Fast MRI Techniques of the Liver and Pancreaticobiliary Tract: Overview and Application

Abstract

In liver and pancreatobiliary MRI, mitigating respiratory motion-related artifacts has always been a major challenge in image acquisition. Motion reduction by breathing control schemes or scan time acceleration by k-space undersampling are two accessible approaches in clinical imaging. Parallel imaging is an indispensable everyday technique with well-known characteristics, but with drawbacks that limit acceleration factors to ≤4. Compressed sensing exploits the data sparsity of MR images, and pseudorandomly undersamples k-space data to iteratively reconstruct images using sophisticated complex computations within highly accelerated scanning time. Albeit, this is with long reconstruction time and complexity in parameter optimization. Deep learning reconstruction uses pretrained and validated convolutional neural networks to reconstruct undersampled data, with the main tasks being image acceleration, denoising, and superresolution. While promising, deep learning reconstruction requires further testing and practical experience with model stability, generalizability, and output image fidelity.

INTRODUCTION

Liver MRI is essential for the detection, characterization, and quantification of focal and diffuse liver diseases and it offers excellent soft-tissue contrast without radiation exposure. However, liver MRI faces unique challenges, including motion artifacts from respiration and arterial pulsation, as well as the need to image a relatively large volume. Respiratory motion can cause substantial liver displacement, leading to ghosting, blurring, and degradation of the diagnostic image quality. While breathing motion control techniques such as breath-holding, respiratory triggering, and respiratory gating are commonly used, another critical strategy is accelerating image acquisition. Acceleration can be achieved through image reconstruction from undersampled data using advanced reconstruction methods. Clinical liver MRI uses parallel imaging (PI), compressed sensing (CS), and deep learning reconstruction (DLR) for this purpose. Although effective, these methods require careful parameter optimization to balance the signal-to-noise ratio (SNR), scan time, and other unique imaging characteristics. This brief review summarizes the clinically available acceleration techniques for liver MRI and highlights key considerations for their application.

APPROACHES TO MITIGATE RESPIRATORY MOTION ARTIFACTS IN LIVER MRI

Breath-holding, respiratory triggering, and respiratory gating approaches were employed to minimize the respiratory motion artifacts. Breath-holding for approximately 15–20 seconds is the simplest and most widely used method when patient cooperation is possible. However, this approach can result in severe motion artifacts in uncooperative patients or those with limited breath-holding capacity. To reduce breath-holding time, compromises in both in-plane and through-plane spatial resolutions are often made, while a prolonged echo train length can lead to image blurring (1).

In respiratory triggering, the signal acquisition is synchronized with a specific phase of the respiratory cycle. In respiratory gating, data are continuously acquired and retrospectively selected based on a predefined respiratory threshold (2,3,4). Although these methods can reduce motion artifacts and enable high spatial resolution, inconsistent breathing may prolong the scan time and introduce misregistration-related image degradation.

For patients with a markedly limited breath-holding capacity and irregular respiratory patterns, free-breathing acquisition with multiple signal averages can be used. In diffusion-weighted imaging (DWI) at high b-values, multiple averages are required to compensate for a low SNR (5). However, blurring and signal loss may persist in regions with severe motion. Motion-robust techniques, such as radial k-space sampling or view-sharing, offer improved motion tolerance but pose challenges, including complex parameter optimization, frequent streak artifacts, weak T1 signals, particularly in contrast-enhanced studies, and inherent image blurriness (1,3,4,6).

PARALLEL IMAGING (PI)

In an MRI, the scan time is proportional to the number of phase-encoding steps required to fill the k-space. PI reduces the scan time by undersampling the phase-encoding lines in an equidistant manner (Fig. 1) (7,8,9). However, undersampling introduces aliasing artifacts owing to the reduced field of view and imperfect spatial localization of the signal. Aliasing artifacts are corrected using coil sensitivity maps or auto-calibration signals from multichannel phased-array coils (9,10,11,12). Phased-array coils comprise multiple independent receiver channels or elements, each with a distinct spatial sensitivity profile, enabling the localization of MRI signals despite reduced phase-encoding steps. Image reconstruction is performed in either the image domain or the k-space.

