PaperlayerView exact revision on arXiv

Source-linked AI summary

MRI Super-Resolution with Deep Learning: A Comprehensive Survey

Mohammad Khateri, Serge Vasylechko, Morteza Ghahremani, Liam Timms, Deniz Kocanaogullari, Simon K. Warfield, Camilo Jaimes, Davood Karimi, Alejandra Sierra, Jussi Tohka, Sila Kurugol, Onur Afacan

arXiv:2511.16854v3eess.IVcs.AIcs.CVeess.SP

TL;DR

High-resolution MRI is valuable but costly to acquire, motivating computational super-resolution from lower-resolution scans. This survey organizes DL-based MRI SR methods, reviews their techniques and resources, and identifies challenges that shape future research.

  • Problem

    High-resolution MRI acquisition involves trade-offs and costs, while MRI SR research faces limited paired LR-HR data, scarce reliable HR ground truth, small samples, and domain shifts.

    Source: §I · p. 1, A. Learning Approach · p. 27

  • Method

    The survey reviews DL-based MRI SR across its methods and applications, develops a taxonomy, and curates open-access resources.

    Source: §I · pp. 1–2

  • Results

    The survey organizes MRI SR methods and consolidates advances in learning strategies, image-quality assessment, benchmark datasets, and deployment considerations.

    Source: §I · pp. 1–2

  • Takeaways & Limitations

    Progress requires greater emphasis on unsupervised and self-supervised learning, data-efficient adaptation, and methods that address unpaired data and unseen domains.

    Source: A. Learning Approach · p. 27

Abstract

from arXiv · show

High-resolution (HR) magnetic resonance imaging (MRI) is crucial for many clinical and research applications. However, achieving it remains costly and constrained by technical trade-offs and experimental limitations. Super-resolution (SR) presents a promising computational approach to overcome these challenges by generating HR images from more affordable low-resolution (LR) scans, potentially improving diagnostic accuracy and efficiency without requiring additional hardware. This survey reviews recent advances in MRI SR techniques, with a focus on deep learning (DL) approaches. It examines DL-based MRI SR methods from the perspectives of computer vision, computational imaging, inverse problems, and MR physics, covering theoretical foundations, architectural designs, learning strategies, benchmark datasets, and performance metrics. We propose a systematic taxonomy to categorize these methods and present an in-depth study of both established and emerging SR techniques applicable to MRI, considering unique challenges in clinical and research contexts. We also highlight open challenges and directions that the community needs to address. Additionally, we provide a collection of essential open-access resources, tools, and tutorials, available on our GitHub: https://github.com/mkhateri/Awesome-MRI-Super-Resolution. IEEE keywords: MRI, Super-Resolution, Deep Learning, Computational Imaging, Inverse Problem, Survey.

I. INTRODUCTION

MRI super-resolution can recover detail from faster, lower-resolution scans, but MRI-specific trade-offs and recent advances motivate a focused survey. The paper organizes DL-based methods across formulations and perspectives, reviews their approaches, and identifies resources and research directions.

I. INTRODUCTION

  • The survey addresses sparse MRI-specific coverage by reviewing DL-based MRI SR, proposing a systematic taxonomy, and discussing recent advances, challenges, and open-access resources.Its scope spans theoretical foundations, architectures, evaluation, datasets, applications, and learning strategies.Source: §I · pp. 1–2

A. Super-Resolution Formulation

  • MRI SR seeks to recover an unknown HR image from a degraded LR observation, an ill-posed inverse problem because multiple HR images can yield the same observation.The degradation may include geometric transformation, blur, undersampling, and noise; the data are assumed already reconstructed from k-space into images.Source: A. Super-Resolution Formulation · p. 2

B. Super-Resolution Approaches

  • SR methods divide broadly into model-based approaches that specify the degradation mapping and learning-based approaches that learn an LR-to-HR mapping.The survey frames both as ways to approximate the inverse mapping from LR observations to HR images.Source: A. Super-Resolution Formulation · p. 2

1) Model-Based:

  • Model-based SR expresses degradation through geometric transformation, blur, undersampling, and additive noise, then estimates the HR image by optimizing a likelihood-plus-prior objective.The stated forward-model assumption treats the transformation, blur kernel, and undersampling operator as known.Source: 1) Model-Based · pp. 2–3

2) Learning-Based:

  • Learning-based methods regress from LR to HR images, while DL methods learn priors and degradation models from paired data rather than requiring manual design.The described DL objective trains predictions against LR-HR pairs and applies the trained network to unseen LR images.Source: 2) Learning-Based · p. 3

C. Perspectives on DL-based Super-Resolution

1) End-to-End Data-Driven Perspective:

  • End-to-end data-driven SR directly maps LR images to HR counterparts, with performance depending on dataset scale, quality, diversity, and realism.Supervised, unsupervised, or self-supervised learning can be used depending on data context and availability.Source: 1) End-to-End Data-Driven Perspective · p. 3

2) Physics-Driven Perspective:

MRI super-resolution is presented as a computational response to acquisition trade-offs and as an image-to-image translation task from LR to HR images. The section surveys MRI-specific SR scenarios alongside physics-driven methods that combine imaging physics with learned priors.

