Authors: Li Feng, Hersh Chandarana
Categories: Article, Abdominal MRI, Rapid MRI, Accelerated imaging, Dynamic imaging, Respiratory motion
Source: Journal of magnetic resonance imaging : JMRI
Doi: 10.1002/jmri.29750
Authors: Li Feng, Hersh Chandarana
MRI is widely used for the diagnosis and management of various abdominal diseases involving organs such as the liver, pancreas, and kidneys. However, one major limitation of MRI is its relatively slow imaging speed compared to other modalities. In addition, respiratory motion poses a significant challenge in abdominal MRI, often requiring patients to hold their breath multiple times during an exam. This requirement can be particularly challenging for sick, elderly, and pediatric patients, who may have reduced breath-holding capacity. As a result, rapid imaging plays an important role in routine clinical abdominal MRI exams. Accelerated data acquisition not only reduces overall exam time but also shortens breath-hold durations, thereby improving patient comfort and compliance. Over the past decade, significant advancements in rapid MRI have led to the development of various accelerated imaging techniques for routine clinical use. These methods improve abdominal MRI by enhancing imaging speed, motion compensation, and overall image quality. Integrating these techniques into clinical practice also enables new applications that were previously challenging. This paper provides a concise yet comprehensive overview of rapid imaging techniques applicable to abdominal MRI and discusses their advantages, limitations and potential clinical applications. By the end of this review, readers are expected to learn the latest advances in accelerated abdominal MRI and explore new frontiers in this evolving field.
Magnetic resonance imaging (MRI) is an important clinical tool that offers superior soft-tissue contrast and diverse contrast mechanisms without radiation exposure for diagnosing and managing various abdominal diseases. Compared to ultrasound and computed tomography (CT), MRI provides greater specificity in characterizing focal abdominal lesions, such as hepatocellular carcinoma (HCC),^1^ and enables both morphological and functional assessments across a broad spectrum of conditions, from oncologic to non-oncologic and from acute to chronic diseases.^1,2^ These advantages have made MRI an invaluable imaging modality for abdominal applications in clinical practice.
Despite these advantages, the clinical adoption of abdominal MRI still falls behind ultrasound and CT, which remain the first-line choices for diagnosing many abdominal diseases in the clinical setting.^3^ This is attributed to several reasons, including the limited accessibility, high cost, and slow imaging speed of MRI, as well as its complex imaging workflow. Among these, slow imaging speed is one of the major barriers in routine MRI exams, which results in long scan times, reduced patient comfort, motion-induced artifacts, and low clinical efficiency.
In current clinical practice, a typical abdominal MRI exam can take 30 minutes or longer, requiring patients to remain still and perform multiple breath-holds to minimize respiratory motion artifacts. However, many patients, especially the sick, elderly, and pediatric patients, have limited breath-holding capacity. In these patients, images may be degraded by motion artifacts, or the spatial resolution and/or volumetric coverage may need to be compromised to allow for shorter breath-hold durations. As a result, rapid imaging techniques play a crucial role for improving scan efficiency, patient experience, and overall image quality. By shortening acquisition times and improving image quality, these methods have the potential to broaden the clinical adoption of abdominal MRI and increase its diagnostic utility.
This review paper provides a thorough overview of rapid imaging techniques suitable for abdominal MRI applications, with a primary focus on methods that enable fast data acquisitions through k-space undersampling and simultaneous multi-slice imaging. Following the introduction, the second section outlines the imaging sequences and sampling trajectories currently used in routine clinical abdominal MRI and summarizes their challenges and limitations. The third section discusses state-of-the-art rapid imaging techniques for both static and dynamic abdominal MRI. The fourth section explores the clinical applications and impact of accelerated MRI in abdominal imaging, while the last section summarizes the review and discusses potential future research directions. By the end of this review, readers – including MRI physicists and clinicians – are expected to have a good understanding of the fundamental principles, advantages, challenges, limitations, and clinical applications of different accelerated imaging techniques for abdominal MRI.
In current clinical practice, an abdominal MRI exam typically combines 2D and 3D acquisitions with different types of image contrast. Typical sequences for abdominal MRI (a) fat-suppressed T1-weighted imaging using a 3D gradient-echo (GRE) sequence performed before and following the injection of an exogenous contrast agent; (b) T2-weighted imaging using a 2D single-shot fast spin-echo (SS-FSE) sequence and/or a 2D multi-shot fast spin-echo (MS-FSE) sequence for lesion detection and evaluation; (c) in-phase and out-of-phase imaging using a 2D or 3D multi-echo GRE sequence for assessing fat content; (d) diffusion-weighted imaging (DWI) using a 2D echo-planar imaging (EPI) sequence for evaluating tissue diffusivity and lesion characterization. For specific applications, such as MR cholangiopancreatography (MRCP), imaging is performed using a 3D FSE sequence to achieve strong T2-weighted contrast for visualizing the bile duct and pancreatic duct. This is typically performed with respiratory triggering to enable free-breathing data acquisition with nearly isotropic resolution.^4^
In abdominal MRI, gadolinium-based contrast agents (GBCAs) are widely used to improve lesion detection and characterization. This is achieved through contrast-enhanced multi-phase imaging, also known as dynamic contrast-enhanced MRI (DCE-MRI), to acquire multiple fat-suppressed 3D T1-weighted images at different time points before and after the intravenous administration of a contrast agent to capture arterial, venous, and delayed phases. Beyond traditional qualitative assessment, quantitative MRI, such as proton density fat fraction (PDFF) quantification, R2* relaxation mapping, and magnetic resonance elastography (MRE), have been increasingly applied in abdominal MRI with well-documented accuracy.^5–7^
Standard Cartesian sampling, where k-space data are acquired line by line, as shown in Figure 1a, remains the predominant imaging trajectory in abdominal MRI due to its robustness against system imperfections such as off-resonance effects and gradient delays. It enables relatively simple image reconstruction and provides great compatibility with different types of imaging sequences. EPI can be seen as a specialized form of Cartesian sampling that allows the acquisition of multiple k-space lines after each radiofrequency (RF) excitation. This greatly improves imaging efficiency compared to standard Cartesian sampling, making EPI particularly suited for applications such as DWI, where time-consuming diffusion encoding is required before data acquisition. However, EPI is susceptible to artifacts caused by off-resonance, gradient delays, and other factors, requiring additional corrections during image reconstruction.