Fig. 1. Diagram of k-space sampling pattern. Each line represents a phase encoding step, and the dotted lines in the center indicate the center of the k-space.

A. In full sampling, all phase encoding lines are sampled.

B. In parallel imaging, phase encoding lines are skipped, or undersampled, in an equidistant manner.

C. In compressed sensing, the k-space is sampled randomly; however, more is still sampled in the center.

In image-domain reconstruction, aliasing artifacts are corrected after Fourier transformation. The process begins by generating coil sensitivity maps from a low-spatial-resolution full-field-of-view acquisition for each coil receiver element. The undersampled main-pulse sequence is then acquired to produce aliased images from each receiver element. The coil sensitivity maps are used in a reconstruction matrix inversion process to “unwrap” and combine the individual aliased images, yielding a full-field-of-view image without aliasing (8,9,10). Sensitivity encoding (SENSE; Philips Healthcare), modified sensitivity encoding (mSENSE; Siemens Healthineers), and array spatial sensitivity encoding technique (ASSET; GE Healthcare) re based on an image domain reconstruction approach.

In k-space domain reconstruction, aliasing artifacts are addressed before image reconstruction. This approach predicts missing k-space data by leveraging the coil sensitivity profiles. An undersampled pulse sequence is first acquired, along with additional autocalibration signals collected near the center of the k-space. These autocalibration signals cover the same field of view as the final image but at a lower resolution. They are used to generate coil-specific weighting factors, which, in turn, estimate the missing k-space data for each coil. While acquiring more autocalibration signals can improve image quality, it also extends the scan time by several seconds, requiring careful optimization, particularly for breath-hold scans (6). After completing the k-space reconstruction, a Fourier transformation is applied to produce single-coil images, which are then combined to generate the final image (8,9,10). Commercially available k-space domain reconstruction techniques include generalized auto-calibrating partial parallel acquisition (GRAPPA; Siemens Healthineers) and auto-calibrating reconstruction for Cartesian imaging (ARC; GE Healthcare).

PI is widely used for accelerating liver MRI acquisition. It can be applied across all essential sequences and combined with fast imaging techniques such as CS. A key limitation of the PI is the reduction in the SNR as the acceleration factor increases. SNR loss is inversely proportional to the square root of the acceleration factor regardless of whether acceleration is applied in 2D or 3D imaging (Fig. 2) (13). This SNR loss results from fewer signal acquisitions, inefficiencies in the geometric factor (g-factor), and reduced coil sensitivity toward the body center. The performance of PI relies heavily on the configuration and sensitivity profile of phased-array coils, and suboptimal setups can lead to inconsistent image quality and residual aliasing artifacts. Such artifacts, including ghosting inside or outside the imaged area, become more prevalent when the acceleration factor exceeds the geometric encoding capability of the coil. These artifacts may arise from inaccurate coil sensitivity maps (in image-domain reconstruction) or erroneous coil-weighting factor calculations (in k-space-domain reconstruction). In clinical practice, the acceleration factor is typically limited to ≤4 to maintain diagnostic quality (14).

Fig. 2. Effect of the acceleration factor on coronal single-shot fast spin-echo images (echo time, 85 ms) from the same patient.

A. With an acceleration factor of 3, the edge of the liver and margins of the intrahepatic structures are slightly blurred due to the long echo train length.

B. When the acceleration factor is increased to 5, the edges are sharper, but the image appears noisier due to a decrease in the signal-to-noise ratio.

While the achievable acceleration factor theoretically scales 1:1 with the number of independent receiver elements in a multichannel phased-array coil, practical considerations, such as financial cost, device weight, and physical space, constrain the number of elements to ≤64 (10,15).

