High-resolution dynamic imaging of the thorax and abdomen—such as cardiac or liver MRI—has historically relied on a fragile assumption: that the patient can hold their breath on command. To prevent respiratory motion from blurring fine anatomical structures, standard clinical protocols require patients to pause their breathing for 15 to 25 seconds at a time.
However, this requirement excludes the very populations who often need these diagnostics most. Patients with advanced heart failure, respiratory distress, pediatric patients, and individuals with cognitive impairments like dementia frequently cannot maintain stable breath-holds, as noted by Reuben Jyong Kiat Foo et al. in their study on scan quality limitations (Association Between Dementia and Optical Coherence Tomography Scan Quality). When a patient fails to comply with breathing instructions, the resulting data is heavily corrupted by motion artifacts, rendering scans clinically unusable.
To resolve this bottleneck, a major research direction has emerged to transition medical imaging from a paradigm of patient compliance to one of algorithmic resilience. By combining ultra-fast acquisition sequences, mathematical optimization, and real-time motion tracking, modern techniques allow patients to breathe naturally throughout the scan while still producing sharp, diagnostic-quality images.
Outrunning Motion via Ultra-Fast Acquisition
The most direct way to eliminate the need for a breath-hold is to make the image acquisition process faster than the physical motion itself. If a single frame can be captured in tens of milliseconds, the diaphragm's movement during that window becomes negligible, effectively freezing the respiratory motion.
To achieve this, researchers drastically undersample k-space—the spatial frequency domain where raw MRI data is collected. Instead of filling k-space line by line, fast sequences collect only a small fraction of the data. Martin Uecker demonstrated this in Real-time cardiovascular magnetic resonance at high temporal resolution: radial FLASH with nonlinear inverse reconstruction, utilizing highly undersampled radial trajectories combined with a nonlinear inverse reconstruction algorithm. This approach achieved imaging speeds of up to 50 frames per second, completely bypassing the need for both breath-holding and external electrocardiogram (ECG) gating.
Similarly, Samuel T Ting et al. introduced a specialized spatiotemporal sampling pattern called VISTA, paired with an optimized parallel imaging reconstruction framework (Sub-30 ms real-time, free-breathing cardiac imaging with SPIRiT). By capturing images in under 30 milliseconds, the authors successfully isolated rapid cardiac motion and respiratory cycles without introducing the "ghosting" artifacts typical of slower, conventional sequences.
[ Traditional MRI Scan ] ---> Takes ~20 seconds ---> Requires Breath-Hold to avoid blur
[ Ultra-Fast Radial ] ---> Takes ~25 ms ---> Freezes motion; no breath-hold needed
Exploiting Data Sparsity and Redundancy
When hardware constraints prevent instantaneous imaging, researchers rely on Compressed Sensing (CS) to reconstruct full-fidelity images from sparse, free-breathing data. Compressed sensing works because dynamic physiological processes are highly redundant; the difference between consecutive frames in a video sequence is mostly zero (sparse).
Algorithms can exploit this mathematical sparsity to solve the underdetermined problem of missing k-space data. For example, Yongjun Kwak et al. accelerated phase-contrast blood flow imaging by utilizing the "complex difference" between flow-encoded images as a sparsifying transform (Accelerated phase contrast imaging using compressed sensing with complex difference sparsity). Because the background anatomy remains static between the two encodings, subtracting them isolates the moving blood, creating a highly sparse image that can be perfectly reconstructed from a fraction of the usual data.
Alternatively, the "dead time" between traditional breath-holds can be turned into an asset. Seunghoon Nam et al. developed a method that collects unconstrained, free-breathing data during these rest periods (Improved accelerated breath-hold radial cine image reconstruction by acquiring additional free-breathing data between breath-holds). This extra data acts as a low-resolution reference map, guiding the compressed sensing algorithm to accurately fill in the missing k-space lines of highly accelerated scans.
Active Tracking and Self-Gating
Rather than simply outrunning or compressing motion, an alternative strategy is to actively track the respiratory cycle and adapt the scanner's behavior on the fly. This can be achieved through hardware-free tracking mechanisms embedded directly within the imaging sequence.
Real-Time Slice Following
As a patient breathes, the diaphragm pushes the heart and liver up and down along the superior-inferior axis. Kévin Moulin et al. implemented a "slice-following" sequence for cardiac diffusion imaging (In vivo free‐breathing DTI and IVIM of the whole human heart using a real‐time slice‐followed SE‐EPI navigator‐based sequence: A reproducibility study in healthy volunteers). The sequence uses a rapid navigator echo to measure liver displacement in real time. If the liver moves by 5 mm, the scanner automatically shifts the excitation slice by a corresponding amount, ensuring that the exact same cross-section of tissue is imaged throughout the respiratory cycle.
Data-Driven Self-Gating
Instead of using external respiratory belts or separate navigator echoes, self-gating extracts motion signals directly from the core imaging data. Because the center of k-space represents the total signal intensity of the imaged volume, its value fluctuates periodically as the lungs fill with air and empty. Shuo Li et al. validated this approach for 3D cardiac cine imaging, demonstrating that these internal, data-driven signals can sort free-breathing data into distinct respiratory phases with an accuracy that matches traditional hardware gating (The accuracy and effectiveness of self-gating signals in free-breathing 3D cardiac cine MRI).
Retrospective Alignment and Deep Learning
The final line of defense against respiratory motion occurs after data collection is complete. If data is acquired continuously during free breathing, retrospective computational alignment can untangle the motion post-hoc.
For quantitative tissue characterization, such as $T_2$ mapping, images must be acquired with varying contrasts to calculate precise relaxation times. However, respiratory motion shifts the anatomy between these contrast changes, corrupting the pixel-by-pixel mathematical fit. To solve this, the ARCTIC framework, evaluated by Sébastien Roujol et al., uses image-based registration to mathematically realign $T_2$-weighted images before fitting the relaxation curves, restoring high reproducibility to free-breathing cardiac maps (Motion correction for free breathing quantitative myocardial t2 mapping: impact on reproducibility and spatial variability).
[ Misaligned Free-Breathing Frames ]
│ (Varying contrasts + respiratory shifts)
▼
[ ARCTIC Image Registration ]
│ (Mathematical alignment of anatomy)
▼
[ Precise Pixel-by-Pixel Fit ] ---> Sharp, accurate quantitative map
Where classical mathematical alignment fails or proves too slow for clinical workflows, deep learning models can step in to suppress artifacts instantly. Yunan Wu et al. designed "J-net," a two-stage deep learning framework tailored for liver dynamic contrast-enhanced MRI (Image-based Motion Artifact Reduction on Liver Dynamic Contrast Enhanced MRI). The network uses a residual architecture to isolate and subtract complex motion blur from the image, followed by a perceptual loss stage to restore fine structural details in less than a second.
By shifting the burden of motion management from the patient's physical control to advanced mathematical and computational frameworks, these combined innovations make high-fidelity diagnostic imaging accessible to critically ill, pediatric, and uncooperative patient populations.
Key Foundational Work
EXAMINATION OF IMAGE DIAGNOSIS SYSTEM AT HIGH LEVEL EMERGENCY MEDICAL SERVICE (I. Emergency Medicine and Radiological Technology) (1983) — contributed 2 concept elements to this direction.