Despite its widespread use and inherent robustness, Cartesian sampling has limitations that can affect its performance in abdominal MRI. It is highly sensitive to respiratory motion, which poses significant challenges during abdominal scans. Failed breath-holds can result in ghosting artifacts and blurring to degrade the overall image quality. Moreover, the acceleration potential of Cartesian sampling is restricted to the phase-encoding dimension for 2D imaging and to the phase-encoding and slice-encoding dimensions for 3D imaging. This limits the overall acceleration capability of Cartesian sampling.
Non-Cartesian sampling, where k-space data are acquired on a non-Cartesian grid, has gained considerable interest for abdominal MRI over the past decade. Radial and PROPELLER sampling, as shown in Figure 1b, are two non-Cartesian trajectories that have been increasingly adopted over the past decade due to their increased robustness to motion. In abdominal MRI, radial sampling is typically employed for 3D T1-weighted imaging using a stack-of-stars sampling scheme that combines radial sampling for in-plane imaging and Cartesian sampling for through-plane imaging, as shown in Figure 1b.^8,9^ This approach offers several advantages over full 3D radial sampling, including effective fat suppression and flexible slice coverage or resolution. PROPELLER sampling, also known under different names such as BLADE (Siemens Healthineers) or MultiVane (Philips Healthcare), can be used for 2D T2-weighted fast spin-echo imaging.^10^ In this method, each echo train acquires a blade of k-space data rotated by a preset angle. Since each blade covers the center of k-space, PROPELLER imaging offers increased robustness to motion and is particularly advantageous for free-breathing T2-weighted imaging in abdominal applications.^11^ Figure 2 presents representative examples from the study by Block et al^9^ to show that Cartesian sampling is highly sensitive to motion artifacts, leading to compromised image quality in patients unable to hold their breath. In contrast, radial sampling offers improved motion robustness, allowing patients to breathe freely during MRI scans.
Despite its advantages, non-Cartesian sampling comes with its own set of challenges. Reconstruction of non-Cartesian images requires a complex process called gridding, which involves interpolating k-space data from the non-Cartesian grid to a Cartesian grid before applying a fast Fourier transform (FFT) to generate an image. In addition, non-Cartesian sampling is more sensitive to off-resonance effects and gradient delay errors, often necessitating additional corrections during image reconstruction to ensure high-quality images. Therefore, the choice between Cartesian and non-Cartesian sampling for abdominal MRI needs to be guided by the specific requirements of an application and patient conditions. Table 1 provides a brief summary of the advantages and limitations of Cartesian and non-Cartesian sampling for application in abdominal MRI.
Current abdominal MRI protocols are inherently complex and time-consuming due to the need to acquire multiple contrast-weighted images across many breath-holds. This poses significant challenges for patients, technologists, and radiologists. In current clinical practice, each breath-hold scan typically lasts about 10 to 15 seconds, and if a patient is unable to maintain a breath-hold, the acquisition may need to be repeated until an adequate image is obtained. This issue is particularly critical when a contrast agent needs to be administered, as repeated scans may not be feasible. Therefore, new MRI techniques that can improve data acquisition efficiency, reduce breath-hold durations, or eliminate the need for breath-holds are highly desired. These advancements would streamline the imaging process, improve patient experience, and minimize motion artifacts while maintaining diagnostic image quality.
In current abdominal MRI protocols, acceleration imaging is typically achieved through in-plane and/or through-plane k-space undersampling or, in the case of 2D imaging, by simultaneously exciting multiple image slices. A variety of advanced acceleration imaging techniques are now available on clinical MRI scanners to shorten scan time in abdominal MRI, including partial Fourier imaging, view sharing imaging, parallel imaging and its variants, simultaneous multi-slice (SMS) imaging, compressed sensing, and deep learning. These methods can also be combined to leverage their strengths and mitigate their individual limitations. This section provides an overview of these techniques and discusses their advantages and challenges for clinical implementation. Table 2 provides a brief summary of the advantages and limitations of these techniques, including their acceleration capability and computational cost.
Partial Fourier imaging is one of the earliest fast imaging methods introduced in the 1980s and 1990s,^12^ and it remains widely used in clinical MRI due to its simplicity and effectiveness in reducing scan times. This technique leverages the conjugate symmetry of k-space, a property where the Fourier transform of a real-valued signal (e.g., an MRI image) exhibits complex conjugate symmetry. Partial Fourier imaging can be applied along the phase-encoding dimension in 2D imaging or along both the phase-encoding and slice-encoding dimensions in 3D imaging to enable image reconstruction from approximately half of the acquired data. However, conjugate symmetry only applies to purely real images. When the underlying image contains phase variations from factors such as motion, flow, or field inhomogeneities, the symmetry property is not valid. To address this, partial Fourier images need to be reconstructed using phase-constrained algorithms that estimate and correct for phase variations.^13^ Alternatively, a simple zero-filling reconstruction can be performed, where the missing k-space data are filled with zeros to accelerate reconstruction at the cost of reduced image quality.
