displayed (e.g., "512×512" from a 256×256 acquisition). The acquired ("Base resolution") is separately documented in the DICOM header.

4. Typical Value Ranges

4.1 In-Plane Interpolation Factors

4.2 Slice Interpolation in 2D

4.3 3D Reformatting (No Interpolation — True Resolution)

5. Parameter Interaction Ecosystem

5.1 Parameter Relationships Matrix

5.2 The Acquisition Matrix — Reconstruction Matrix Distinction

The most critical concept of this page:

ACQUISITION MATRIX (N_x × N_y)
  ↓ Determines: true spatial resolution, SNR, scan time
  ↓ Set by: FOV + target voxel size

FOURIER TRANSFORM (reconstruction)
  ↓ + Zero-filling (optional)

RECONSTRUCTION MATRIX (M_x × M_y)
  ↓ Determines: display pixel size, apparent smoothness, Gibbs ringing
  ↓ Set by: acquisition matrix × interpolation factor
  
M_x = N_x × ZF_x    (where ZF_x = zero-fill factor in x)
M_y = N_y × ZF_y

TRUE PIXEL SIZE (display): FOV / M_x
TRUE RESOLUTION: FOV / N_x  (ALWAYS the acquisition matrix, not reconstruction)

6. Effects on Image Appearance

6.1 Increasing Reconstruction Matrix (More Zero-Filling)

Visual smoothing: sharp pixel boundaries between adjacent voxels are smoothed by the sinc interpolation. The image looks less "pixelated" or "blocky." This is visually appealing but diagnostically neutral.

Gibbs ringing reduction: as described in Section 2.2.2, zero-filling reduces but does not eliminate Gibbs ringing at high-contrast boundaries. The reduction is approximately proportional to the zero-fill factor. At ×2 zero-filling, ringing amplitude ≈ 4.5% vs 9% without zero-filling.

Apparent edge sharpness: paradoxically, sinc interpolation can make edges appear sharper than in the non-zero-filled image, because the sinc overshoot creates a subtle bright–dark fringe at high-contrast boundaries. This is not true resolution gain — it is a display artefact of the interpolation function.

Small lesion conspicuity: for a lesion at or near the resolution limit (size ≈ 1–2 voxels), zero-filling can make it more visible by reducing the "pixelated" appearance that might cause it to be dismissed as noise. However, this is a perceptual, not a physical, improvement.

No change in contrast-to-noise ratio: the tissue contrast (CNR = signal_tissue1 − signal_tissue2) / noise is determined entirely by the acquisition parameters. Reconstruction matrix does not change CNR.

6.2 Decreasing Reconstruction Matrix (Fewer Pixels Than Acquired)

Reducing the reconstruction matrix below the acquisition matrix (downsampling) discards acquired k-space information — it reduces resolution below the acquisition limit. This is never intentional in routine clinical MRI but may occur in error (e.g., incorrect DICOM export settings). It produces an obviously blurry image.

6.3 Slice Interpolation Appearance Effects

Smoother MPR transitions: with slice interpolation, oblique and curved reformats show smooth continuity rather than the staircase appearance of large gap or thick slices.

Blurring in z-direction: slice interpolation (particularly linear interpolation) introduces mild blurring in the z-direction. Fine z-direction detail that was present in the adjacent slices may be averaged away in the synthetic interpolated slice.

No new information: a synthetic interpolated slice shows a weighted average of adjacent acquired slices — it cannot reveal anatomy that was between the acquired slices and not visible in either.

7. Effects on Acquisition Time

7.1 Zero-Filling / Reconstruction Matrix

Zero-filling and reconstruction matrix changes are post-acquisition computations with zero scan time impact. The scanner adds zeros to k-space and computes the larger Fourier transform during reconstruction — this occurs while the scan is running (pipeline reconstruction) or immediately after scan completion, adding seconds to minutes of reconstruction time but adding no time to the data acquisition.