2) Physics-Driven Perspective:

  • The image-to-image translation perspective treats SR as learning a mapping from LR images to HR counterparts with enhanced anatomical detail and visibility.The framework also addresses differences in resolution, contrast, modality, and imaging techniques.Source: 3) Image-to-Image Translation… · p. 3

A. Key Scenarios in MRI Super-Resolution; 1) In-Plane Super-Resolution:

2) Through-Plane Super-Resolution:

  • Slice-to-volume reconstruction (SVR) combines motion-affected anisotropic stacks from varying orientations to reconstruct an HR isotropic volume, jointly addressing motion correction and through-plane SR.The complementary spatial coverage of stacks is especially useful in fetal and neonatal imaging; a cited study also demonstrates single-stack DL SVR feasibility.Source: 3) Slice-to-Volume Reconstruction · p. 4, 2) Through-Plane Super-Resolution · p. 4

3) Slice-to-Volume Reconstruction:

  • Isotropic SR enhances volumes that are low resolution in all three dimensions, a poorly conditioned inverse problem because no axis retains high-frequency structural cues.Most approaches therefore use fully supervised learning with external HR isotropic datasets.Source: 4) Isotropic Volumetric Super-Resolution · p. 4

5) Super-Resolution Across Field Strengths:

  • Across field strengths, SR is described as narrowing the image-quality gap at low field and supporting portable, resource-limited MRI; the section also notes FDA-cleared low-field clinical systems.The cited examples include ultra-low-field and portable systems.Source: 5) Super-Resolution Across Field… · p. 4

6) Temporal Super-Resolution:

  • Temporal SR reconstructs temporally enhanced images from low-temporal-resolution data, aiming to improve dynamic-process fidelity without faster acquisition protocols.Applications include cardiac and functional imaging, 4D flow MRI, and respiration-resolved 4D MRI.Source: 6) Temporal Super-Resolution · p. 4

7) Multi-Contrast MRI Super-Resolution:; 8) Cross-contrast MRI Super-Resolution:

IV. PHYSICS-DRIVEN SUPER-RESOLUTION; A. Deep Plug-and-Play

  • Physics-driven SR combines imaging physics with learned priors; plug-and-play (PnP) replaces handcrafted MAP regularization with a pretrained denoiser in iterative optimization.PnP can work without ground-truth data, but repeated forward-model and denoiser execution adds computational and memory overhead, and λ and K require manual tuning.Source: §IV · p. 5, A. Deep Plug-and-Play · p. 5

B. Deep Unfolding

The section reviews physics-informed reconstruction methods alongside learning paradigms for MRI super-resolution, from paired-data supervision to unpaired and scan-specific approaches.

B. Deep Unfolding

  • Deep unfolding turns optimization iterations into trainable network layers, alternating data-consistency enforcement with prior-based refinement.This preserves an interpretable optimization structure and offers faster inference than traditional iterative methods.Source: B. Deep Unfolding · p. 5, Figure 3 · p. 6

C. Deep Equilibrium

  • Deep equilibrium models solve an implicit fixed point rather than unrolling a fixed number of iterations, enabling constant memory but not guaranteeing convergence.Implicit differentiation also adds computational overhead.Source: C. Deep Equilibrium · p. 6, Figure 4 · p. 6

V. LEARNING APPROACHES IN MRI SR

  • MRI SR methods are grouped by learning paradigms, learning strategies, and data-efficient or adaptive methods, which can overlap in one method.For example, a method can be both self-supervised and multi-task.Source: §V · p. 6

1) Supervised Learning:

  • Supervised MRI SR uses physically acquired, registered LR-HR pairs, but obtaining and aligning these scans is difficult and few studies use truly paired data.Motion, scanner variability, cost, and non-rigid anatomical changes complicate data collection and registration.Source: 1) Supervised Learning · p. 6, A. Learning Paradigms · p. 6

2) Unsupervised:

  • Unsupervised degradation-based methods synthesize LR scans from HR data, but simplified degradations can differ from clinical acquisitions and hinder generalization.K-space truncation is commonly used, while protocol, SNR, coil-sensitivity, and calibration differences contribute to the domain gap.Source: 2) Unsupervised · pp. 6–7, Figure 6 · p. 7
  • Domain-transfer methods learn mappings between unpaired LR and HR domains, including frameworks that correct LR inputs before upscaling without explicit degradation knowledge.Other approaches align both image and latent-feature spaces for unpaired or misaligned MRI data.Source: 2) Unsupervised · p. 7
  • Self-supervised MRI SR creates training pairs from the input scan, with methods supporting anisotropic-to-isotropic reconstruction and other MRI tasks without external supervision.Scan-specific degradation learning models input-specific degradation; later work combines scan-specific SR with an unsupervised transformer.Source: 3) Self-Supervised Learning · p. 7

3) Self-Supervised Learning:

This section surveys complementary MRI SR learning strategies and data-efficient approaches, then introduces loss formulations used to train reconstruction networks.