Partial Fourier imaging is compatible with all Cartesian MRI sequences and contrast types, and it is particularly useful in 2D SS-FSE imaging and DWI, where it shortens the echo train length (ETL) and readout duration to minimize T2/T2* blurring and geometric distortion.^14,15^ It is also used to reduce breath-hold durations for imaging move organs. However, the undersampling capability of partial Fourier imaging is relatively modest, typically providing less than a twofold reduction in scan time. Despite this limitation, it remains a valuable tool in abdominal MRI and is usually combined with other rapid imaging techniques to achieve better performance.
View sharing is a simple yet effective method commonly used in dynamic MRI, especially in DCE-MRI or DCE-MR angiograph (MRA).^16–18^ The principle behind view sharing imaging is that dynamic changes, such as contrast enhancement following the injection of an exogenous contrast agent, are mostly captured in the low-frequency components of k-space, while high-frequency components contribute to finer image details. Therefore, by sampling the central k-space region more frequently than the periphery, high-frequency data can simply be shared across adjacent frames to increase imaging speed and temporal footprints. The main advantages of view-sharing techniques include fast reconstruction speed and easy clinical implementation. However, data sharing across multiple temporal frames may lead to temporal blurring, since shared high-frequency data may not accurately capture dynamic changes along time. To address this issue, optimized k-space sharing strategies can be implemented, where data sharing is confined to a short temporal window.^16,17^
Several clinically available view-sharing techniques are used across different MRI vendors, including the TWIST (Time-resolved angiography With Stochastic Trajectories) method from Siemens Healthineers,^16^ the TRICKS (Time-Resolved Imaging of Contrast KineticS) method from GE Healthcare,^17^ and the TRAK (Time-Resolved Angiography using Keyhole) method from Philips Healthcare.^18^ These methods are mainly applied to organs with minimal motion, such as the brain, breast, or extremity, since motion or inconsistent breath-holding can cause ghosting artifacts. DISCO (DIfferential Subsampling with Cartesian Ordering) is an improved view-sharing technique^19^ that uses pseudo-random variable-density k-space sampling to disperse motion-induced artifacts. The performance of DISCO has been demonstrated in DCE-MRI of the liver and is routinely available on clinical MRI scanners (GE Healthcare).
Non-Cartesian trajectories, such as radial sampling, have also been integrated with view-sharing approaches to improve motion robustness. A notable example is the KWIC (K-space Weighted Image Contrast) technique,^20^ which implements radial sampling with adaptive data-sharing weights across different k-space regions to enable broader sharing in peripheral regions compared to the center. This method has been successfully employed in DCE-MRI of the liver and other abdominal organs to provide improved motion robustness.^21^ However, while view-sharing techniques remain a simple and effective tool for clinical use, more advanced MRI reconstruction methods, as will be discussed in the following subsections, now offer improved motion robustness and better temporal fidelity, making them a more preferred alternative in clinical practice.
Parallel imaging is one of the most widely used acceleration techniques in clinical MRI^22–26^ and serves as a cornerstone for reducing scan time in abdominal exams. The idea of parallel imaging lies in the use of multicoil arrays with distinct coil sensitivities to emulate certain gradient-encoded measurements. This allows for the omission of certain k-space measurements to accelerate data acquisition, and images can be reconstructed by filling in the missing information using algorithms that incorporate coil sensitivity profiles.
Parallel imaging techniques fall into two broad k-space-based methods and image-space-based methods.^25^ K-space-based methods first estimate the unmeasured k-space data using a linear combination of coil sensitivities, followed by a standard FFT to reconstruct the image. GRAPPA (GeneRalized Autocalibrating Partial Parallel Acquisition),^24^ one of the most widely used k-space-based methods, employs an autocalibration process, where a small portion of fully sampled k-space data, known as autocalibration signals (ACS), is used to calculate coefficients for the linear combination of coil sensitivities. The coefficients are then applied to interpolate the missing k-space data. In contrast, image-space-based methods, such as SENSE (SENSitivity Encoding),^23^ first convert the undersampled k-space data into images with overlapping signal from different spatial locations known as aliasing. The reconstruction process then resolves the aliasing artifacts using coil sensitivity profiles.
A well-known limitation of parallel imaging is its impact on signal-to-noise ratio (SNR), particularly at higher acceleration rates.^23,25^ SNR reduction primarily arises from two k-space undersampling and coil geometry that is referred to as the geometry factor or g-factor.^23^ The g-factor amplifies noise when the sensitivity profiles of different coil elements are similar, which reduces the capability of parallel imaging in removing aliasing artifacts. This results in a trade-off between achievable acceleration rate and SNR, as higher acceleration rates generally lead to stronger noise amplification and, consequently, lower image quality. In clinical practice, standard parallel imaging is typically limited to an acceleration rate of 2 to 3 to maintain a good balance between scan time and image quality.
CAIPIRINHA (Controlled Aliasing in Parallel Imaging Results in Higher Acceleration) is an advanced parallel imaging technique to minimize noise amplification and residual aliasing artifacts for 3D imaging. Unlike conventional parallel imaging that collects undersamples k-space in a regular pattern, CAIPIRINHA strategically shifts undersampling patterns across the phase-encoding and slice-encoding directions.^27^ This controlled aliasing helps distribute undersampling artifacts more broadly across the entire image, thus enabling more effective utilization of coil sensitivities to resolve aliased artifacts with reduced noise amplification.