Exception: on very old systems (pre-2010) without pipeline reconstruction, computing a 512×512 FFT instead of 256×256 added significant post-processing delay. On modern systems, this is negligible.

7.2 Slice Interpolation

Slice interpolation is always a post-processing computation — it adds no scan time. It may add seconds to the image reconstruction queue before the DICOM images are available on the workstation.

7.3 3D Reformatting

3D reformatting of isotropic acquisitions is a post-acquisition computation. Modern workstations perform this in near-real-time (< 30 seconds for a full brain MPRAGE reformat into three planes). It does not affect scan time.

8. Effects on SNR and CNR

8.1 SNR per Pixel vs SNR per Voxel

This distinction is critical and commonly confused:

SNR per pixel (display pixel, reconstruction matrix):

• Decreases with zero-filling: SNR_pixel_ZF = SNR_pixel_no-ZF / √(ZF_factor_x × ZF_factor_y)
• At ×2 zero-filling (both directions): SNR_pixel = SNR_original / 2

SNR per voxel (acquisition voxel, true spatial unit):

• Unchanged by zero-filling: the total signal in each acquisition voxel is redistributed across multiple display pixels but the total remains the same
• SNR per voxel = FOV/N × per-voxel-signal / noise — independent of reconstruction matrix

The diagnostic significance: radiologists assess images in terms of tissue distinction (which depends on CNR per voxel, not per display pixel). Zero-filling does not change diagnostic SNR. The visual appearance of noise changes (the noise texture appears finer, more "powdery" after zero-filling) but the fundamental signal quality is unchanged.

8.2 DLR as a "Meaningful" Reconstruction Matrix Enlargement

Deep learning reconstruction (DLR — see Section 18 of the parent page 9501) operates on the acquired k-space data (or on the initial Fourier-reconstructed image) and applies a neural network that has learned to produce high-quality images from low-SNR or low-resolution inputs. Unlike zero-filling:

• DLR can generate plausible high-frequency image content beyond the Nyquist limit of the acquisition (though this is extrapolation, not measurement)
• DLR improves SNR per voxel at the acquisition matrix level
• DLR changes the noise texture in ways that differ from zero-filling

The distinction: zero-filling redistributes existing information to more display pixels; DLR generates new image content (plausible but not measured). Both increase the display matrix, but with fundamentally different diagnostic implications.

9. Artefacts Associated with Reconstruction Matrix and Interpolation

10. Behaviour Across Sequence Families

Spin Echo and Turbo Spin Echo

Zero-filling is routinely applied (×2 in most clinical protocols). The sinc interpolation effectively reduces the visible impact of the finite acquisition matrix — particularly for TSE sequences where the moderate ETL already introduces mild T2 blurring, which has a similar appearance to zero-filling smoothing.

Gradient Echo

Standard ×2 zero-filling is applied to all body GRE sequences (VIBE, LAVA). In DCE dynamic series, the reconstruction matrix is fixed at the protocol level; changing it during a dynamic series would introduce inconsistency between phases. All phases must use the same reconstruction matrix for kinetic analysis and subtraction.

Inversion Recovery (STIR, FLAIR, MPRAGE)

For MPRAGE, the reconstruction matrix is 256×256 (ADNI standard) — the same as the acquisition matrix. No zero-filling is applied in ADNI-compliant MPRAGE because volumetric tools (FreeSurfer) are validated on specific acquisition and reconstruction matrix combinations; zero-filling changes the voxel intensity distributions and may affect automated segmentation.

EPI (DWI, DSC, fMRI)

EPI acquisitions use aggressive zero-filling (×4 is common for DWI: acquired 96×96, displayed 384×384). This is a pragmatic choice: EPI has inherently low geometric accuracy from susceptibility distortion, and displaying a 96×96 image has no aesthetic appeal in clinical practice. The ×4 zero-filled image looks clinically acceptable. However, the true spatial resolution remains at the 96×96 acquisition matrix — when lesion size is measured from DWI, measurements should use the voxel dimensions from the acquisition matrix, not the display pixel size.