COMPRESSED SENSING (CS)

In MRI, considerable data redundancy allows for substantial k-space undersampling without compromising the image quality. CS accelerates imaging by undersampling, exploiting image sparsity, incoherence, and nonlinear iterative (16,17,18,19). An image is considered sparse when only a small subset of voxels contains the pertinent information required to reconstruct an image relative to the total number of voxels (16). In the abdomen, T2-weighted MR cholangiopancreatography (MRCP) exhibits sparsity in the imaging domain with suppressed background tissue and relevant signals concentrated in the biliary system. Multiphasic contrast-enhanced studies demonstrate temporal sparsity, in which minimal changes occur between time points (10). Sparsity transformations, such as wavelet transformations, can produce a sparse representation of an image that can be compressed with minimal information loss (17). In CS, pseudo-random k-space undersampling is performed; sampling is randomized but denser near the center, introducing incoherent noise-like artifacts (17,19). Nonlinear iterative reconstruction removes these artifacts by enforcing sparsity in the transform domain while maintaining data consistency. The regularization factor controls the balance between noise suppression and data fidelity, and the final images are generated through multiple reconstruction iterations (10,17). Commercial implementations of CS include Compressed Sensing (Siemens Healthineers), HyperSense (GE Healthcare), and compressed SENSE (Philips Healthcare).

CS can be seamlessly integrated with PI, enabling acceleration factors greater than four while preserving image quality. In particular, CS has proven to be effective in highly sparse sequences, such as T2-weighted MRCP, allowing breath-hold thin-slice 3D MRCP acquisition (Fig. 3) (14,20,21,22). CS also enhanced T1-weighted contrast-enhanced imaging and high-resolution hepatobiliary phase imaging, demonstrating better image quality, a higher SNR (Fig. 4) (23,24), improved focal lesion detection, and superior biliary visualization (25,26,27). CS performs well with motion-robust non-Cartesian sampling, making continuous free-breathing acquisition increasingly practical in clinical liver imaging (28,29,30,31).

Fig. 3. 3D MRCP accelerated with compressed sensing (A, B) and parallel imaging (C, D) from the same patient.

A-D. Representative thin-slice image (A) and maximum intensity projection image (B) of a single, 17 second breath-hold compressed sensing accelerated 3D MRCP with an acceleration factor of 25. Overall, the image provides excellent biliary visualization and good structural sharpness. However, note that noise-like artifacts remains in the background of the thin-slice image (arrows). Thin-slice image (C) and maximum intensity projection image (D) of a respiratory-triggered MRCP acquired using 3D fast spin-echo acquisition with an acceleration factor of 3. The scan time is more than 3 minutes.

MRCP = MR cholangiopancreatography

Fig. 4. Compressed sensing-accelerated breath-hold 3D T1 gradient-echo triple arterial phase images of hepatocellular carcinoma. The breath-hold time is 14 seconds, with a spatial resolution of 1 × 1 × 3 mm³. From images left to right, note the gradual enhancement of the left portal vein as the phase progresses to the late arterial phase, the optimal timing for detecting arterial phase hyperenhancement in hepatocellular carcinoma.

Conversely, CS is computationally intensive because of its reliance on complex optimization processes, particularly when applied to non-Cartesian trajectories, often resulting in prolonged reconstruction times. This extended computation poses challenges for real-time implementation of the scanner, requiring a balance between image fidelity and reconstruction efficiency. Additionally, CS depends heavily on handcrafted priors, which may not be generalizable across different scan types or patient populations, limiting its flexibility (32). For example, while increasing the regularization factor can suppress image noise, excessive regularization may produce overly smooth blurred images, potentially obscuring fine structural details (17). Optimal regularization typically requires manual tuning through trial and error. Furthermore, despite the ability of iterative reconstruction to reduce incoherent and aliasing artifacts, residual artifacts may persist, particularly at high acceleration factors (11).

DEEP LEARNING RECONSTRUCTION (DLR)

DLR algorithms for MRI have primarily been developed using supervised learning. To train convolutional neural networks (CNNs), pairs of input images with noise and high SNR ground truth images are used (33). Similar to PI and CS, DLR reconstructs images from undersampled k-space data. However, unlike traditional methods that require mathematical reconstruction for every scan, DLR replaces this process with a pre-trained, optimized model through prior training and validation sessions (34,35). Consequently, the reconstruction time is significantly reduced (35). In addition, DLR models can replace some or all of the signal-processing steps during reconstruction and can be applied during post-processing after conventional reconstruction.

