Reconstruction and Post-Processing Physics

Version 1.0 — July 2026

1. Executive Summary

This page covers, in depth, the eleventh and final topic group listed in Section 4.11 of the MRI Physics — Fundamentals and Principles master page: the mathematical and physical basis of modern image reconstruction. Three topics are covered: the mathematical/physical basis of SENSE and GRAPPA parallel imaging reconstruction, compressed sensing, and the physical principles behind AI/deep-learning reconstruction. With this page, all eleven topic groups originally laid out in the master page’s roadmap (Section 4) have now been published as dedicated child pages, completing the initial build-out of the MRI Physics cluster. Reconstruction is, in a real sense, where every other topic in this cluster converges: k-space sampling strategy (companion Signal Localisation and Image Formation child page), hardware coil-array capability (companion Magnetic Field, Hardware, and Homogeneity child page), and the fundamental resolution/SNR/scan-time trade-off (also introduced there) all become concrete, quantitative constraints only once a specific reconstruction algorithm is chosen to turn acquired data into a final image.

2. Relationship to the Master Page

The master page introduces reconstruction and post-processing physics only as a listed topic heading (Section 4.11), without detail, and several companion child pages have already referenced this topic group as the natural home for material beyond their own scope: the RF Pulses and Pulse Sequence Physics child page (adiabatic pulse SAR trade-offs), the Signal Localisation and Image Formation (k-Space) child page (non-Cartesian and highly undersampled reconstruction validity), the Physics of MRI Artefacts child page (partial Fourier’s interaction with truncation ringing), the SAR, Bioeffects, and MRI Safety Physics child page (whole-body vs local SAR), and the Chemical Shift and Fat/Water Physics child page (quantitative multi-peak fat spectral modelling). This page assumes, as direct prerequisites, the k-Space child page’s treatment of k-space sampling and undersampling, and the Hardware child page’s treatment of the phased array (Roemer et al., already cited there) as the physical basis of multi-channel receive data. What this page adds: the specific mathematical mechanisms by which SENSE, GRAPPA, compressed sensing, and deep-learning methods each recover a full image from deliberately incomplete or otherwise transformed data.

3. Mathematical and Physical Basis of SENSE and GRAPPA

3.1 The Shared Underlying Problem

Both SENSE and GRAPPA, introduced practically on the existing Parallel Imaging parameter page and physically motivated by the phased-array hardware described in the companion Magnetic Field, Hardware, and Homogeneity child page, address the same underlying mathematical problem: when k-space is undersampled below the Nyquist rate established in the companion Physics of MRI Artefacts child page (Section 4.1 there), the resulting image would ordinarily suffer aliasing — but a multi-channel receive array’s several coil elements, each with a distinct, independently measurable spatial sensitivity pattern, provides additional spatial information that a single-channel acquisition lacks, which can in principle be used to resolve the aliasing mathematically rather than by acquiring more raw data.

3.2 SENSE — Image-Domain Unfolding

Pruessmann and colleagues’ 1999 SENSE (SENSitivity Encoding) method works directly in the image domain: because each aliased pixel in an undersampled reconstruction is actually a weighted superposition of a small number of true, non-aliased pixel locations (the specific locations determined by the acceleration factor and each coil element’s known sensitivity at each of those locations), a system of linear equations — one per coil element, with the true pixel values as unknowns — can be solved directly to “unfold” the aliased image back into its true, unaliased form, provided the coil sensitivity maps are known accurately and the number of independent coil elements is at least as great as the acceleration factor [1].

3.3 GRAPPA — k-Space Domain Interpolation

Griswold and colleagues’ 2002 GRAPPA (GeneRalized Autocalibrating Partially Parallel Acquisitions) method takes a mathematically distinct approach, working directly in k-space rather than image space: missing k-space lines (skipped due to undersampling) are estimated directly from the acquired lines using a set of linear weights derived from a small region of fully-sampled “autocalibration” k-space data acquired as part of the same scan, exploiting the same underlying coil-sensitivity-encoded spatial information as SENSE, but without requiring an explicit, separately-measured sensitivity map [2]. Because GRAPPA reconstructs missing k-space data directly, standard Fourier reconstruction (introduced in the companion Signal Localisation and Image Formation (k-Space) child page) can then be applied unchanged to the completed k-space dataset — a practical advantage in workflow integration that partly explains GRAPPA’s widespread adoption alongside SENSE-family methods.