DWI and ADC maps: the ADC map is calculated from the acquired DWI signal values, then zero-filled for display. The ADC value within a voxel is unchanged by zero-filling — the same number is distributed across more display pixels. Zero-filling does not change ADC measurements.

DSC Perfusion

Perfusion maps (CBV, CBF, MTT) are calculated on the acquired EPI matrix (typically 64×64 or 96×96), then zero-filled or spatially smoothed (Gaussian filter) for display. The temporal SNR required for perfusion quantification is determined by the acquisition matrix, not the display matrix. A 64×64 acquisition with ×8 zero-filling displayed at 512×512 looks cleaner but still has the fundamental perfusion map resolution of 64×64.

DCE

Same principles as GRE above. The kinetic analysis (Tofts model, maximum slope) operates on the acquired matrix. Zero-filling is applied consistently to all dynamic phases. For subtraction images (post minus pre), the zero-filling must be identical for both phases to ensure pixel-wise subtraction is geometrically consistent.

ASL

Perfusion maps from ASL are typically displayed at larger reconstruction matrices (acquired 64×64 → displayed 256×256 or larger). The inherently low ASL SNR means that the fundamental image quality is determined by the acquired matrix; zero-filling provides cosmetic improvement. The perfusion quantification (CBF in mL/100g/min) is computed on the acquired matrix.

bSSFP (CISS/FIESTA)

For inner ear and brachial plexus 3D CISS at sub-millimetre isotropic resolution, the reconstruction matrix equals the acquisition matrix (no zero-filling needed — the acquired resolution is already sufficient). For clinical brain 3D CISS at 0.8 mm isotropic: a 256×256 acquisition with zero-filling to 512×512 is common.

Spectroscopy

In MRS, "interpolation" refers to the spectral resolution and point spread function of the spectral display — not spatial image interpolation. Spatial spectral maps (MRSI) may be zero-filled for display in the spatial dimensions, smoothed by a Gaussian spatial filter, or presented at the native spectral voxel resolution. Gaussian smoothing of MRSI maps is a common alternative to zero-filling, producing blurred but smoothly varying maps suitable for displaying metabolite distributions.

11. Field Strength Behaviour

Zero-filling and reconstruction matrix are field-strength independent computations — the same mathematical operations apply at 0.55T, 1.5T, 3T, and 7T. However, the practical value of zero-filling differs with field strength:

At 0.55T: the lower SNR forces lower acquisition matrices in many protocols. Zero-filling ×4 is commonly applied to produce displayable images that don't look excessively pixelated. The genuine Gibbs ringing reduction from zero-filling has higher clinical value at 0.55T than at 3T, because the lower acquisition matrix at 0.55T produces more pronounced ringing at any given anatomy.

At 7T: the high intrinsic SNR enables very high acquisition matrices — zero-filling is often not applied because the acquired matrix is already sufficient for smooth display.

12. Vendor-Specific Implementation

Siemens

The "Interpolation" toggle in the contrast tab applies ×2 zero-filling in both in-plane directions simultaneously. The resulting reconstruction matrix is displayed in the protocol card but the underlying acquisition matrix (called "Base resolution" in Siemens) is the true resolution parameter. The DICOM tag "Acquisition Matrix" (0018,1310) contains the acquired matrix; the "Rows" and "Columns" tags (0028,0010/0011) contain the reconstruction matrix.

For 3D sequences (MPRAGE, SPACE, VIBE): the "Interpolation" toggle operates on the partition (z) direction separately from the in-plane directions. In-plane interpolation and z-interpolation can be independently enabled.

GE

GE's ZIP (Zero-fill Interpolation) family is the most granular implementation:

• ZIP2: ×2 in-plane (both directions)
• ZIP4: ×4 in-plane
• ZIP512: ×2 in the frequency direction specifically (useful for matching 256-matrix acquisitions to a 512-pixel display without unnecessarily upsampling the phase direction)

ZIP is applied to all standard clinical GE protocols by default. The true acquisition matrix is documented in the DICOM header under "Acquisition Matrix." Technologists should verify whether ZIP is active and document it in the protocol notes.

