More IMPATIENT: A Gridding-Accelerated Toeplitz-based Strategy for Non-Cartesian High-Resolution 3D MRI on GPUs

A clinically viable vendor-independent and device-agnostic solution for accelerated cardiac MRI reconstruction

BACKGROUND AND OBJECTIVE

Recent research has reported methods that reconstruct cardiac MR images acquired with acceleration factors as high as 15 in Cartesian coordinates. However, the computational cost of these techniques is quite high, taking about 40 min of CPU time in a typical current machine. This delay between acquisition and final result can completely rule out the use of MRI in clinical environments in favor of other techniques, such as CT. In spite of this, reconstruction methods reported elsewhere can be parallelized to a high degree, a fact that makes them suitable for GPU-type computing devices. This paper contributes a vendor-independent, device-agnostic implementation of such a method to reconstruct 2D motion-compensated, compressed-sensing MRI sequences in clinically viable times.

METHODS

By leveraging our OpenCLIPER framework, the proposed system works in any computing device (CPU, GPU, DSP, FPGA, etc.), as long as an OpenCL implementation is available, and development is significantly simplified versus a pure OpenCL implementation. In OpenCLIPER, the problem is partitioned in independent black boxes which may be connected as needed, while device initialization and maintenance is handled automatically. Parallel implementations of both a groupwise FFD-based registration method, as well as a multicoil extension of the NESTA algorithm have been carried out as processes of OpenCLIPER. Our platform also includes significant development and debugging aids. HIP code and precompiled libraries can be integrated seamlessly as well since OpenCLIPER makes data objects shareable between OpenCL and HIP. This also opens an opportunity to include CUDA source code (via HIP) in prospective developments.

RESULTS

The proposed solution can reconstruct a whole 12-14 slice CINE volume acquired in 19-32 coils and 20 phases, with an acceleration factor of ranging 4-8, in a few seconds, with results comparable to another popular platform (BART). If motion compensation is included, reconstruction time is in the order of one minute.

CONCLUSIONS

We have obtained clinically-viable times in GPUs from different vendors, with delays in some platforms that do not have correspondence with its price in the market. We also contribute a parallel groupwise registration subsystem for motion estimation/compensation and a parallel multicoil NESTA subsystem for l1-l2-norm problem solving.

Accelerating B-spline interpolation on GPUs: Application to medical image registration

BACKGROUND AND OBJECTIVE

B-spline interpolation (BSI) is a popular technique in the context of medical imaging due to its adaptability and robustness in 3D object modeling. A field that utilizes BSI is Image Guided Surgery (IGS). IGS provides navigation using medical images, which can be segmented and reconstructed into 3D models, often through BSI. Image registration tasks also use BSI to transform medical imaging data collected before the surgery and intra-operative data collected during the surgery into a common coordinate space. However, such IGS tasks are computationally demanding, especially when applied to 3D medical images, due to the complexity and amount of data involved. Therefore, optimization of IGS algorithms is greatly desirable, for example, to perform image registration tasks intra-operatively and to enable real-time applications. A traditional CPU does not have sufficient computing power to achieve these goals and, thus, it is preferable to rely on GPUs. In this paper, we introduce a novel GPU implementation of BSI to accelerate the calculation of the deformation field in non-rigid image registration algorithms.

METHODS

Our BSI implementation on GPUs minimizes the data that needs to be moved between memory and processing cores during loading of the input grid, and leverages the large on-chip GPU register file for reuse of input values. Moreover, we re-formulate our method as trilinear interpolations to reduce computational complexity and increase accuracy. To provide pre-clinical validation of our method and demonstrate its benefits in medical applications, we integrate our improved BSI into a registration workflow for compensation of liver deformation (caused by pneumoperitoneum, i.e., inflation of the abdomen) and evaluate its performance.