3.4 Shared Physical Limitation — the g-Factor

Both methods share a fundamental, physically-grounded SNR penalty beyond the simple “fewer averages” SNR loss expected from acquiring less data (introduced in the companion Signal Localisation and Image Formation (k-Space) child page’s Section 7.2): the geometry (g-) factor, which quantifies additional, spatially-varying noise amplification arising specifically from how well-conditioned (versus how similar to one another) the coil sensitivity patterns are at each unfolded pixel location. Regions where coil sensitivities are similar to one another (providing less independent spatial information to resolve aliasing) show higher g-factor penalty and correspondingly worse local SNR than regions where coil sensitivities differ substantially — a direct, physically unavoidable consequence of the underlying mathematics, not a fixable implementation detail, and the reason parallel imaging acceleration cannot be increased indefinitely without a corresponding, non-uniform SNR cost.

4. Compressed Sensing

4.1 The Core Insight — Exploiting Image Sparsity

Lustig, Donoho, and Pauly’s landmark 2007 description of compressed sensing MRI introduced a mathematically distinct undersampling-recovery strategy from parallel imaging: rather than relying on multi-channel coil sensitivity information (Section 3), compressed sensing exploits the fact that most MR images have a sparse representation — most of their information content can be captured by relatively few non-zero coefficients — in some suitable mathematical transform domain (for example, a wavelet transform for typical anatomical images, or the raw pixel domain itself for inherently sparse images such as MR angiograms) [3]. Provided k-space is undersampled in a sufficiently incoherent (pseudo-random, rather than regular/periodic) pattern, and reconstruction uses an appropriate non-linear recovery algorithm that enforces this known sparsity as a mathematical constraint, an image can be recovered from substantially fewer k-space samples than the Nyquist criterion would otherwise require — a genuinely different mathematical mechanism from parallel imaging’s coil-sensitivity-based approach, exploitable independently of, and in combination with, parallel imaging acceleration.

4.2 Why Incoherent (Not Regular) Undersampling Matters

A critical, specific physical/mathematical requirement distinguishes compressed sensing from simple regular undersampling: regular, periodic undersampling (as used in standard parallel imaging, Section 3) produces coherent, structured aliasing artefact that directly overlaps true image structure in a predictable way, while pseudo-random, incoherent undersampling produces aliasing that instead resembles diffuse, noise-like interference rather than structured replication — a form of “artefact” that sparsity-enforcing non-linear reconstruction algorithms can distinguish from genuine image content far more effectively than they could distinguish coherent aliasing. This is the specific mathematical reason compressed sensing sequences use deliberately irregular (rather than simply sparser-but-still-regular) k-space sampling patterns.

5. Physical Principles Behind AI/Deep-Learning Reconstruction

5.1 From Hand-Crafted Priors to Learned Priors

Both parallel imaging (Section 3) and compressed sensing (Section 4) rely on a mathematically explicit, hand-specified model of what makes an image “correct” — known coil sensitivities, or a hand-chosen sparsifying transform and regularisation strength. Deep-learning reconstruction methods, exemplified by Hammernik and colleagues’ 2018 variational network approach, replace this hand-specified model with a learned one: a neural network, trained on large datasets of paired undersampled and fully-sampled MR data, learns — implicitly, through its trained parameters — a far more complex, data-derived representation of what constitutes a physically and anatomically plausible MR image than any hand-crafted sparsity assumption could easily encode [4].

5.2 Physics-Informed Architecture — the Unrolled Optimisation Approach

Rather than treating reconstruction as an unconstrained, purely data-driven image-to-image translation problem, the variational network approach specifically embeds the known MR physics — the same Fourier/k-space relationship and coil sensitivity encoding underlying Sections 3-4 — directly into the network architecture itself, structured as a fixed number of “unrolled” iterations of a classical gradient-descent-style optimisation scheme, with only the prior model (what an anatomically plausible image should look like) and certain data-weighting terms learned from training data rather than hand-specified. This physics-informed architectural choice is not incidental: by constraining the network to remain consistent with the genuinely-measured k-space data at every iteration (data consistency), rather than allowing an unconstrained network to freely hallucinate plausible-looking but physically unsupported image content, this approach directly addresses the physical-plausibility concern already flagged as an open question in the companion Signal Localisation and Image Formation (k-Space) child page’s evidence-gaps discussion (Section 11 there).