GE-specific slice interpolation: GE workstations (AW) offer dedicated slice interpolation for 2D stacks in the reformatting toolbox — the acquired slices are not changed; synthetic slices are inserted at the post-processing workstation stage.

Philips

Philips distinguishes explicitly between "Scan matrix" (acquisition) and "Reconstruction matrix" (display). The interface allows independent setting of both. When the reconstruction matrix exceeds the scan matrix, zero-filling is automatically applied. Philips provides a direct readout of the resulting voxel sizes in all three dimensions on the protocol design screen — the clearest implementation for verifying true resolution vs display resolution.

Canon

Standard zero-filling toggle ("Interpolation" or equivalent). Canon systems by default apply ×2 zero-filling on brain and MSK sequences. The "Recon. matrix" field shows the display matrix; the "Scan matrix" shows the acquired matrix.

United Imaging

UIH systems display both acquisition and reconstruction matrix explicitly in the scan planning interface. Default zero-filling is ×2 for most clinical protocols, consistent with industry standard practice.

DICOM tags for reconstruction vs acquisition matrix:

• Acquisition Matrix: DICOM (0018,1310) — [freq_rows, freq_cols, phase_rows, phase_cols]
• Image Rows / Columns (display): DICOM (0028,0010) / (0028,0011)
• Pixel Spacing (display pixel size): DICOM (0028,0030) — reflects reconstruction matrix / FOV
• Acquisition Pixel Spacing: derived from FOV / acquisition matrix — requires calculation

13. Practical Optimisation Strategies

13.1 Clinical Optimisation Recipes

14. Parameter Extremes

14.1 No Interpolation (Reconstruction = Acquisition Matrix)

When reconstruction matrix equals the acquisition matrix, the image represents the raw Fourier-transformed data at the true voxel resolution. For coarse acquisitions (DWI at 128×96), the result appears blocky and pixelated. This is appropriate for: ADNI MPRAGE (mandated no-interpolation for volumetric tools); ADC map generation (measurements should be at native resolution); any quantitative MRI where reconstruction artefacts must be minimised.

14.2 Extreme Interpolation (×8 or more)

Very large zero-filling factors (reconstructed matrix 8× the acquisition matrix, e.g., 64×64 → 512×512) produce images that look deceptively smooth and high-resolution. The actual spatial detail is unchanged — only 64×64 independent measurements exist. Each display pixel represents a sub-voxel region with signal synthesised entirely by sinc interpolation. This extreme zero-filling is rarely clinically useful and can mislead inexperienced readers about the true resolution.

14.3 Slice Interpolation to Very Thin Synthetic Slices

Interpolating 5 mm slices to create 0.5 mm synthetic slices (×10 interpolation) produces smooth-appearing MPR reformats that completely hide the true 5 mm z-resolution. A slice of 0.5 mm thickness from such interpolation contains no more anatomical information than the original 5 mm slice — it is mathematical smoothing, not physical measurement. For clinical reporting, the true slice thickness (5 mm) must be stated, not the interpolated thickness.

15. Common Optimisation Errors

16. MRI Technologist Pearls

Always read the acquisition matrix, not the display matrix: on any scanner console, check the "base resolution" or "scan matrix" field — not the reconstruction matrix. The acquisition matrix determines what you actually measured. DICOM tag 0018,1310 contains the acquired matrix; the rows/columns tags contain the display matrix.

Zero-filling is included in reconstruction time, not scan time: the displayed image creation time (after the last TR) includes the Fourier transform and any zero-filling. On modern scanners this adds < 10 seconds. If a reconstruction seems to be taking longer than expected, zero-filling factor is not the cause — check for other post-processing steps (DLR, distortion correction, subtraction).