1) Multi-Task Learning:

  • Multi-task MRI SR jointly trains related tasks, including denoising, reconstruction, motion correction, artifact removal, segmentation, and registration.The approach leverages shared representations across tasks.Source: 1) Multi-Task Learning · p. 8

2) Multi-Modal MRI Learning:

  • Multi-modal MRI SR fuses information across contrasts, such as using high-frequency T1 structure to guide T2 reconstruction.Proposed approaches include multiscale fusion, guided filtering, and Transformer-based contextual matching across contrasts.Source: 2) Multi-Modal MRI Learning · p. 8

3) Curriculum Learning:

  • Curriculum learning introduces increasingly difficult training tasks, which can improve robustness to high scaling factors, noise, and blurring.The stated outcome is better reconstruction quality and generalization.Source: 3) Curriculum Learning · p. 8

4) Prompt Learning:

  • Prompt learning uses contextual cues such as degradation characteristics to guide reconstruction, while proposed multimodal prompts use stable-diffusion priors in image restoration.Using scan parameters and acquisition protocols as MRI prompts is described as a potential application, not an existing MRI method.Source: 4) Prompt Learning · p. 8

5) Ensemble Learning (Multi-Supervision)

  • MRI SR ensembles aggregate networks trained with different priors, architectures, or datasets, while federated learning supports collaborative training without sharing raw patient data.RepFL combines perturbed client-data replicas and heterogeneous models for graph super-resolution, reporting better performance than conventional federated methods.Source: 5) Ensemble Learning (Multi-Supervision) · p. 8, 6) Federated Learning · p. 8

6) Federated Learning:

  • Reinforcement-learning SR uses iterative resolution-enhancement decisions but is less common, while mixture-of-experts models route inputs to specialized subnetworks.Defining RL states, actions, and rewards is challenging; MoE-DiffSR uses gated experts with a diffusion backbone on multi-scanner brain and knee datasets.Source: 7) Reinforcement Learning Models (RLs) · p. 8, 8) Mixture of Experts (MoE) · p. 8

C. Data-Efficient and Adaptive Methods

  • Data-efficient approaches include transfer learning, contrastive losses for multi-contrast texture realism, and meta-learning that adapts to new imaging tasks with minimal data.Delta-WKV dynamically adjusts weights at inference and is reported to achieve state-of-the-art MRI SR quality while reducing computational cost.Source: 1) Transfer Learning and Zero-Shot · p. 8, 3) Meta learning · p. 9

A. Loss Functions

  • Voxel-wise losses compare image intensities, whereas feature-based perceptual and style losses encourage similarity in representations from a pretrained network.The ℓ2 loss penalizes larger errors more severely than ℓ1; the survey reports ℓ1's superior performance across different restoration tasks and notes that ℓp losses can smooth high-frequency detail.Source: 1) Pixel-/Voxel-wise Loss · pp. 9–10, §VI · p. 9, 2) Feature-wise Loss · p. 10

2) Feature-wise Loss:

The section describes feature-wise losses for perceptual and style fidelity, adversarial and cycle-consistent training, and prior-based constraints that guide MRI SR reconstructions.

2) Feature-wise Loss:

  • Perceptual loss compares high-level image representations to encourage outputs that resemble targets in human perception.Style loss uses Gram-matrix features to encourage preservation of textures and other style attributes.Source: 2) Feature-wise Loss · p. 10

3) GAN-based Loss:

  • GAN adversarial loss trains a generator to produce SR images that a discriminator judges as real, but GANs can suffer mode collapse and convergence problems.The section also describes cycle consistency as constraining mappings in both directions to help maintain content fidelity in unpaired restoration.Source: 3) GAN-based Loss · p. 10

4) Prior-based Loss:

  • Total variation loss penalizes neighboring-pixel differences to promote piecewise smoothness, though volumetric MRI extensions pose numerical stability challenges.Practical 3D TV implementations are reported for fetal brain imaging, multi-contrast brain MRI, and self-supervised fMRI SR.Source: 4) Prior-based Loss · p. 10
  • Physics-based losses compare degraded reconstructions with observed measurements to enforce consistency with MRI acquisition principles.The degradation operator is applied to the reconstructed image before comparison with the observation.Source: 4) Prior-based Loss · p. 10
  • Histogram loss matches predicted and target intensity distributions, while edge-preserving loss penalizes differences between their edge maps.The section identifies histogram distances such as L2 or KL divergence and notes the value of preserving anatomical boundaries.Source: 4) Prior-based Loss · pp. 10–11
  • Metric-based losses optimize image-quality criteria such as SSIM rather than relying only on pixel-wise differences.The ideal reference value for SSIM is 1.Source: 5) Metric-based · p. 11

5) Metric-based:

The section surveys metric-based training objectives and architectural choices for MRI SR, from upsampling and feature reuse to attention mechanisms.