RESULTS

Our approach improves the performance of BSI by an average of 6.5×  and interpolation accuracy by 2×  compared to three state-of-the-art GPU implementations. Through pre-clinical validation, we demonstrate that our optimized interpolation accelerates a non-rigid image registration algorithm, which is based on the Free Form Deformation (FFD) method, by up to 34%.

CONCLUSION

Our study shows that we can achieve significant performance and accuracy gains with our novel parallelization scheme that makes effective use of the GPU resources. We show that our method improves the performance of real medical imaging registration applications used in practice today.

Trajectory optimized NUFFT: Faster non-Cartesian MRI reconstruction through prior knowledge and parallel architectures

Purpose

The non‐uniform fast Fourier transform (NUFFT) involves interpolation of non‐uniformly sampled Fourier data onto a Cartesian grid, an interpolation that is slowed by complex, non‐local data access patterns. A faster NUFFT would increase the clinical relevance of the plethora of advanced non‐Cartesian acquisition methods.

Methods

Here we customize the NUFFT procedure for a radial trajectory and GPU architecture to eliminate the bottlenecks encountered when allowing for arbitrary trajectories and hardware. We call the result TRON, for TRajectory Optimized NUFFT. We benchmark the speed and accuracy TRON on a Shepp‐Logan phantom and on whole‐body continuous golden‐angle radial MRI.

Results

TRON was 6–30× faster than the closest competitor, depending on test data set, and was the most accurate code tested.

Conclusions

Specialization of the NUFFT algorithm for a particular trajectory yielded significant speed gains. TRON can be easily extended to other trajectories, such as spiral and PROPELLER. TRON can be downloaded at http://github.com/davidssmith/TRON.

High-resolution dynamic speech imaging with deformation estimation

Dynamic speech magnetic resonance imaging (DSMRI) is a promising technique for visualizing articulatory motion in real time. However, many existing applications of DSMRI have been limited by slow imaging speed and the lack of quantitative motion analysis. In this paper, we present a novel DS-MRI technique to simultaneously estimate dynamic image sequence of speech and the associated deformation field. Extending on our previous Partial Separability (PS) model-based methods, the proposed technique visualizes both speech motion and deformation with a spatial resolution of 2.2 × 2.2 mm(2) and a nominal frame rate of 100 fps. Also, the technique enables direct analysis of articulatory motion through the deformation fields. Effectiveness of the method is systematically examined via in vivo experiments. Utilizing the obtained high-resolution images and deformation fields, we also performed a phonetics study on Brazilian Portuguese to show the method's practical utility.

A survey of GPU-based acceleration techniques in MRI reconstructions

Image reconstruction in magnetic resonance imaging (MRI) clinical applications has become increasingly more complicated. However, diagnostic and treatment require very fast computational procedure. Modern competitive platforms of graphics processing unit (GPU) have been used to make high-performance parallel computations available, and attractive to common consumers for computing massively parallel reconstruction problems at commodity price. GPUs have also become more and more important for reconstruction computations, especially when deep learning starts to be applied into MRI reconstruction. The motivation of this survey is to review the image reconstruction schemes of GPU computing for MRI applications and provide a summary reference for researchers in MRI community.

Accelerated Computing in Magnetic Resonance Imaging: Real-Time Imaging Using Nonlinear Inverse Reconstruction

Purpose

To develop generic optimization strategies for image reconstruction using graphical processing units (GPUs) in magnetic resonance imaging (MRI) and to exemplarily report on our experience with a highly accelerated implementation of the nonlinear inversion (NLINV) algorithm for dynamic MRI with high frame rates.

Methods

The NLINV algorithm is optimized and ported to run on a multi-GPU single-node server. The algorithm is mapped to multiple GPUs by decomposing the data domain along the channel dimension. Furthermore, the algorithm is decomposed along the temporal domain by relaxing a temporal regularization constraint, allowing the algorithm to work on multiple frames in parallel. Finally, an autotuning method is presented that is capable of combining different decomposition variants to achieve optimal algorithm performance in different imaging scenarios.