While parallel imaging is predominantly implemented with Cartesian sampling in clinical practice, it can also be applied to non-Cartesian trajectories, such as radial, spiral, and PROPELLER imaging.^28^ For example, SENSE reconstruction can be adapted to non-Cartesian sampling using iterative algorithms,^29^ and GRAPPA can be applied in PROPELLER imaging to accelerate the acquisition of each k-space blade for reducing both scan time and the echo train length.^30^ In particular, PROPELLER imaging has been demonstrated effective in T2-weighted imaging for patients who cannot hold their breath.^31^
A particularly promising non-Cartesian parallel imaging method for dynamic abdominal MRI is through-time non-Cartesian GRAPPA,^32^ which has been demonstrated for liver imaging.^33^ This technique extends standard GRAPPA reconstruction to radial or spiral sampling by calibrating GRAPPA weights – needed for estimating missing k-space points – using temporal correlations. Through-time GRAPPA first acquires a fully sampled image multiple times, treating these images as the ACS data. The GRAPPA weights are then calibrated for a small k-space segment across these fully sampled images and subsequently applied to undersampled k-space data to reconstruct the missing information. Compared to standard GRAPPA, through-time non-Cartesian GRAPPA enables higher acceleration rates and improved reconstruction performance by exploiting temporal redundancy to improve GRAPPA weight calibration and leveraging the variable-density undersampling property of non-Cartesian sampling schemes.^32^ However, the main limitation of this technique is the restriction to dynamic imaging applications. In addition, it requires fully sampled images for calibrating GRAPPA weights, which increases scan time before acquiring undersampled data. Recently, self-calibrated through-time GRAPPA has been proposed, which removes the need for separate fully sampled calibration data.^34^ However, its clinical application has so far been limited to cardiac MRI, and further research is needed to adapt it for abdominal imaging.
Overall, parallel imaging techniques are predominantly implemented with Cartesian sampling on most MRI scanners. Despite its moderate acceleration capabilities, parallel imaging remains widely used in abdominal MRI, providing a fast and robust solution for accelerating data acquisition. Further details on the basics of parallel imaging can be found in these review papers.^25,26,28^
Conventional multi-slice 2D MRI acquisition collects data sequentially from each slice. Simultaneous multi-slice (SMS) imaging improves this by exciting multiple slices simultaneously for data acquisition to reduce acquisition time.^35–37^ The performance of SMS imaging has been demonstrated in various clinical applications, including abdominal MRI.^38–40^ An SMS sequence typically uses a specialized RF pulse to simultaneously excite multiple slices at different spatial locations, with each slice uniquely modulated by a specific phase or frequency offset to distinguish it from the others. However, because these slices are acquired simultaneously, their signal overlaps, resulting in aliasing along the slice dimension. SMS image reconstruction aims to separate these superimposed slices using advanced algorithms that leverage varying coil sensitivities across the slice dimension to reconstruct each slice from the combined acquisition.
SMS imaging can be treated as an extension of parallel imaging along the slice direction. As in conventional parallel imaging, its performance depends on the geometry of the multicoil array along the slice dimension. When coil sensitivities are similar between excited slices, such as when slices are close together, reconstruction performance can be compromised. To address this, the concept of CAIPIRINHA technique (see parallel imaging) has been adapted for SMS imaging to improve slice separation by strategically shifting aliasing patterns associated with each slice, thus mitigating noise amplification and improving reconstruction quality at high acceleration rates.^36^ One advantage of SMS imaging over traditional parallel imaging is that it avoids the typical SNR penalty associated with k-space undersampling.
Compressed sensing is one of the most influential rapid MRI methods that has demonstrated a broad impact across a wide variety of clinical applications.^41–43^ Compressed sensing MRI aims to reconstruct images from undersampled data by exploiting image sparsity. Here, image sparsity refers to the concept that an image can be represented with only a few dominant coefficients, allowing it to be reconstructed from fewer data samples. An intuitive example of sparse images is MR angiography, where only the vessels contain useful information, while the background mainly consists of noise. Since many images are not inherently sparse, a sparsifying transform, such as the wavelet transform or finite differences, can be applied to represent an image in a domain where it exhibits sparsity, as illustrated in Figure 3a. In addition, compressed sensing MRI requires an incoherent undersampling scheme that preserves image content. This ensures that undersampling-induced aliasing artifacts resemble noise in the image and are distinct from the underlying image information, as shown in Figure 3b. A widely used undersampling approach in compressed sensing MRI is variable-density random undersampling on a Cartesian grid, which strategically samples more data near the center of k-space, where the main energy of the image is concentrated, and sparsely acquiring fewer measurements in the outer regions of k-space.^41^ In addition, non-Cartesian sampling trajectories, such as radial or spiral, are also well-suited for compressed sensing MRI, as they allow undersampling across multiple spatial dimensions to enable higher incoherence and improved reconstruction quality.^44,45^ Image reconstruction in compressed sensing MRI is performed using an iterative algorithm that alternates between enforcing consistency of the reconstructed image with the measured k-space data and enforcing image sparsity until a convergence criterion is met.