5) Metric-based:

  • Metric-based losses optimize perceptual or evaluation-driven criteria such as SSIM, whose ideal reference value is 1.The formulation compares the metric for the prediction and target with its ideal reference value.Source: 5) Metric-based · p. 11

B. Fundamental Network Architectures

  • Upsampling may occur before the network, at its output, or progressively through intermediate layers, and architectures also use residual, dense, or recursive connections.Fig. 7 depicts these arrangements, including long skip connections, feature reuse, and shared layers.Source: 1) Up-sampling Methods · p. 11

1) Up-sampling Methods:

  • Interpolation is simple and efficient but cannot recover high-frequency information, while learned upsampling can recover details and textures.At large scaling factors, interpolation tends to produce overly smooth or blurred results.Source: 1) Up-sampling Methods · p. 11
  • Transposed convolution adaptively expands feature maps but can create checkerboard patterns from uneven overlaps.Sub-pixel layers rearrange convolution-generated channels into HR outputs and can be more efficient than transposed convolutions, though uneven receptive-field distribution can cause artifacts.Source: 1) Up-sampling Methods · p. 11
  • Meta-upscaling predicts convolution weights for each HR position, allowing one model to support arbitrary scaling factors.This avoids separate models for predefined scales.Source: 1) Up-sampling Methods · p. 12

2) Residual Learning:

  • Global residual learning predicts the LR-to-HR residual, while local shortcuts support deeper models by mitigating vanishing gradients and degradation.The two forms can be combined to recover high-frequency details and manage complex transformations.Source: 2) Residual Learning · p. 12

3) Dense Connections:

  • Dense connections reuse earlier features and improve gradient flow, while recursive learning reapplies shared modules to limit parameter counts.Dense designs can incur computational overhead; recursive designs can face gradient and computation challenges, for which residual learning is one mitigation.Source: 3) Dense Connections · p. 12, 4) Recursive Learning · p. 12

5) Attention Mechanisms:

  • Attention mechanisms reweight informative features across channels, spatial locations, time, and relations, with specialized designs for each domain.Examples include channel attention, spatial-channel fusion, long-range self-attention, temporal attention for dynamic MRI, cross-attention across modalities or scales, and graph attention.Source: 5) Attention Mechanisms · p. 12

C. Networks

MRI SR architectures span CNNs, long-range dependency models, generative approaches, continuous representations, and recurrent designs, each offering distinct modeling strengths and trade-offs.

1) Convolutional Neural Networks (CNNs):

  • CNN architectures including U-Net, ResNet, and DenseNet underpin MRI SR, with adaptations supporting detail preservation, contextual modeling, and artifact reduction.Examples include dilated convolutions for wider receptive fields and spatio-temporal residual learning for multicontrast, 3D, and dynamic MRI.Source: C. Networks · p. 13, 1) Convolutional Neural Networks (CNNs) · p. 13

2) Long-Range Dependency Models:

  • ViTs, state-space models, and recurrent-attention hybrids model long-range dependencies, with CNN combinations supplying local detail and spatial inductive biases.State-space models offer linear scaling with sequence length, while the section characterizes CNN-plus-global-context designs as a dominant MRI SR choice.Source: 2) Long-Range Dependency Models · p. 13

3) Generative Adversarial Networks (GANs):

  • GANs use adversarial training to produce perceptually realistic MRI SR outputs, but training instability, mode collapse, and computational overhead remain challenges.The survey describes MRI applications spanning slice-thickness and volumetric SR, frequency-aware designs, and temporal models.Source: 3) Generative Adversarial Networks… · pp. 13–14

4) Diffusion Models (DMs):

  • Diffusion models iteratively reverse a stochastic noising process and are presented as stable, controllable alternatives that can generalize under noise, motion, or low SNR.MRI SR examples use latent priors, partial sampling, and residual shifting; Res-SRDiff is reported to reconstruct in as few as four steps on ultra-high-field brain and prostate MRI.Source: 4) Diffusion Models (DMs) · p. 14

5) Implicit Neural Representations (INRs):

  • Implicit neural representations map coordinates and optional features continuously to signal intensities, enabling HR reconstruction through queries at arbitrary locations.They can avoid fixed-scale upsampling modules, reduce memory use, and represent slices or full volumes.Source: 5) Implicit Neural Representations… · p. 14, Figure 8 · p. 14
  • MRI applications of INRs include volumetric and multi-contrast SR, super-resolution of dynamic acquisitions, and patient-specific representations.Cycle-INR uses local attention-based grid sampling and cycle consistency to address over-smoothing and inter-slice inconsistency in anisotropic 3D images.Source: 5) Implicit Neural Representations… · p. 14

6) Gaussian Splatting (GS):

  • Gaussian splatting represents pixels as spatially adaptive Gaussians, yielding continuous fields that can be evaluated smoothly at arbitrary locations.The survey notes that GS has been used for MRI reconstruction but had not yet been introduced for MRI SR.Source: 6) Gaussian Splatting (GS) · pp. 14–15, Figure 9 · p. 15

C. Networks

  • Recurrent networks model frame-to-frame dependencies for dynamic MRI, supporting temporal consistency and reducing frame-to-frame artifacts.Examples include real-time interventional MRI and self-supervised dynamic fetal MRI SR without paired HR data.Source: 7) Recurrent Neural Networks (RNNs) · p. 15

7) Recurrent Neural Networks (RNNs):

The section reviews recurrent and graph-based MRI SR, foundation-model strategies, and generative AI, then outlines diffusion formulations and conditional DDPM methods.