5.3 What Remains Physically True Regardless of Reconstruction Method

Regardless of which reconstruction approach (parallel imaging, compressed sensing, or deep learning) is used, none of them alter the underlying physical information-content limit established by the actual acquired k-space data and the coil array’s genuine spatial encoding capability (Sections 3.1, 4.1) — each method is a different mathematical strategy for extracting the maximum genuine information from a given, physically fixed acquisition, not a means of acquiring information that was never physically encoded in the first place. This is the physically grounded reason that, however sophisticated a reconstruction algorithm becomes, sufficiently aggressive undersampling combined with insufficient genuine spatial encoding (from too few independent coil elements, or insufficiently sparse/incoherent sampling) eventually reaches a hard information-theoretic limit beyond which no reconstruction method — however advanced — can recover genuinely lost information without risking the “hallucination” of plausible-but-unsupported detail.

6. MRI Technologist and Radiologist Pearls — Common Misconceptions

• “SENSE and GRAPPA are the same algorithm with different names.” They work in genuinely different mathematical domains (image-space unfolding vs k-space interpolation, Sections 3.2-3.3), sharing the same underlying physical information source (coil sensitivity) but not the same reconstruction mathematics.
• “Compressed sensing and parallel imaging exploit the same physical mechanism, just described differently.” They rely on genuinely distinct physical/mathematical resources — coil sensitivity spatial information (Section 3) vs image sparsity in a transform domain (Section 4) — which is precisely why they can be, and routinely are, combined for acceleration beyond what either could achieve alone.
• “AI reconstruction can recover any level of undersampling given enough training data.” As emphasised in Section 5.3, no reconstruction method — including deep learning — can exceed the genuine information-theoretic content of the actual acquired data and coil encoding; sufficiently aggressive undersampling reaches a hard physical limit regardless of reconstruction sophistication.
• “The g-factor penalty in parallel imaging is a software limitation that improves with better algorithms.” As detailed in Section 3.4, it is a physically grounded consequence of coil sensitivity geometry at each specific pixel location — better reconstruction algorithms can approach the theoretical g-factor-limited performance more closely, but cannot eliminate the underlying physical constraint itself.

7. Practical and Clinical Relevance

The physics on this page underlies the acceleration-factor and reconstruction-method choices documented throughout MRIninja’s parameter and sequence pages, the specific SNR/artefact trade-offs associated with different acceleration strategies (directly explaining, via the g-factor, why acceleration-related SNR loss is spatially non-uniform rather than a simple global scaling), and the emerging role of deep-learning reconstruction in enabling higher acceleration factors than parallel imaging or compressed sensing alone, discussed practically wherever accelerated acquisition strategies are referenced elsewhere on MRIninja.

9. Evidence Gaps and Ongoing Debate

• Physical validity limits of aggressive combined acceleration, as already flagged in the companion Signal Localisation and Image Formation (k-Space) child page’s own evidence-gaps discussion, remain an active area of technical debate — particularly the risk of plausible-looking but physically unsupported “hallucinated” detail at high combined acceleration factors using deep-learning reconstruction (Section 5.2-5.3), without full consensus on standardised validation methodology across vendors and research groups.
• Generalisability of learned reconstruction priors across pathology and patient populations not well represented in training data remains a genuinely open, actively researched question, directly relevant to the data-consistency-based mitigation strategies discussed in Section 8.2, and distinct from the more classically characterised, physically-derived limitations of parallel imaging and compressed sensing.
• Optimal combination strategy for hybrid acceleration approaches (Section 8.1) — how best to combine parallel imaging, compressed sensing, and learned priors within a single reconstruction pipeline — remains an active engineering and methodological research area without a single settled, universally adopted framework.

10. Evidence-Based References

A. Guidelines / Consensus / Society Recommendations

No dedicated society guideline exists for foundational reconstruction and post-processing physics as such — this is technical/physics and methodology literature rather than a clinical practice area subject to society guidance. Category A is therefore not populated for this child page.

D. Technical MRI Papers

End of document — Reconstruction and Post-Processing Physics — Child Page under the MRIninja MRI Physics — Fundamentals and Principles master page — v1.0 — July 2026 Parent page: MRI Physics — Fundamentals and Principles