Compressed sensing can accelerate both static and dynamic MRI acquisitions. In static imaging, 3D imaging typically offers greater acceleration potential than 2D imaging due to its inherently higher sparsity and the ability to undersample along both the phase-encoding and slice-encoding dimensions. Dynamic images typically exhibit extensive temporal correlations, which result in greater image sparsity than static images. This allows for higher acceleration rates by exploiting temporal redundancy across the dynamic dimension^46^. Temporal image sparsity can be exploited in different ways, such as by applying an explicit temporal sparsifying transform like temporal finite differences to minimize total variations (TV),^45,47^, or by enforcing a low-rank constraint along the temporal dimension.^48^ In addition, accelerated dynamic MRI allows for different undersampling patterns across dynamic frames. This increases incoherence along the temporal dimension and distributes undersampling-induced aliasing artifacts across the spatial-time (k-t) dimensions to enable higher acceleration rates.
With the widespread availability of multicoil arrays in modern MRI scanners, compressed sensing is usually combined with parallel imaging to further enhance reconstruction performance beyond what each technique can achieve alone.^49–51^ These two fast imaging techniques provide complementary information, as the sparsity constraint in compressed sensing effectively suppresses noise amplification in parallel imaging, while multicoil arrays help reduce incoherent artifacts in compressed sensing to enable higher acceleration. In addition, the performance of dynamic compressed sensing MRI is expected to improve further with non-Cartesian sampling. This is because non-Cartesian trajectories, such as radial and spiral sampling, inherently provide greater incoherence for implementation of compressed sensing reconstruction with improved performance.
Despite these advancements, compressed sensing faces significant challenges in routine clinical implementation, such as slow reconstruction speeds, the need to optimize regularization parameters to balance data fidelity with image sparsity, and the risk of artifacts like image blurring due to over-regularization. Nonetheless, compressed sensing has played a significant role in shaping the field of rapid MRI over the past decades and has proven to be a valuable technique in abdominal MRI. More details about the basics of compressed sensing MRI and its clinical applications can be found in these review papers.^43,52,53^
Deep learning-based MRI reconstruction is one of the latest trends in rapid MRI. Since the initial demonstration of using deep learning for reconstructing undersampled MR images around 2016–2017,^54–56^ this field has seen explosive growth, with various techniques developed and applied to a wide range of clinical applications from standard 2D imaging to high-dimensional and dynamic imaging.^57,58^ Major MRI vendors are now actively integrating deep learning reconstruction into their MRI systems to enable accelerated imaging across different clinical applications, including abdominal MRI.
Deep learning-based reconstruction typically utilizes a database of images from different subjects to train a neural network. The training can be conducted in a supervised manner when (sufficient) fully sampled reference images are available, or via self-supervised methods when such images are unavailable. During training, the network learns to predict artifact-free images from undersampled images with aliasing artifacts by minimizing the difference between predicted images and reference images in supervised settings, or by ensuring consistency with undersampled k-space data in self-supervised settings.^59^ Once trained, the network is expected to generalize to new, unseen datasets in a process known as inference that can be completed within seconds. Deep learning-based image reconstruction can be applied to both Cartesian and non-Cartesian imaging, with higher reconstruction performance often achievable in non-Cartesian sampling due to its greater incoherence property.
Deep learning provides unique advantages in MRI reconstruction compared to other methods like compressed sensing and parallel imaging. One major benefit is real-time or near-real-time image reconstruction once the network is trained. This dramatically reduces reconstruction latency compared to traditional iterative reconstruction methods such as compressed sensing. Furthermore, as a data-driven technique, deep learning enables high acceleration rates even in standard 2D imaging. Its flexibility enables robust implementation across various sampling schemes, sequences, and clinical applications, making it a versatile tool for accelerated MRI. In abdominal MRI, the performance of deep learning-based rapid imaging has been demonstrated in various sequences, including 3D GRE, SS-FSE, 2D FSE imaging, and DWI.^60–63^
Deep learning can also be applied to dynamic MRI, where higher acceleration rates can be achieved by leveraging temporal correlations, like in compressed sensing MRI. In addition, deep learning can be integrated with explicit properties of dynamic images commonly used in iterative reconstruction, such as the low-rank constraint, to further improve reconstruction performance.^64,65^ Like its use in static image reconstruction, deep learning-based dynamic image reconstruction can also be trained using either supervised or self-supervised approaches. However, self-supervised training is particularly advantageous in dynamic imaging, as acquiring fully sampled dynamic datasets is often challenging. Beyond standard dynamic imaging, deep learning also holds promise for accelerating quantitative MRI to facilitate its broader clinical adoption.
The main challenge of deep learning-based MRI reconstruction is the need for substantial training datasets. For supervised training, (sufficient) fully sampled reference images without acceleration are required, which can be feasible for 2D static imaging but is challenging for 3D imaging, dynamic imaging, or quantitative imaging due to larger data volumes and prolonged scan times. To address this, self-supervised training that does not rely on reference images has become increasingly important.^59,66^ In addition, managing discrepancies between training and testing datasets, such as variations in image contrast, SNR, and voxel size, has become a critical aspect of deep learning-based image reconstruction to ensure generalizability and robust performance across different imaging conditions.^67^ Techniques involving domain adaptation and transfer learning are being actively explored to improve the robustness of trained network models. These methods improve the adaptability of deep learning-based reconstruction methods to variation in SNR, acquisition parameters, and imaging protocols across different datasets and clinical settings.^68,69^ Another promising solution is zero-shot training, which requires no training datasets and allows network training directly on individual images to be reconstructed.^70^ While this approach avoids dataset discrepancies, it necessitates scan-specific training for each dataset and results in lengthy reconstruction times, which may limit its clinical translation.