7) Recurrent Neural Networks (RNNs):

VII. FOUNDATION MODELS

  • Foundation models are proposed as adaptable MRI SR systems that draw on broad data-derived priors, semantic conditioning, and multimodal information.The survey highlights anatomically guided multi-task enhancement and graph-based multimodal semantic modeling as representative paradigms.Source: §VII · p. 15, A. Recent Foundation Models for MRI… · p. 15

1) Anatomically Guided Multi-Task Enhancement (BME-X):

  • BME-X predicts soft tissue labels and conditions HR reconstruction on them, using synthetic aligned pairs from more than 13,000 volumes across 19 public datasets.It is reported to be robust to motion-corrupted infant scans and cross-domain scenarios such as 3T→7T SR, with improved SR fidelity and more consistent downstream segmentation and registration.Source: 1) Anatomically Guided Multi-Task… · pp. 15–16

2) Graph-Based Multimodal Semantic Modeling (GraphMSR):

  • GraphMSR combines adaptive image-patch graphs and hierarchical graph attention with medical-language-model text embeddings fused through cross-attention.Its semantic priors include anatomical, image-quality, and pathological descriptors, used to emphasize diagnostically important regions and suppress artefacts.Source: 2) Graph-Based Multimodal Semantic… · p. 16

B. Comparison and Direction

  • BME-X uses explicit tissue priors while GraphMSR uses semantic-graph reasoning; foundation-model errors or biases may propagate to downstream models.The survey frames the two systems as distinct foundation-model strategies for MRI SR.Source: B. Comparison and Direction · p. 16

VIII. GENERATIVE AI

  • Generative AI covers GANs, VAEs, normalizing flows, and diffusion models, which differ in output characteristics and computational demands.The survey associates GANs with sharp outputs but possible hallucinations and instability, VAEs with uncertainty but smooth textures, flows with memory-intensive architectures, and diffusion with fidelity and robustness at high computational cost.Source: §VIII · p. 16

A. Diffusion Models (DMs) for Super-Resolution

  • Diffusion models iteratively reverse a noising process to generate outputs, and MRI SR conditions the reverse process on the observed LR image.The section lists DDPM, score-based, SDE-based, DDIM, and latent diffusion formulations.Source: A. Diffusion Models (DMs) for… · p. 16, Figure 10 · p. 17

1) Denoising Diffusion Probabilistic Models (DDPMs):

  • Conditional DDPMs add Gaussian noise in the forward process and learn an LR-conditioned reverse denoising process that estimates noise rather than directly recovering the clean image.The survey describes MRI adaptations for brain and cardiac SR, including partial diffusion to reduce sampling steps and multi-contrast methods using diffusion priors or modality-specific features.Source: 1) Denoising Diffusion Probabilistic… · pp. 16–17

2) Score-Based Generative Models (SGMs):

The section surveys conditional score and diffusion methods for MRI super-resolution, spanning continuous-time sampling, accelerated inference, and pixel, latent, and frequency-domain formulations. It also describes posterior-specialized diffusion models and mixture-of-experts approaches.

2) Score-Based Generative Models (SGMs):

  • Conditioning score-based sampling on the LR image supports HR reconstruction consistent with measured data, using Langevin dynamics or probability-flow ODEs.MRI examples tailor score learning, recover high frequencies, use wavelet decomposition, or combine ensemble denoisers with spatial self-consistency.Source: 2) Score-Based Generative Models (SGMs) · p. 17

3) Stochastic Differential Equation (SDE)-Based Diffusion Models:

  • SDE-based diffusion formulates the forward process continuously, then approximates the conditional score to support reverse-time sampling with stochastic solvers or deterministic probability-flow ODEs.The framework is presented as useful for uncertainty, adaptive priors, and continuous-time regularization in MRI reconstruction.Source: 3) Stochastic Differential Equation… · pp. 17–18

4) Denoising Diffusion Implicit Models (DDIMs):

  • DDIMs preserve DDPM marginal distributions while allowing sparse-timestep sampling, with η = 0 producing a deterministic process.MRI SR uses DDIMs to accelerate inference; examples include latent sampling, skipping early steps, and residual-based initialization.Source: 4) Denoising Diffusion Implicit Models… · p. 18, Figure 11 · p. 18

5) Latent Diffusion Models (LDMs):

  • Latent diffusion applies denoising in a pretrained autoencoder’s compact representation, reducing computation relative to pixel-space generation and enabling efficient 3D MRI reconstruction.The section describes applications to brain SR, diagnostic-feature preservation, and arbitrary zoom factors; latent diffusion can also combine with DDIM step reduction.Source: 5) Latent Diffusion Models (LDMs) · p. 18

6) Frequency-Domain Diffusion Models:

  • Frequency-domain diffusion corrupts and denoises transformed MRI data, targeting high-frequency components important for diagnostic detail.Wavelet DDPM and frequency-aware SGM examples suppress noise and artifacts or improve perceptual quality and sampling efficiency; these approaches can combine with latent or DDIM acceleration.Source: 6) Frequency-Domain Diffusion Models · p. 18
  • Region-aware mixture-of-experts diffusion routes anatomy to specialized denoisers, aiming to preserve fine detail and improve robustness across contrasts and sites.Expert calibration adds complexity and may introduce boundary artifacts.Source: 7) Mixture of Diffusion Models (MDMs) · p. 18

7) Mixture of Diffusion Models (MDMs):

The section explains how diffusion priors can support inverse-problem reconstruction without paired training data, then contrasts VAEs, normalizing flows, and GANs. It closes with clinical and computational challenges of generative MRI super-resolution.

8) Diffusion Models for Imaging Inverse Problems via Posterior Estimation:

  • Posterior sampling combines a learned denoising prior with measurement fidelity to approximate p(x | y), enabling inverse-problem reconstruction without retraining when paired LR-HR data are unavailable.The figure distinguishes pixel, latent, and frequency diffusion spaces; iterative methods alternate denoising and measurement-consistency steps.Source: 8) Diffusion Models for Imaging… · pp. 18–19, Figure 12 · p. 19

1) Variational Autoencoders (VAEs):

  • VAEs encode HR targets and LR inputs into a probabilistic latent representation, then decode HR images; ELBO training balances reconstruction fidelity and latent regularization.VAEs offer fast inference and structured latent spaces, but Gaussian decoders can over-smooth textures and VAEs generally lag diffusion in perceptual quality and fine-detail reconstruction.Source: 1) Variational Autoencoders (VAEs) · p. 19, Figure 13 · p. 19

2) Normalizing Flow Models (NFMs):

  • Conditional normalizing flows model p(x | y) through invertible transformations, enabling exact likelihood estimation and diverse HR outputs from sampled latent variables.MRI examples include anatomically consistent, diverse reconstructions and improved fidelity across contrasts; deep architectures can raise memory and computational costs.Source: 2) Normalizing Flow Models (NFMs) · p. 19, Figure 14 · p. 20

3) Generative Adversarial Networks (GANs):

C. Challenges with GenAI

  • Generative MRI SR models face hallucination risks and trade-offs: GANs may collapse modes, VAEs may blur detail, and diffusion models are less interpretable and directly controllable.Most generative SR models are designed for 2D; 3D extension raises memory demands and requires spatial continuity across volumes.Source: C. Challenges with GenAI · p. 20

IX. PERFORMANCE EVALUATION

The section reviews subjective, objective, and task-based assessment of super-resolved MRI, noting that no single measure captures every relevant quality dimension. It highlights MRI-specific IQA needs and the tension between distortion fidelity and perceptual quality.

A. Image Quality Assessments

1) Peak Signal-to-Noise Ratio (PSNR):

2) Structural Similarity Index (SSIM):

  • SSIM evaluates luminance, contrast, and structure over local windows, while multi-scale SSIM assesses quality across resolutions.Its components can be weighted to adjust their relative importance.Source: 2) Structural Similarity Index (SSIM) · p. 21

3) Visual Information Fidelity (VIF):

  • VIF estimates preserved visual information relative to a reference using natural-scene and human-vision models, and is reported to correlate more strongly with subjective evaluations than PSNR or SSIM.Its scores typically range from 0 to 1, with higher values indicating better perceptual quality.Source: 3) Visual Information Fidelity (VIF) · p. 21

4) Contrast-to-Noise Ratio (CNR):

5) Task-based Evaluation:

7) Advancing MRI-Specific IQA:

  • MRI-specific no-reference IQA includes a deep-learning metric aligned with radiologists’ ratings and a fetal brain mask inlier fraction concordant with experts.The latter enabled adaptive scan termination; clinical validation and multi-center applications are also discussed.Source: 7) Advancing MRI-Specific IQA · p. 22
  • A 35-metric benchmark using 14,700 expert scores found that IQA performance varies with network architecture and some perceptual metrics align closely with radiologists.Other work found many no-reference metrics poorly reflect clinical relevance, especially localized anatomical errors and downstream suitability; distortion and perceptual measures can trade off, motivating balanced evaluation.Source: 7) Advancing MRI-Specific IQA · p. 22, 8) Perception-Distortion Trade-off · p. 22, Figure 16 · p. 22

8) Perception-Distortion Trade-off:

The section organizes MRI SR datasets by modality, anatomy, and field strength, and identifies public repositories that could broaden evaluation. It highlights data diversity as useful for examining generalization, domain shift, and reliability.