Zero-filling consistency for subtraction sequences: for any sequence where subtraction is performed (breast DCE subtraction, ARIA-E FLAIR subtraction, post-minus-pre T1), verify that the reconstruction matrix (and therefore the display pixel spacing) is identical for both pre-contrast and post-contrast acquisitions. Inconsistent interpolation produces subtraction artefacts.

Slice interpolation label in MPR: when presenting MPR reformats from 2D acquisitions with slice interpolation, add a note to the image or the report: "Images reformatted with slice interpolation; true slice thickness = X mm." Many PACS systems do not distinguish interpolated from acquired slice thickness in the image header.

DLR and ADC: if DLR is applied to DWI images before ADC map generation, the ADC values may differ from standard FFT reconstruction. For longitudinal studies measuring ADC changes, use a consistent reconstruction pipeline (DLR always on, or DLR always off). Mixing DLR and non-DLR ADC maps introduces systematic measurement variability.

The "Zip512" shortcut: for GE protocols where the frequency matrix is 256 and the phase matrix is 192, applying ZIP2 (×2) gives 512×384. Applying ZIP512 instead gives 512×192 — only the frequency direction is upsampled. This may be appropriate when the phase direction resolution is already adequate but the frequency display could benefit from smoothing.

17. Real Clinical Examples

Example 1: DWI Lesion Sizing — Measurement Error from Display Matrix

Clinical scenario: acute ischaemic stroke; DWI acquisition at 128×96 at FOV 240×180 mm. True voxel: 1.875×1.875 mm. After ZIP2 (×2) zero-filling: display at 256×192 → display pixel: 0.94×0.94 mm. After ZIP4 (×4): display at 512×384 → display pixel: 0.47×0.47 mm.

Problem: the radiologist measures the DWI lesion using the PACS measurement tool on the ZIP4 image. The reported area uses 0.47 mm pixels — suggesting sub-voxel accuracy. The measurement variability is actually ±1 acquisition voxel = ±1.875 mm.

Consequence: a lesion measured as "8.4 mm × 6.2 mm" from ZIP4 display pixels should be reported as "approximately 8–9 mm × 6–7 mm" — one acquisition voxel accuracy. The ZIP4 display gives the appearance of precision that the acquisition does not support.

Correction: always use the acquisition voxel size for lesion dimension precision estimates. For this protocol: "Lesion approximately 2× acquisition voxel in each direction" is the meaningful precision.

Example 2: MPRAGE for AD Volumetry — Zero-Filling Invalidates FreeSurfer

Clinical scenario: a dementia clinic acquires MPRAGE for hippocampal volumetry using FreeSurfer. Protocol: ADNI-standard 1 mm isotropic, 256×256×192. However, the technologist had inadvertently enabled the "Interpolation" toggle on the Siemens scanner → the images are reconstructed at 512×512 (×2 zero-filled).

Problem: FreeSurfer's hippocampal segmentation was validated and normalised on ADNI-protocol images at 256×256 — specifically non-interpolated. The zero-filled 512×512 images produce different surface fitting behaviour, altered cortical thickness estimates, and systematically different hippocampal volume measurements compared with the ADNI reference dataset.

Result: the hippocampal volume for this patient is reported as 2.1 mL (using the 512×512 images). The ADNI normative database comparison is invalid — the patient's result cannot be correctly compared to the reference.

Correction: disable interpolation for ADNI-compliant MPRAGE; set reconstruction matrix = acquisition matrix = 256×256.

Lesson: for quantitative protocols (volumetry, ADC measurements, T1/T2 mapping), zero-filling must be specified in the protocol and must match the validation conditions of the measurement tool.

Example 3: Slice Interpolation for Liver MPR — Clinical Utility and Limitation

Clinical scenario: post-Whipple resection; surveillance liver MRI at 3.5 mm axial slices (liver DCE VIBE). The surgeon requests coronal MPR for biliary anastomosis assessment.

Without slice interpolation: coronal MPR shows staircase appearance at 3.5 mm steps. The biliary anastomosis appears as a series of blocks rather than a smooth duct — difficult to assess continuity.

