MRI Super-Resolution with Deep Learning: A Comprehensive Survey

TL;DR

This survey synthesizes recent deep learning advances in MRI super-resolution, framing SR as an ill-posed inverse problem and categorizing methods by model-based, data-driven, and physics-informed perspectives. It covers a spectrum of SR scenarios (in-plane, through-plane, SVR, isotropic, across field strengths, temporal, multi-contrast, cross-contrast, and cross-modality), then delves into physics-driven frameworks (PnP, unfolding, equilibrium) and learning strategies (supervised, unsupervised, self-supervised; multi-task, multi-modal, curriculum, prompts, ensembles, FL, RL, MoE). The review catalogs network architectures (CNNs, Transformers, GANs, diffusion models, INRs, Gaussian splatting, GCNs), foundation models, and generative AI for SR, plus evaluation standards, datasets, and clinical considerations. It emphasizes challenges in real data, data efficiency, interpretability, and clinical validation, and it outlines future directions including foundation models, diffusion-based priors, and standardized benchmarks to accelerate clinical translation.

Abstract

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.

Figures (16)

• Figure 1: Taxonomy of MRI super-resolution methods.
• Figure 2: Perspectives on MRI super-resolution methods: (a) Data-driven approach that learns LR-to-HR mapping purely from data; (b) Physics-informed approach that incorporates the underlying imaging physics while mapping the LR-to-HR; and (c) Image-to-image translation approach that translates across LR and HR domains, such as modalities or contrasts.
• Figure 3: Deep unfolding models provide a physics-informed approach to imaging inverse problems by integrating the forward model with a neural network within an iterative reconstruction framework, as illustrated in Algorithm \ref{['alg:deep_unfolding']}. Each unrolled layer corresponds to a single iteration of an optimization algorithm zhang2020deepmonga2021algorithm.
• Figure 4: Deep Equilibrium models provide a physics-driven approach to imaging inverse problems by incorporating the forward model and a neural network within a fixed-point formulation gilton2021deep.
• Figure 5: Learning paradigms for MRI super-resolution: (a) Supervised learning, where the network is trained on physically acquired and well-aligned pairs of LR-HR scans; (b) Unsupervised learning, where no LR-HR pairs are available: 1) generating synthetic LR images from HR scans to create LR-HR training pairs, and 2) scenarios where only LR scans and unpaired examples from the HR target domain are accessible; and (c) Self-supervised learning, where training pairs are generated directly from the available LR scans.