In summary, deep learning-based image reconstruction has become one of the most effective rapid MRI techniques in modern MRI scanners, addressing many limitations presented in previous acceleration methods. Unlike conventional techniques, which often involve trade-offs between acceleration rates, image quality, and noise amplification, deep learning enables higher acceleration rates while preserving spatial resolution, image quality, and SNR. It also provides flexibility across different sampling strategies, making it compatible with both Cartesian and non-Cartesian imaging. More details about the basics of deep learning-based MRI reconstruction and its clinical applications can be found in these review papers.^71,72^
The integration of various accelerated imaging techniques into abdominal MRI has significantly improved scan efficiency, motion robustness, and image quality, making MRI more clinically available for a broader range of patients. In addition, faster abdominal MRI acquisitions also unlock opportunities to explore new applications that were previously limited by long scan times or motion sensitivity. This section presents an overview of the applications of accelerated imaging techniques in abdominal MRI. The first subsection discusses how different acceleration methods can be applied and optimized for specific abdominal MRI sequences, and the second subsection explores emerging clinical applications in abdominal MRI that can benefit from the advancements in rapid imaging techniques.
The choice of accelerated imaging techniques in abdominal MRI depends on the underlying imaging sequences and clinical requirements. Different MRI sequences benefit from specific acceleration strategies tailored to their unique needs. This section discusses how rapid imaging methods are selected and applied to different abdominal MRI sequences, including T1-weighted imaging, in-phase/out-of-phase imaging, T2-weighted and MRCP imaging, DWI, and dynamic MRI, along with their advantages and challenges.
T1-weighted imaging in abdominal MRI is typically performed using a fat-suppressed 3D GRE sequence, while in-phase/out-of-phase imaging is acquired with a multi-echo GRE sequence (either as 2D or 3D acquisition). Both sequences are conventionally acquired under breath-hold conditions. As a result, shortening scan time is essential for improving patient comfort and compliance. Partial Fourier imaging and parallel imaging are both widely employed in these two sequences to enable faster acquisition speed in abdominal MRI while maintaining diagnostic quality. In particular, partial Fourier imaging is often applied along both the phase-encoding and slice-encoding dimensions, which enables the acquisition of a complete 3D image within a single breath-hold.^73^ More recently, advanced rapid imaging techniques, particularly deep learning-based reconstruction, have demonstrated great potential in further reducing breath-hold durations without compromising image quality.^63^ This makes T1-weighted and in-phase/out-of-phase imaging more accessible for patients with limited breath-holding capacity.
Standard T2-weighted imaging in abdominal MRI is performed using a 2D FSE sequence, which can be implemented as either SS-FSE imaging or MS-FSE imaging, each with respective advantages and limitations. SS-FSE acquires all k-space data in a single echo train, which significantly reduces scan times and motion artifacts, making it particularly beneficial for patients with limited breath-holding capacity. However, SS-FSE suffers from lower spatial resolution due to longer ETL and increased T2 blurring due to T2 decay across the echo train. This limits the primary use of SS-FSE to initial anatomical assessments in clinical abdominal MRI, especially for patients unable to hold their breath. In contrast, MS-FSE acquires k-space data over multiple RF excitations, which results in longer scan time but enables higher spatial resolution and improves SNR with reduced T2 blurring. Different rapid MRI techniques, including partial Fourier imaging, parallel imaging, and SMS imaging, can be used to accelerate T2-weighted abdominal MRI. Partial Fourier imaging and parallel imaging, in particular, play a crucial role in shortening the ETL in SS-FSE to help mitigate T2 blurring. In addition, deep learning has also been increasingly used to accelerate T2-weighted imaging in recent years, which enables superior performance compared to more traditional acceleration methods. It is particularly useful in SS-FSE imaging to shorten the ETL, thereby reducing T2 blurring and specific absorption rate (SAR).^74^ Figure 4 shows a comparison of standard SS-FSE and deep learning-enabled SS-FSE protocols for bowel imaging.^75^ Deep learning allows for reduced ETL and shorter total acquisition time. The shorter ETL helps minimize T2 blurring in SS-FSE imaging.
In addition to standard T2-weighted imaging, MRCP imaging is commonly used in abdominal MRI to evaluate the biliary and pancreatic ducts. MRCP is typically performed using a 3D FSE sequence to achieve strong T2-weighted contrast with nearly isotropic spatial resolution, which results in prolonged scan times. To accommodate this, MRCP is often performed as a free-breathing sequence with respiratory triggering to minimize motion artifacts. Rapid MRI techniques that have been applied to accelerate MRCP imaging include partial Fourier imaging, parallel imaging, compressed sensing, and deep learning. In particular, both compressed sensing and deep learning have demonstrated the feasibility of highly-accelerated MRCP within a single breath-hold, achieving an acceleration rate of up to 17x.^76–78^ This significantly reduces the scan time and can be particularly useful in patients who can comply with breath-hold instructions. Figure 5 shows a comparison of standard and deep learning-accelerated free-breathing MRCP using respiratory triggering from the study by Bane et al, where deep learning enables faster acquisition and improves visual SNR to improve image quality.
DWI is widely used in abdominal MRI for lesion detection and characterization but is inherently limited by long acquisition times and low SNR. To compensate for the low SNR, clinical DWI is commonly performed with multiple averages, which improves image quality at the expense of increased scan time. The most frequently used sequence for abdominal DWI is 2D single-shot EPI (SS-EPI), which enables efficient diffusion image acquisition. However, SS-EPI is highly susceptible to geometric distortion and signal loss due to off-resonance, particularly at high field strengths (e.g., 3T) and in areas with significant susceptibility variations, such as near air-tissue interfaces.