B. Benchmark Datasets

  • Benchmark datasets are organized by imaging modality, anatomical focus, and field strength, with MRI sequences spanning structural, diffusion, angiographic, and functional imaging.Examples include T1w, T2w, PD, FLAIR, DWI/DTI, MRA, and fMRI.Source: B. Benchmark Datasets · p. 23

1) fastMRI:

  • fastMRI provides fully sampled and retrospectively undersampled multi-coil knee and brain k-space data for reconstruction, acceleration, and super-resolution research.Its sequences include knee PD and PD-FS and brain T1w, T2w, and FLAIR.Source: 1) fastMRI · p. 23

2) IXI:

  • IXI contains nearly 600 healthy-adult MRI scans from three London hospitals at 1.5T and 3T, while HCP provides about 1,200 healthy adults’ structural and functional MRI data at 3T and partly 7T.IXI includes T1w, T2w, PD, DWI, and MRA; HCP includes T1w, T2w, fMRI, and DWI.Source: 2) IXI · p. 23, 3) HCP · p. 23

3) HCP:

  • dHCP provides neonatal structural and functional MRI from 558 subjects at 3T using motion-robust protocols and a dedicated neonatal head coil.NAMIC provides multimodal brain MRI from over 20 healthy subjects at 1.5T and 3T, and supports super-resolution research.Source: 4) dHCP · p. 23, 5) NAMIC · p. 23

6) CRL:

  • CRL’s fetal brain atlas contains motion-robust in-utero T2w scans spanning 21–38 gestational weeks, while its EPI collection contains 111 fetal MRI sessions spanning 25–35 weeks.The EPI data support motion-correction and fetal fMRI research.Source: 6) CRL · p. 23, 7) EPI · p. 23

8) BraTS:

  • BraTS supplies multimodal glioma MRI with expert tumor-subregion annotations; other public repositories remain underused despite offering data diversity for SR generalization and domain-shift evaluation.Examples include OpenNeuro, MSD, ADNI, and TCIA, with multi-site or heterogeneous acquisitions and, in some cases, clinical or genomic information.Source: 8) BraTS · p. 23, 9) Other Public Repositories Suitable… · pp. 23–24

X. GLOBAL ACCESS AND CLINICAL APPLICATIONS

The section connects MRI super-resolution to access and clinical applications while reviewing limits, inverse-imaging relationships, uncertainty, and practical challenges in real data and model design.

A. Technology Distribution, Economic Significance, and the Importance of Super-Resolution

  • 65,000 MRI scanners worldwide correspond to roughly 7 per million; high-income countries average over 25, while low- and middle-income countries have around 1.Approximately two-thirds of the global population has little or no MRI access; ultra-low-field systems such as 0.055T paired with SR are described as a cost-effective option.Source: A. Technology Distribution, Economic… · p. 24

B. Applications of MRI Super-Resolution

  • MRI SR spans medical-imaging applications and can enhance spatial resolution without increasing acquisition time, supporting anatomical visualization, diagnostic accuracy, and quantitative analysis.Table VI is identified as summarizing representative application areas, but the supplied passage does not enumerate them.Source: B. Applications of MRI Super-Resolution · p. 24, Table VI · p. 25

A. Theoretical Limits of Bandwidth Extrapolation in Super-Resolution

  • Classical SR recovery beyond sampling limits is unstable without strong priors, while learning-based methods can synthesize plausible high-frequency detail that may not be strictly accurate.Even minimal noise can cause unbounded amplification in bandwidth extrapolation; under sparsity assumptions, partial SR is possible, but stability worsens rapidly as the resolution factor increases.Source: A. Theoretical Limits of Bandwidth… · p. 24

B. Denoising, Deblurring, Inpainting, and Super-Resolution: A View of Inverse Imaging

  • Denoising, deblurring, inpainting, and SR share an inverse-imaging goal, but SR is generally the most ill-posed because its degradation effects are compounded.The degradation operator can combine spatial transformation, blur, downsampling, and noise; the active components determine the inverse problem.Source: B. Denoising, Deblurring, Inpainting,… · pp. 24–25

C. Uncertainty Quantification in Super-Resolution

  • Bayesian uncertainty estimation, GAN-based uncertainty modeling, and equivariant bootstrapping are among approaches discussed for improving the interpretability and robustness of SR outputs.The integration of uncertainty into DL-based SR remains an open challenge.Source: C. Uncertainty Quantification in… · p. 25

A. Challenges with Real Data and Solutions

  • Real-world MRI SR faces imperfect or missing ground truth and variable image quality and scale, prompting strategies such as reference-free training, misalignment-aware models, domain adaptation, and continuous-scale SR.These challenges hinder direct application of supervised learning.Source: A. Challenges with Real Data and… · p. 25, Table VII · p. 26

B. Deviations from General Deep Learning Practices

  • Batch normalization may impair fine-texture preservation and introduce artifacts in SR, so many recent architectures omit it to retain fidelity and reduce memory use.Omitting batch normalization can also allow deeper or wider network designs.Source: 1) Batch Normalization (BN) · p. 25

2) Dropout:

  • Dropout can disrupt fine-detail learning, though selective use in final layers has improved performance under multi-degradation settings, while standard dropout has also been reported to degrade high-frequency reconstruction.Kong et al. report the selective-use result; Xu et al. report the degradation and propose an adaptive variant for SR.Source: 2) Dropout · pp. 25–26

C. Efficiency

The section reviews efficiency and preprocessing considerations, then surveys open resources and research directions spanning data, learning, network design, generative models, and interpretability.