Results

The algorithm is successfully ported to a multi-GPU system and allows online image reconstruction with high frame rates. Real-time reconstruction with low latency and frame rates up to 30 frames per second is demonstrated.

Conclusion

Novel parallel decomposition methods are presented which are applicable to many iterative algorithms for dynamic MRI. Using these methods to parallelize the NLINV algorithm on multiple GPUs, it is possible to achieve online image reconstruction with high frame rates.

Rapid compressed sensing reconstruction of 3D non-Cartesian MRI

PURPOSE

Conventional non-Cartesian compressed sensing requires multiple nonuniform Fourier transforms every iteration, which is computationally expensive. Accordingly, time-consuming reconstructions have slowed the adoption of undersampled 3D non-Cartesian acquisitions into clinical protocols. In this work we investigate several approaches to minimize reconstruction times without sacrificing accuracy.

METHODS

The reconstruction problem can be reformatted to exploit the Toeplitz structure of matrices that are evaluated every iteration, but it requires larger oversampling than what is strictly required by nonuniform Fourier transforms. Accordingly, we investigate relative speeds of the two approaches for various nonuniform Fourier transform kernel sizes and oversampling for both GPU and CPU implementations. Second, we introduce a method to minimize matrix sizes by estimating the image support. Finally, density compensation weights have been used as a preconditioning matrix to improve convergence, but this increases noise. We propose a more general approach to preconditioning that allows a trade-off between accuracy and convergence speed.

RESULTS

When using a GPU, the Toeplitz approach was faster for all practical parameters. Second, it was found that properly accounting for image support can prevent aliasing errors with minimal impact on reconstruction time. Third, the proposed preconditioning scheme improved convergence rates by an order of magnitude with negligible impact on noise.

CONCLUSION

With the proposed methods, 3D non-Cartesian compressed sensing with clinically relevant reconstruction times (<2 min) is feasible using practical computer resources. Magn Reson Med 79:2685-2692, 2018. © 2017 International Society for Magnetic Resonance in Medicine.

A review of GPU-based medical image reconstruction

Tomographic image reconstruction is a computationally demanding task, even more so when advanced models are used to describe a more complete and accurate picture of the image formation process. Such advanced modeling and reconstruction algorithms can lead to better images, often with less dose, but at the price of long calculation times that are hardly compatible with clinical workflows. Fortunately, reconstruction tasks can often be executed advantageously on Graphics Processing Units (GPUs), which are exploited as massively parallel computational engines. This review paper focuses on recent developments made in GPU-based medical image reconstruction, from a CT, PET, SPECT, MRI and US perspective. Strategies and approaches to get the most out of GPUs in image reconstruction are presented as well as innovative applications arising from an increased computing capacity. The future of GPU-based image reconstruction is also envisioned, based on current trends in high-performance computing.

Spatially-segmented undersampled MRI temperature reconstruction for transcranial MR-guided focused ultrasound

Background

Volumetric thermometry with fine spatiotemporal resolution is desirable to monitor MR-guided focused ultrasound (MRgFUS) procedures in the brain, but requires some form of accelerated imaging. Accelerated MR temperature imaging methods have been developed that undersample k-space and leverage signal correlations over time to suppress the resulting undersampling artifacts. However, in transcranial MRgFUS treatments, the water bath surrounding the skull creates signal variations that do not follow those correlations, leading to temperature errors in the brain due to signal aliasing.

Methods

To eliminate temperature errors due to the water bath, a spatially-segmented iterative reconstruction method was developed. The method fits a k-space hybrid signal model to reconstruct temperature changes in the brain, and a conventional MR signal model in the water bath. It was evaluated using single-channel 2DFT Cartesian, golden angle radial, and spiral data from gel phantom heating, and in vivo 8-channel 2DFT data from a FUS thalamotomy. Water bath signal intensity in phantom heating images was scaled between 0-100% to investigate its effect on temperature error. Temperature reconstructions of retrospectively undersampled data were performed using the spatially-segmented method, and compared to conventional whole-image k-space hybrid (phantom) and SENSE (in vivo) reconstructions.