DLR can be categorized into three types: k-space learning, image domain, and direct mapping (36,37). In k-space learning type, CNNs are trained to enhance k-space data and use detailed raw data from hardware, such as phase and coil sensitivity information. A CNN is embedded within the reconstruction process and operates directly on raw data (35,36). This approach may be more effective when combined with other undersampling methods, such as PI and CS, which also leverage coil sensitivity data for image reconstruction (37). However, it is generally not feasible to retrospectively reconstruct previously acquired images.

In the image domain type, also known as image-to-image DLR, CNNs are applied after Fourier transformation of the undersampled data. Most image restoration applications, including denoising, artifact reduction, and super-resolution, fall under this category (11,38). Because it uses reconstructed images as input, image-domain DLR is less affected by variations in imaging conditions or hardware and allows the retrospective processing of acquired images.

The direct mapping method reconstructs images end-to-end from the undersampled k-space to the final image. This method has the potential to mitigate errors caused by field inhomogeneity, eddy currents, phase distortions, and regridding. However, it is not yet widely available for clinical use (36,37).

Various vendor-specific or agnostic DLR solutions are available in onboard or cloud-based environments: k-space-based learning type include Sonic DL (GE healthcare) and Deep Resolve Gain/Boost/ Sharp (Siemens Healthineers), and image domain type include AIR Recon DL (GE Healthcare), SmartSpeed (Philips Healthcare), AiCE (Canon Medical Systems), DeepRecon (United Imaging Healthcare), Deep Learning Reconstruction (Fujifilm Healthcare), SwiftMR (AIRS Medical Inc), IQMR (Medic Vision-Imaging Solutions Ltd), and SubtleMR (Subtle Medical Inc) (37). These solutions can accelerate all key sequences in routine liver MRIs.

The most immediate effect of DLR-driven denoising is an improvement in SNR, which in turn enhances visualization of anatomical structures and detection of pathological conditions (Fig. 5). Alternatively, DLR can shorten the scan time while maintaining SNR. Several studies have shown superior image quality, improved focal lesion detection with DLR-accelerated single-shot fast spin-echo T2-weighted liver imaging, and reduced scan time (38,39,40,41). In liver MRI, DWI is critical for lesion detection and typically requires respiratory triggering to achieve high image quality. DLR enables substantial scan time reduction while improving the image quality and SNR, with comparable focal lesion detection rates and apparent diffusion coefficient values (Fig. 6) (42,43). In a multiphasic contrast-enhanced study, DLR was shown to enhance the SNR, improve lesion conspicuity (44,45,46), and shorten the breath-hold time during the arterial phase (Fig. 7) (45). For hepatobiliary phase imaging, DLR provided isovoxel, high-resolution imaging within a single breath-hold, resulting in improved image quality and focal lesion detection (47,48).

Fig. 5. Single-shot fast spin-echo images of hepatocellular carcinoma in the caudate lobe before (A) and after deep learning reconstruction (B) (echo time, 80 seconds; number of breath-holds, 2; acceleration factor, 3). Before deep learning reconstruction, image noise obscures the tumor margin and intrahepatic structures. After reconstruction, the tumor appears more conspicuous with cleared margin.

Fig. 6. Examples of key imaging parameters and scan times in conventional and DL-accelerated liver DWI (b = 800 s/mm²) (upper row) and matching ADC maps (bottom row) in a case of multiple hepatic metastasis from pancreatic cancer with peritoneal seeding. In free-breathing acquisition, DL reconstruction delivers an excellent signal-to-noise ratio and enhanced lesion conspicuity, while reducing scan time when compared with conventional acquisition (solid arrows). With breath-hold DL-accelerated acquisition, scan time savings are prioritized, resulting in a slightly lower focal lesion signal compared with the other methods. The ADC values of the representative lesion in segment 3 of the liver (blank arrows) are 625.2 × 10^-3 mm²/s, 530.9 × 10^-3 mm²/s, and 825.4 × 10^-3 mm²/s, for free-breathing non-DL DWI, free-breathing DL DWI, and breath-hold DL DWI, respectively.