C. Efficiency

D. Preprocessing

  • Preprocessing needs depend on acquisition details, anatomy, and SR objective; steps such as denoising or smoothing can erase details SR is meant to reconstruct.Bias fields, misalignments, distortions, and segmentation choices also need consideration in relation to the study goal.Source: D. Preprocessing · p. 26

E. Open-Source and Proprietary Software

  • The survey provides a curated list of open-source and proprietary MRI SR tools, implementations, datasets, and official repositories through its GitHub resource.The conclusion also notes that open-source resources were curated to facilitate further research.Source: E. Open-Source and Proprietary Software · p. 26, §XIII · p. 27

XIII. CRITICAL OUTLOOK AND FUTURE DIRECTIONS

  • The survey synthesizes MRI SR across computer vision, computational imaging, inverse problems, and MR physics, with a taxonomy spanning supervision, architecture, and physics coupling.It also consolidates learning strategies, image-quality assessment, benchmark datasets, and deployment considerations.Source: §XIII · p. 27

A. Learning Approach

  • Paired LR–HR data, reliable ground truth, training samples, and cross-site or protocol shifts constrain real-world MRI SR; unsupervised, self-supervised, transfer, and few-/zero-shot approaches remain directions for addressing these constraints.The passage states that transfer and few-/zero-shot adaptation have received limited attention.Source: A. Learning Approach · p. 27

B. Network Design

  • Future network design should balance fidelity, robustness, efficiency, and interpretability rather than optimizing only PSNR/SSIM, which may not reflect clinical utility.The section also identifies regression to the mean under ℓp-norm losses as a source of overly smoothed reconstructions and potential loss of clinically relevant fine details [231].Source: B. Network Design · p. 27

C. Generative AI

  • Extending generative AI's 2D MRI SR gains to 3D remains difficult because of data dimensionality, GPU memory demands, and the need for spatial coherence; hallucinations also threaten diagnostic reliability.Suggested directions include volumetric consistency and combining generative realism with physics-based constraints –[481].Source: C. Generative AI · pp. 27–28

D. Foundation Models

  • Foundation models may support generalization across sites, vendors, and protocols, while physics-aligned interpretability, calibrated uncertainty, and failure detection are proposed to support transparency and regulatory approval.The passage notes that foundation-model application to MRI SR remains limited; paired LR–HR acquisitions can also capture realistic degradation patterns beyond simple downsampling or k-space truncation.Source: D. Foundation Models · p. 28, E. Interpretability · p. 28, 1) Data Acquisition · p. 28

1) Data Acquisition:

This section describes data and benchmarking needs, MRI-physics considerations, application areas, and proposed competition and clinical-validation practices for MRI SR.

1) Data Acquisition:

  • Paired LR–HR acquisitions can capture realistic noise, motion, and system-specific degradation beyond simplistic downsampling or k-space truncation; diversity in anatomy, pathology, and scanners can enhance generalization and clinical relevance.Curation, annotation, quality control, standardization, and metadata are also described as important to dataset clinical value.Source: 2) Data Curation · p. 28

G. Benchmarking

  • Preprocessing differences can produce discrepant results even on the same datasets, motivating standardized benchmarks with diverse data, harmonized preprocessing, and documented degradation models.The proposed benchmark scope includes anatomies, pathologies, scanners, and acquisition settings.Source: G. Benchmarking · p. 28

H. MR Physics in Forward Model

  • MRI slice profiles are often non-rectangular and can cause through-plane blurring and partial-volume effects; modeling them in the degradation operator can improve physical consistency, especially for through-plane SR and slice-to-volume reconstruction.The blur can be decomposed into an in-plane point spread function and a through-plane slice-profile kernel.Source: H. MR Physics in Forward Model · p. 28

I. Application to Diffusion MRI (dMRI)

  • Diffusion MRI SR must preserve diffusion-specific information, with proposed directions including physics-aware models, task-driven objectives, and cross-domain constraints from structural MRI or biophysical models.The passage identifies tractography, connectomics, and microstructural biomarker estimation as downstream tasks.Source: §I · p. 28

J. Current and Future Clinical applications

K. Competitions

  • No dedicated MRI SR competition platform exists, limiting standardized evaluation and reproducibility; proposed tracks cover methods, image-quality assessment, benchmark datasets, and preprocessing standards.The proposed parallel tracks could yield reference implementations, standardized pipelines, and reproducible baselines.Source: K. Competitions · p. 29

L. Clinical Validation

  • Current reliance on PSNR and SSIM may miss clinical utility, and most algorithms are tested on synthetic rather than real data; validation should include clinical data and task-relevant assessment.Proposed approaches include multi-center blinded reader studies, downstream diagnostic tasks, hallucination checks with expert review, and standardized public validation benchmarks.Source: L. Clinical Validation · pp. 29–30

Related papers

Loading 2511.16854v3…