Results

At 100% water bath signal intensity, 3 ×-undersampled spatially-segmented temperature reconstruction error was nearly 5-fold lower than the whole-image k-space hybrid method. Temperature root-mean square error in the hot spot was reduced on average by 27 × (2DFT), 5 × (radial), and 12 × (spiral) using the proposed method. It reduced in vivo error 2 × in the brain for all acceleration factors, and between 2 × and 3 × in the temperature hot spot for 2-4 × undersampling compared to SENSE.

Conclusions

Separate reconstruction of brain and water bath signals enables accelerated MR temperature imaging during MRgFUS procedures with low errors due to undersampling using Cartesian and non-Cartesian trajectories. The spatially-segmented method benefits from multiple coils, and reconstructs temperature with lower error compared to measurements from SENSE-reconstructed images. The acceleration can be applied to increase volumetric coverage and spatiotemporal resolution.

Compressed sensing MRI reconstruction from 3D multichannel data using GPUs

PURPOSE

To accelerate iterative reconstructions of compressed sensing (CS) MRI from 3D multichannel data using graphics processing units (GPUs).

METHODS

The sparsity of MRI signals and parallel array receivers can reduce the data acquisition requirements. However, iterative CS reconstructions from data acquired using an array system may take a significantly long time, especially for a large number of parallel channels. This paper presents an efficient method for CS-MRI reconstruction from 3D multichannel data using GPUs. In this method, CS reconstructions were simultaneously processed in a channel-by-channel fashion on the GPU, in which the computations of multiple-channel 3D-CS reconstructions are highly parallelized. The final image was then produced by a sum-of-squares method on the central processing unit. Implementation details including algorithm, data/memory management, and parallelization schemes are reported in the paper.

RESULTS

Both simulated data and in vivo MRI array data were tested. The results showed that the proposed method can significantly improve the image reconstruction efficiency, typically shortening the runtime by a factor of 30.

CONCLUSIONS

Using low-cost GPUs and an efficient algorithm allowed the 3D multislice compressive-sensing reconstruction to be performed in less than 1 s. The rapid reconstructions are expected to help bring high-dimensional, multichannel parallel CS MRI closer to clinical applications. Magn Reson Med 78:2265-2274, 2017. © 2017 International Society for Magnetic Resonance in Medicine.

High-resolution whole-brain DCE-MRI using constrained reconstruction: Prospective clinical evaluation in brain tumor patients

PURPOSE

To clinically evaluate a highly accelerated T1-weighted dynamic contrast-enhanced (DCE) MRI technique that provides high spatial resolution and whole-brain coverage via undersampling and constrained reconstruction with multiple sparsity constraints.

METHODS

Conventional (rate-2 SENSE) and experimental DCE-MRI (rate-30) scans were performed 20 minutes apart in 15 brain tumor patients. The conventional clinical DCE-MRI had voxel dimensions 0.9 × 1.3 × 7.0 mm(3), FOV 22 × 22 × 4.2 cm(3), and the experimental DCE-MRI had voxel dimensions 0.9 × 0.9 × 1.9 mm(3), and broader coverage 22 × 22 × 19 cm(3). Temporal resolution was 5 s for both protocols. Time-resolved images and blood-brain barrier permeability maps were qualitatively evaluated by two radiologists.

RESULTS

The experimental DCE-MRI scans showed no loss of qualitative information in any of the cases, while achieving substantially higher spatial resolution and whole-brain spatial coverage. Average qualitative scores (from 0 to 3) were 2.1 for the experimental scans and 1.1 for the conventional clinical scans.

CONCLUSIONS

The proposed DCE-MRI approach provides clinically superior image quality with higher spatial resolution and coverage than currently available approaches. These advantages may allow comprehensive permeability mapping in the brain, which is especially valuable in the setting of large lesions or multiple lesions spread throughout the brain.

