In medical imaging modalities like Magnetic Resonance Imaging (MRI) and Computed Tomography (CT), clinicians face a fundamental physical bottleneck: the trade-off between acquisition time and image quality. To capture fine anatomical structures or map complex physiological processes, scanners traditionally need to collect a massive volume of raw data points. However, keeping a patient inside a scanner for an extended period is costly, uncomfortable, and highly susceptible to patient motion, which introduces blurring and artifacts that degrade diagnostic utility.
To break this bottleneck, a prominent research direction focuses on computational reconstruction from accelerated scans. Instead of physically capturing every data point required by classical signal processing theorems, modern techniques intentionally undersample the data—drastically cutting scan times—and use advanced algorithmic priors to accurately fill in the missing information.
This approach works because medical images are not random noise; they contain massive structural redundancies. By exploiting mathematical sparsity, cross-modality correlations, and deep generative priors, these methods trade physical scanning time for computational inference time.
Exploiting Sparsity and Optimal Sampling
The foundation of accelerated imaging lies in the realization that anatomical structures can be represented compactly. In classical signal processing, the Nyquist-Shannon theorem dictates uniform sampling at a rate proportional to the highest frequency. However, if an image is sparse in a transform domain (such as wavelets or gradients), it can be perfectly reconstructed from far fewer measurements using compressed sensing.
To maximize the efficiency of compressed sensing, the sampling patterns themselves must be optimized. Matthias W. Seeger et al. demonstrated that a scalable Bayesian experimental design framework can automatically discover "irregular" scan sequences. Rather than sampling data on a rigid grid, these irregular patterns prioritize the most informative regions of raw data space, producing high-fidelity images in a fraction of the time.
Traditional Grid Sampling (Slow) Optimized Irregular Sampling (Fast)
X X X X X X X . X . . X
X X X X X X ======> . X . X . .
X X X X X X X . . . X X
When these optimized sampling strategies are paired with sparse reconstruction algorithms, the acceleration gains are clinically transformative. For instance, Hussein Rayatzadeh et al. applied compressed sensing to cardiac MRI, achieving a 3-fold acceleration that allowed high-resolution 3D imaging of the heart. Building on this, Orestis Vardoulis et al. showed that highly accelerated compressed sensing can compress a full 3D left atrial volume scan into a single patient breath-hold, eliminating respiratory motion artifacts entirely.
This principle extends beyond MRI to other modalities like CT. Guangming Zang et al. introduced a proximal optimization framework that incorporates a 3D Structure Tensor Prior, enabling the reconstruction of thin, manifold-like structures from highly sparse projection angles, effectively surpassing the physical resolution limits of the scanner hardware.
Borrowing Structure via Multi-Contrast Learning
In a typical clinical workflow, patients rarely receive just one sequence; they are scanned across multiple modalities or contrasts (e.g., T1-weighted and T2-weighted MRI). A T1 scan might be fast to acquire at high resolution, whereas a T2 scan provides critical diagnostic contrast but requires a much longer acquisition time.
Instead of treating these scans as independent problems, multi-contrast super-resolution frameworks use the high-resolution scan as an anatomical "cheat sheet" to reconstruct the accelerated, low-resolution scan. Because both scans capture the exact same underlying anatomy, their edge boundaries, tissue interfaces, and structural shapes are highly correlated.
[Fast Scan: High Res T1] ----> Extract Structural Edges \
===> [Reconstructed High Res T2]
[Slow Scan: Low Res T2] ----> Extract Target Contrast /
Early deep learning approaches formalized this using sparse coding. Ling Shao et al. developed a framework called WEENIE that uses joint convolutional sparse coding to simultaneously perform super-resolution and cross-modality synthesis from weakly supervised data.
To capture more complex, non-linear dependencies between contrasts, researchers shifted toward deep progressive architectures. Qing Lyu et al. designed a multi-contrast network that integrates features progressively across a high-dimensional latent space, ensuring that shared structural details are transferred without distorting the target scan's unique contrast characteristics. Similarly, Chun-Mei Feng et al. developed the Multi-Stage Integration Network (MINet), which explicitly models the multi-contrast relationships across multiple hidden layers to sharpen blurry, rapidly acquired scans.
This cross-modality guidance is particularly potent in diffusion MRI, where mapping tissue microstructure normally requires hundreds of time-consuming angular measurements. Jian Cheng et al. demonstrated that using 3D tissue probability maps derived from a standard T1 scan can guide the reconstruction of highly undersampled diffusion data, achieving a 15x acceleration and making advanced neuroimaging clinically viable.
Learning Global Context and Self-Similarity
While multi-contrast methods rely on external references, an image also contains massive internal redundancy. Anatomical structures exhibit non-local self-similarity—meaning a patch of tissue in one part of an organ often shares patterns, textures, or symmetries with patches elsewhere in the same scan. Traditional interpolation algorithms fail because they only look at immediate, local pixel neighbors, which smooths out high-frequency details.
To exploit these internal patterns, Yinghua Li et al. combined overcomplete dictionaries with non-local similarity constraints to classify image patches and reconstruct fine brain structures from repeating internal motifs. Modern neural networks scale this intuition by using attention mechanisms to look across the entire image volume. Xiaobin Hu et al. proposed a Feedback Graph Attention Convolutional Network (FB-GACN) that maps image patches as nodes in a graph, allowing the network to leverage long-range anatomical symmetries to strip away noise and motion artifacts.
To reason about these relationships more abstractly, Yulun Zhang et al. developed the Squeeze and Excitation Reasoning Attention Network (SERAN). SERAN "squeezes" global image components into distinct visual primitives, uses a graph network to reason about how these primitives interact across the entire scan, and then uses that global understanding to reconstruct sharp, localized details.
[Noisy/Low-Res Input]
│
▼
[Squeeze to Primitives] ──► [Graph Reasoning over Global Context]
│ │
▼ ▼
[Local Feature Extraction] ◄───────────────┘ (Adaptive Excitation)
│
▼
[Sharp, High-Res Output]
When clean training data is entirely unavailable, generative models can serve as pure statistical priors to separate anatomy from noise. Tiange Xiang et al. introduced DDM², a self-supervised framework that utilizes generative diffusion models. By mapping a noisy, accelerated scan to an intermediate state in a diffusion Markov chain, the model leverages its learned geometric priors to denoise the image without ever seeing a "ground truth" clean scan.
Direct Parameter Mapping
This shift from analytical reconstruction to deep-learning-driven estimation also redefines quantitative imaging, where scanners measure physical tissue properties (like $T_1$ or $T_2$ relaxation times) across multiple sequential images. Instead of performing slow scans to get clean data for traditional curve-fitting, single-author works have shown that networks can directly map parameters from highly accelerated, noisy data. Patrick J. Bolan demonstrated this by training convolutional networks on synthetic data to stabilize prostate $T_2$ parameter estimation in low-signal environments. Similarly, Umberto Gava developed a model-free deep learning approach to generate brain perfusion maps from truncated CT data, bypassing the extensive temporal sampling required by traditional physics equations. Where systematic biases do emerge from fast protocols, post-processing networks can act as calibrators; Adam M. Saunders showed that a patch-based ResNet can successfully map and correct tissue-dependent biases in rapid quantitative $T_1$ sequences, aligning them with slow, gold-standard clinical measurements.