Several acceleration imaging techniques are routinely used in clinical abdominal DWI, including partial Fourier imaging, parallel imaging, and SMS imaging. Partial Fourier imaging and parallel imaging both help shorten the ETL in SS-EPI, thereby minimizing geometric distortion. However, the impact of these techniques on overall scan time reduction is limited, because DWI acquisition requires a time-consuming diffusion encoding process before data acquisition. SMS imaging is a more effective technique for increasing imaging speed in DWI by enabling simultaneous acquisition of multiple slices without compromising spatial coverage.^38–40^ In addition, deep learning-based methods have also shown great promise in further reducing scan times in DWI. This is achieved by increasing acceleration rates and reducing the number of signal averages needed during data acquisition, while applying deep learning-based denoising to maintain sufficient SNR.^79^ Figure 6 shows a comparison of standard DWI and SMS DWI for liver imaging from the study by Taron et al,^79^ where SMS DWI enables faster data acquisition without compromising image quality. Figure 7 shows another comparison of SMS DWI and deep learning-accelerated SMS DWI of the liver from the study, where deep learning allows for additional scan time reduction without affecting image quality.
Most abdominal MRI exams performed in clinical practice today do not involve continuous dynamic imaging. Even in DCE-MRI, different contrast phases are typically acquired separately under multiple breath-holds in clinical protocols, rather than as a continuous dynamic time series. However, DCE-MRI exhibits strong temporal correlations, which can be leveraged to accelerate data acquisition and improve both spatial and temporal resolution. Over the past decade, various studies have demonstrated the feasibility of highly-accelerated DCE-MRI in the abdomen using rapid dynamic MRI methods, including view sharing,^21^ dynamic compressed sensing^45,80^ and deep learning-based reconstruction.^58,81^ These techniques exploit the temporal redundancy of dynamic images to achieve higher acceleration rates than static imaging while preserving diagnostic quality. The ability to acquire higher temporal resolution DCE-MR images has also facilitated perfusion quantification and expanded its clinical utility for assessing tumor vascularity, liver fibrosis, and renal function.^82^ Beyond DCE-MRI, rapid dynamic MRI techniques have also shown increasing potential for emerging applications that are not yet routinely performed, such as MR relaxometry and flow imaging, as will be discussed in the next subsection.
The advances in accelerated MRI techniques have not only improved the efficiency of conventional abdominal imaging protocols but also enabled new applications that were previously limited by long acquisition times and motion sensitivity. These emerging applications hold great potential to expand the role of abdominal MRI in clinical practice, providing new diagnostic and quantitative information that could improve disease detection, characterization, and management. This section explores how these accelerated imaging methods have been applied in advanced clinical applications beyond routine clinical protocols.
High-resolution imaging (<1 mm spatial resolution) provides significant benefits in detecting small lesions, assessing fine anatomical structures, and improving diagnostic accuracy in abdominal MRI. However, achieving high spatial resolution has traditionally been challenging due to long acquisition times, reduced SNR, and increased motion sensitivity. Among various acceleration methods, deep learning-based reconstruction has emerged as a promising solution to enable high-resolution imaging without the typical trade-offs in SNR or scan time. The time savings from higher acceleration rates can be leveraged to increase spatial resolution and improving lesion delineation while maintaining clinical feasibility.^83^ In addition, deep learning can also be employed for joint image reconstruction and denoising to effectively compensate for the SNR loss associated with high spatial resolution acquisitions.
High-resolution abdominal MRI can be useful for different applications, such as (a) improving the detection and characterization of small lesions in patients at risk for HCC or metastatic disease, and (b) improving the visualization of the pancreatic duct and small cystic lesions for early diagnosis of malignancies in pancreatic imaging. Figure 8 shows a comparison of standard SS-FSE and deep learning-enabled SS-FSE for imaging the pancreas. Deep learning provides higher spatial resolution to enhance the delineation of pancreatic vessels and parenchyma, without extending scan time or affecting SNR. Figure 9 shows another example to compare standard breath-hold volumetric interpolated breath-hold examination (BH-VIBE) (a 3D GRE sequence from Siemens Healthineers) with deep learning-enabled BH-VIBE. Deep learning enables higher spatial resolution to improve the delineation of liver vessels and parenchyma without increasing scan time or compromising SNR. As accelerated imaging techniques continue to evolve, high-resolution abdominal MRI is expected to become more clinically adopted for evaluation of different disease conditions.
While rapid imaging techniques have helped reduce breath-hold durations in clinical abdominal MRI scans, some patients may not be able to hold their breath at all. In such cases, free-breathing imaging techniques, which allow patients to breathe freely during the scan, are highly desirable. However, a key limitation of free-breathing acquisitions is prolonged scan times. Therefore, rapid MRI techniques play an essential role in free-breathing imaging to reduce scan time while maintaining image quality.