High-frame-rate full-vocal-tract 3D dynamic speech imaging

PURPOSE

To achieve high temporal frame rate, high spatial resolution and full-vocal-tract coverage for three-dimensional dynamic speech MRI by using low-rank modeling and sparse sampling.

METHODS

Three-dimensional dynamic speech MRI is enabled by integrating a novel data acquisition strategy and an image reconstruction method with the partial separability model: (a) a self-navigated sparse sampling strategy that accelerates data acquisition by collecting high-nominal-frame-rate cone navigator sand imaging data within a single repetition time, and (b) are construction method that recovers high-quality speech dynamics from sparse (k,t)-space data by enforcing joint low-rank and spatiotemporal total variation constraints.

RESULTS

The proposed method has been evaluated through in vivo experiments. A nominal temporal frame rate of 166 frames per second (defined based on a repetition time of 5.99 ms) was achieved for an imaging volume covering the entire vocal tract with a spatial resolution of 2.2 × 2.2 × 5.0 mm³ . Practical utility of the proposed method was demonstrated via both validation experiments and a phonetics investigation.

CONCLUSION

Three-dimensional dynamic speech imaging is possible with full-vocal-tract coverage, high spatial resolution and high nominal frame rate to provide dynamic speech data useful for phonetic studies. Magn Reson Med 77:1619-1629, 2017. © 2016 International Society for Magnetic Resonance in Medicine.

3D multislab, multishot acquisition for fast, whole-brain MR elastography with high signal-to-noise efficiency

PURPOSE

To develop an acquisition scheme for generating MR elastography (MRE) displacement data with whole-brain coverage, high spatial resolution, and adequate signal-to-noise ratio (SNR) in a short scan time.

THEORY AND METHODS

A 3D multislab, multishot acquisition for whole-brain MRE with 2.0 mm isotropic spatial resolution is proposed. The multislab approach allowed for the use of short repetition time to achieve very high SNR efficiency. High SNR efficiency allowed for a reduced acquisition time of only 6 min while the minimum SNR needed for inversion was maintained.

RESULTS

The mechanical property maps estimated from whole-brain displacement data with nonlinear inversion (NLI) demonstrated excellent agreement with neuroanatomical features, including the cerebellum and brainstem. A comparison with an equivalent 2D acquisition illustrated the improvement in SNR efficiency of the 3D multislab acquisition. The flexibility afforded by the high SNR efficiency allowed for higher resolution with a 1.6 mm isotropic voxel size, which generated higher estimates of brainstem stiffness compared with the 2.0 mm isotropic acquisition.

CONCLUSION

The acquisition presented allows for the capture of whole-brain MRE displacement data in a short scan time, and may be used to generate local mechanical property estimates of neuroanatomical features throughout the brain.

Recommendations for real-time speech MRI

Real-time magnetic resonance imaging (RT-MRI) is being increasingly used for speech and vocal production research studies. Several imaging protocols have emerged based on advances in RT-MRI acquisition, reconstruction, and audio-processing methods. This review summarizes the state-of-the-art, discusses technical considerations, and provides specific guidance for new groups entering this field. We provide recommendations for performing RT-MRI of the upper airway. This is a consensus statement stemming from the ISMRM-endorsed Speech MRI summit held in Los Angeles, February 2014. A major unmet need identified at the summit was the need for consensus on protocols that can be easily adapted by researchers equipped with conventional MRI systems. To this end, we provide a discussion of tradeoffs in RT-MRI in terms of acquisition requirements, a priori assumptions, artifacts, computational load, and performance for different speech tasks. We provide four recommended protocols and identify appropriate acquisition and reconstruction tools. We list pointers to open-source software that facilitate implementation. We conclude by discussing current open challenges in the methodological aspects of RT-MRI of speech.