With ×2 slice interpolation: synthetic 1.75 mm slices inserted between acquired slices. The coronal MPR shows a smooth, continuous biliary anastomosis, clearly distinguishable from post-operative fibrosis.

Limitation: the interpolated coronal MPR looks like a 1.75 mm z-resolution image but is actually 3.5 mm. A small bile leak producing a <2 mm biloma between slices would not be detected — it falls between two acquired slices and would be blended away by interpolation.

Clinical decision: document the true acquired slice thickness (3.5 mm) in the report and note "interpolated coronal reformats for display." If a bile leak is suspected, request dedicated thin-slice MRCP or 3D T2 SPACE acquisition.

Example 4: Zero-Filling for Routine Brain TSE — Standard Clinical Practice

Clinical scenario: routine brain T2 TSE; acquisition 256×192 at FOV 240×180 mm → true voxel: 0.94×0.94 mm. Without zero-filling: display at 256×192 (0.94 mm pixels). With ×2 zero-filling: display at 512×384 (0.47 mm display pixels).

Effect on Gibbs ringing: without zero-filling at the brain-CSF interface (frontal horns, quadrigeminal cistern), the truncation ringing produces bright bands of approximately 9% signal height — visible as subtle alternating lines adjacent to CSF. With ×2 zero-filling: ringing bands at 4.5% — less visible; rarely confused with periventricular signal change.

Effect on readability: the ×2 zero-filled image is the universally preferred clinical display. It provides smooth rendering of sulcal and gyral anatomy, clearly visible cortical margins, and reduced artefacts at CSF boundaries.

No diagnostic impact on lesion detection: a 4 mm white matter lesion (4×4 mm = 16 acquisition voxels) is equally detectable in both reconstructions. The zero-filling is a display improvement; the physical measurement is identical.

Standard practice: apply ×2 zero-filling to all routine clinical brain TSE sequences as the universal default. Report resolution in terms of the acquisition matrix (0.94 mm in-plane, 4 mm slice) rather than the display matrix.

18. Visual Educational Material

18.1 Acquisition vs Reconstruction Matrix Concept

k-SPACE (acquired N × N)          IMAGE SPACE
                                  
No zero-filling (M = N):
F(N×N) → IFT → image(N×N)         Pixelated; Gibbs ringing amplitude ~9%
                                   True spatial resolution = FOV/N

With ×2 zero-filling (M = 2N):
F(N×N)+zeros → F(2N×2N) → IFT → image(2N×2N)
                                   Smoother; Gibbs ~4.5%; 4× more display pixels
                                   True spatial resolution STILL = FOV/N
                                   
TRUE RESOLUTION: always set by N (acquisition)
DISPLAY QUALITY: improved by zero-filling (zero-fill factor M/N)

What CHANGES with zero-filling: display pixel size, Gibbs ringing, smoothness
What DOES NOT CHANGE: true resolution, SNR per voxel, CNR, tissue contrast

18.2 Slice Interpolation vs True 3D Resolution

2D ACQUISITION (5 mm slices, 1 mm gap):

Acquired:      [S1] . [S2] . [S3] . [S4] . [S5]
                5mm  1mm  5mm  1mm  5mm  1mm  5mm  ← true coverage

Interpolated:  S1 s12 s13 s14 s15 S2 s23 s24 s25 S3 ...
(×5 interp.)       synthetic        synthetic
                     
MPR looks: smooth (but z-resolution = 5mm; gaps = 1mm; no info in between)

3D ACQUISITION (1 mm isotropic):

Partitions:    [P1][P2][P3][P4][P5][P6][P7][P8][P9]... ×160
                1mm every partition; full Fourier z-encoding; no gaps

MPR from 3D:   TRUE 1 mm z-resolution in any reformat plane
                (ACTUAL measurement, not interpolation)

→ 3D isotropic ≠ 2D + slice interpolation (physically and diagnostically)