One example of a rapid free-breathing imaging technique is GRASP (Golden-angle RAdial Sparse Parallel) MRI,^45^ which is clinically available for free-breathing DCE-MRI. GRASP enables continuous data acquisition while providing the flexibility to retrospectively reconstruct dynamic images at different temporal resolutions. Over the past years, several improved variants based on the GRASP framework have also been developed to further improve motion robustness and reconstruction quality with application in abdominal MRI.^80,84–86^ Figure 10 shows an example of GRASP DCE-MRI from the study of Chen,^80^ which enables continuous, free-breathing dynamic contrast-enhanced MRI with flexible temporal resolution. In addition to radial sampling, PROPELLER imaging is also routinely available for free-breathing T2-weighted imaging and can be further accelerated using different rapid imaging techniques to improve scan efficiency.^11,31^
In addition to non-Cartesian sampling schemes, free-breathing imaging can also be performed using Cartesian sampling trajectories with advanced self-navigation or motion-sorting algorithms.^87–89^ Furthermore, various motion sensors, such as the PilotTone device,^90^ ultrasound-based motion sensors,^91^ and time-of-flight cameras,^92^ are now available for clinical investigation and evaluation in abdominal MRI. The integration of deep learning-based reconstruction with free-breathing imaging, which is now an active area of research, is expected to further increase the robustness and imaging efficiency, making these techniques more practical for routine clinical use across a broader range of patients.
Quantitative MRI has long been an area of interest in the field of MRI research, and it is increasingly being adopted in routine clinical abdominal MRI. For example, PDFF quantification and R2* relaxation mapping can be performed using multi-echo GRE sequences to assess hepatic fat and iron levels in the liver, respectively.^5,7^ MRE has become a clinical standard for assessing liver fibrosis with well-documented accuracy,^6^ and it has also shown promise in differentiating between benign and malignant liver lesions based on variations in tissue stiffness.^93^ More recently, studies have shown that T1 mapping may play an important role in evaluating chronic liver diseases, including inflammation and fibrosis progression.^94–98^
Quantitative MRI techniques typically require the acquisition of multiple images with varying contrast weightings, followed by fitting these images to a quantification model. As a result, quantitative imaging can be viewed as a dynamic imaging process, where contrast variations represent different time points in the acquisition and exhibit strong temporal correlations. These temporal correlations can be leveraged using dynamic imaging techniques, such as compressed sensing and deep learning-based reconstruction, to achieve high acceleration rates. These approaches hold great promise to achieve reduced scan time, making quantitative MRI more clinically feasible or more accurate for parameter estimation. Figure 11 shows two clinical cases demonstrating the use of PDFF and R2* mapping to quantitatively assess liver fat and iron loading, where data acquisition was performed using CAIPIRINHA-accelerated breath-hold 3D multiecho GRE imaging with an acceleration rate of 3. Figure 12 shows three different cases demonstrating the use of compressed sensing-accelerated water-only T1 mapping of the liver for assessing chronic liver diseases. Water-only T1 values from different patients exhibit a good correlation with MR elastography-derived tissue stiffness, which is related to the disease stage of liver fibrosis.
In recent years, low-field MRI (≤0.55T) has gained renewed interest as a cost-effective and accessible imaging solution, particularly for use in resource-limited settings.^99,100^ Historically, low-field MRI has been constrained by lower SNR and longer scan times, which limit its clinical utility. With the recent advances in rapid MRI techniques, particularly deep learning-based reconstruction and denoising, image quality at lower field strengths have been improved significantly, and several recent studies have demonstrated the feasibility of abdominal MRI at 0.55T and its unique advantages beyond cost considerations.^101,102^ Key benefits for low-field abdominal MRI include (a) shorter T1 recovery and longer T2/T2* relaxation times, which lead to more efficient signal recovery and slower signal decay, (b) improved robustness against susceptibility artifacts in DWI, which makes it particularly advantageous in regions prone to distortion from air-tissue interfaces, and (c) lower operational cost that can potentially increase the accessibility in resource-limited regions and outpatient settings.^103^ As deep learning-based denoising and reconstruction techniques continue to develop, low-field MRI is expected to expand beyond its current applications, offering a promising cost-effective alternative for abdominal imaging, particularly in settings where high-field MRI is not readily available.^103^
This paper provides an overview of various rapid MRI techniques available for accelerated abdominal MRI in current clinical practice to help readers understand their principles, benefits, challenges, and applications. We first reviewed MRI sequences commonly used in clinical abdominal exams and discussed the need for increased imaging speed in routine abdominal MRI. Most rapid imaging techniques currently employed in clinical settings have proven effective in abdominal MRI, with significant potential to reduce scan time and improve patient comfort. In addition, we explored emerging clinical applications that could benefit from the advances in rapid MRI techniques, including high-resolution imaging, free-breathing imaging, quantitative MRI, and low-field MRI. With the increasing integration of advanced rapid imaging techniques into clinical practice, these novel applications have already shown promise in improving disease detection, characterization, and management.
There are several future directions for further improving rapid imaging in abdominal MRI. First, the development of self-supervised training techniques for deep learning-based reconstruction is expected to increase the clinical utility and applications. This is particularly important in abdominal MRI, where acquiring fully sampled reference data can be challenging due to different types of motion. Second, quantitative MRI techniques may see increasing clinical adoption, which will allow us to further evaluate the value of different imaging biomarkers to improve disease diagnosis, prognosis, and management. Third, faster imaging speeds could enable real-time 3D or 4D MRI with high spatial resolution (<1 mm) and high temporal resolution (<1 sec). This is expected to improve motion robustness and ultimately eliminate the need for explicit motion compensation by capturing respiratory motion in real time. Lastly, real-time and interactive MRI, particularly with the growing interest in low-field MRI, presents another promising research direction. This could potentially enable on-the-fly adjustment of scan parameters based on preliminary image quality assessment and also facilitate the use of MRI towards image-guided treatment.
Overall, the continued advancement of rapid MRI techniques is expected to pave the way for faster, more patient-friendly, and more informative abdominal MRI exams and different MRI scans in general, ultimately improving clinical decision-making and patient outcomes.