ADC = apparent diffusion coefficient, DL = deep learning, DWI = diffusion-weighted imaging

Fig. 7. Deep learning-accelerated multiphasic 3D T1 gradient-echo dynamic enhancement images of a hemangioma in segment 7 of the liver. Images in the upper row are triple arterial phase images with a breath-hold time of 15 seconds, and images in the bottom row are portal venous phase, transitional phase (breath-hold time, 10 seconds each), and hepatobiliary phase (breath-hold time, 14 seconds). The acceleration factor is 6, with an interpolated voxel size of 0.5 × 0.5 × 3 mm³, except for the hepatobiliary phase where the voxel size is optimized to 0.7 × 0.7 × 1 mm³ for isovoxel imaging.

Although promising, DLR remains relatively new in clinical practice, and hands-on experience in interpreting and optimizing DLR images may be necessary. A commonly encountered issue is the paradoxical prominence of artifacts owing to denoising and a high SNR (36). Although DLR effectively suppresses Gaussian noise, other complex artifacts and intrinsic MR noise may persist and become more conspicuous after denoising (36). An elevated SNR can further emphasize these residual artifacts (Fig. 8). As typical DLR algorithms do not allow modification of the denoising level, training with datasets spanning multiple noise levels is essential for DLR to handle MR images with variable noise levels (36). Additionally, minor and unnoticeable image or sampling domain changes can destabilize DLR algorithms and introduce new artifacts (49).

Fig. 8. Fast spin-echo images of the liver before (A) and after (B) deep learning reconstruction. After deep learning reconstruction, the overall image noise is reduced. However, repetitive curved shape artifacts (arrows, likely respiratory motion artifacts) have become more prominent within the liver parenchyma and spleen.

Another major concern is the potential instability in representing small structures, as some algorithms may fail to reconstruct them under certain conditions (Fig. 9) (49). The omission of small lesions can result in false-negative interpretations, adversely affecting clinical decision-making and patient outcomes. Although DLR studies have primarily focused on enhancing anatomical image quality, limited attention has been paid to depicting pathological features (36). While focal lesion detection sensitivity has been reported to improve in liver MRI, the effect of DLR on detecting small lesions and maintaining lesion conspicuity requires further investigation. A side-by-side comparison with conventional reconstruction images remains necessary during interpretation to ensure that DLR does not obscure subtle pathology.

Fig. 9. Fast spin-echo images of the liver before (A) and after (B) deep learning reconstruction. After deep learning reconstruction, the overall image noise is reduced, giving the liver a smooth and artificial texture. At the same time, fine linear T2 hyperintense structures present in the native image (arrows) have become less conspicuous (dotted arrows).

The role of quantitative MRI in liver imaging is expanding, particularly for assessing hepatic fat, iron, and fibrosis (50,51,52). As quantitative imaging is expected to become more widely adopted, the application of DLR in image acquisition, generation (53), and interpretation (54) is expected to grow. Therefore, it is necessary to verify the quantitative consistency between DLR and non-DLR images, and to periodically reassess the accuracy of quantitative analyses in clinical practice. Furthermore, revalidation of the quantitative metrics is necessary when DLR algorithms are updated with expanded training datasets (36).

Additional issues include variability in image texture and noise characteristics across different DLR models and limited generalizability across vendor-specific implementations.

CONCLUSION

This review discussed respiratory control techniques and imaging acceleration methods to reduce respiratory and motion artifacts in liver and pancreatobiliary MRI. Both widely established and, more recently, clinically adopted techniques have been described, along with their key characteristics and considerations for application. Although a high SNR and minimal artifacts remain ideal for MR imaging, factors such as scan time, reconstruction time, and image reliability influenced by reconstruction stability must also be carefully considered.

Supplementary Materials

Korean translation of this article is available with the Online-only Data Supplement at https://doi.org/10.3348/jksr.2025.0004.