18.3 Zero-Filling Decision Tree

IS THE IMAGE INTENDED FOR QUANTITATIVE MEASUREMENTS?
│
├── YES (ADC, T1/T2 map, volumetry)
│   → Check: does the measurement tool specify a reconstruction matrix?
│   ├── YES (e.g., ADNI MPRAGE: no ZF) → Match the specified reconstruction
│   └── NO → Apply minimal ZF (×1 or ×2); document in protocol
│
└── NO (clinical visual assessment)
    → Apply standard ×2 ZF for all sequences
    → Apply ×4 ZF for EPI (DWI, fMRI, perfusion) for presentable display
    → Use consistent ZF for all phases of dynamic/subtraction series
    
REPORT: Always state acquired resolution (FOV/acquisition_matrix)
        Not display resolution (FOV/reconstruction_matrix)

19. Evidence Gaps and Ongoing Debate

DLR vs zero-filling — diagnostic equivalence: DLR-based reconstruction provides visually superior images to zero-filled standard FFT reconstruction, particularly for high-matrix acquisitions and thin-slice 3D protocols. Whether DLR-improved images translate to measurably better diagnostic performance (sensitivity, specificity for specific pathology classes) compared with zero-filled standard reconstruction has been studied for specific applications (brain, prostate, knee) but not comprehensively across all clinical domains.

ADC value stability across reconstruction methods: the impact of DLR (as an alternative to zero-filling) on ADC values has been studied in limited phantom and patient series. Systematic ADC biases from DLR have been reported. The threshold at which these biases become clinically significant (affecting diagnosis, staging, or treatment response assessment) is not established.

Standardisation of interpolation reporting in radiology: there is no international standard for how DICOM headers report interpolated vs acquired image dimensions. Different vendors encode the "Slice Thickness" and "Pixel Spacing" DICOM tags inconsistently — some report the acquired values, others report the interpolated display values. This creates ambiguity in multi-centre studies and when images are transferred between institutions. No consensus body has addressed this standardisation gap.

Optimal zero-filling factor for AI lesion detection: AI lesion detection algorithms trained on zero-filled images may perform differently on non-zero-filled acquisitions, and vice versa. The optimal reconstruction matrix for AI-assisted brain lesion, prostate lesion, and breast lesion detection has not been systematically validated across vendors and reconstruction strategies.

20. Miscellaneous and Future Directions

Historical context: the zero-filling technique was formally described in the MRI context by Reeves and Simons in 1985 and subsequently refined by multiple groups in the late 1980s. The Gibbs ringing problem — and zero-filling as a partial solution — was recognised in the earliest clinical MRI literature because the brain-CSF interface artefacts were immediately apparent in clinical brain images. The Fermi filter and Hamming window alternatives were described in parallel as competing approaches to the same truncation problem.

Beyond sinc: advanced interpolation methods: mathematical interpolation methods beyond zero-filling (sinc) include:

• Hermite spline interpolation: matches both value and derivative at acquired points → smoother without the overshoot of sinc
• Lanczos interpolation: a truncated sinc kernel with reduced side lobes
• Model-based reconstruction: uses anatomical priors from an atlas to fill in sub-voxel detail — the boundary between "interpolation" and "AI reconstruction"

Synthetic MRI and virtual reconstruction matrices: synthetic MRI (T1/T2 quantitative imaging + synthetic image generation) produces images at arbitrary TE/TR combinations from a single acquisition. The synthetic images are generated at the native resolution of the quantitative acquisition and can be displayed at any reconstruction matrix without additional scan time. This may decouple the display matrix from the acquisition process in future clinical workflows.

AI super-resolution: neural networks trained on paired low-resolution/high-resolution image datasets can generate plausible high-resolution images from low-resolution inputs. This goes beyond zero-filling — the network generates new spatial frequency content. Super-resolution for MRI is an active research area, with initial validation for brain, prostate, and cardiac applications. The boundary between AI super-resolution and DLR denoising is methodologically blurred.

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