High-resolution dynamic speech imaging with joint low-rank and sparsity constraints

PURPOSE

To enable dynamic speech imaging with high spatiotemporal resolution and full-vocal-tract spatial coverage, leveraging recent advances in sparse sampling.

METHODS

An imaging method is developed to enable high-speed dynamic speech imaging exploiting low-rank and sparsity of the dynamic images of articulatory motion during speech. The proposed method includes: (a) a novel data acquisition strategy that collects spiral navigators with high temporal frame rate and (b) an image reconstruction method that derives temporal subspaces from navigators and reconstructs high-resolution images from sparsely sampled data with joint low-rank and sparsity constraints.

RESULTS

The proposed method has been systematically evaluated and validated through several dynamic speech experiments. A nominal imaging speed of 102 frames per second (fps) was achieved for a single-slice imaging protocol with a spatial resolution of 2.2 × 2.2 × 6.5 mm(3) . An eight-slice imaging protocol covering the entire vocal tract achieved a nominal imaging speed of 12.8 fps with the identical spatial resolution. The effectiveness of the proposed method and its practical utility was also demonstrated in a phonetic investigation.

CONCLUSION

High spatiotemporal resolution with full-vocal-tract spatial coverage can be achieved for dynamic speech imaging experiments with low-rank and sparsity constraints.

Co-registration and distortion correction of diffusion and anatomical images based on inverse contrast normalization

Diffusion MRI provides quantitative information about microstructural properties which can be useful in neuroimaging studies of the human brain. Echo planar imaging (EPI) sequences, which are frequently used for acquisition of diffusion images, are sensitive to inhomogeneities in the primary magnetic (B0) field that cause localized distortions in the reconstructed images. We describe and evaluate a new method for correction of susceptibility-induced distortion in diffusion images in the absence of an accurate B0 fieldmap. In our method, the distortion field is estimated using a constrained non-rigid registration between an undistorted T1-weighted anatomical image and one of the distorted EPI images from diffusion acquisition. Our registration framework is based on a new approach, INVERSION (Inverse contrast Normalization for VERy Simple registratION), which exploits the inverted contrast relationship between T1- and T2-weighted brain images to define a simple and robust similarity measure. We also describe how INVERSION can be used for rigid alignment of diffusion images and T1-weighted anatomical images. Our approach is evaluated with multiple in vivo datasets acquired with different acquisition parameters. Compared to other methods, INVERSION shows robust and consistent performance in rigid registration and shows improved alignment of diffusion and anatomical images relative to normalized mutual information for non-rigid distortion correction.

Accelerating patch-based directional wavelets with multicore parallel computing in compressed sensing MRI

Compressed sensing MRI (CS-MRI) is a promising technology to accelerate magnetic resonance imaging. Both improving the image quality and reducing the computation time are important for this technology. Recently, a patch-based directional wavelet (PBDW) has been applied in CS-MRI to improve edge reconstruction. However, this method is time consuming since it involves extensive computations, including geometric direction estimation and numerous iterations of wavelet transform. To accelerate computations of PBDW, we propose a general parallelization of patch-based processing by taking the advantage of multicore processors. Additionally, two pertinent optimizations, excluding smooth patches and pre-arranged insertion sort, that make use of sparsity in MR images are also proposed. Simulation results demonstrate that the acceleration factor with the parallel architecture of PBDW approaches the number of central processing unit cores, and that pertinent optimizations are also effective to make further accelerations. The proposed approaches allow compressed sensing MRI reconstruction to be accomplished within several seconds.

Correction of excitation profile in Zero Echo Time (ZTE) imaging using quadratic phase-modulated RF pulse excitation and iterative reconstruction

Zero-echo Time (ZTE) imaging is a promising technique for magnetic resonance imaging (MRI) of short-T2 tissue nuclei in tissues. A problem inherent to the method currently hindering its translation to the clinic is the presence of a spatial encoding gradient during excitation, which causes the hard pulse to become spatially selective, resulting in blurring and shadow artifacts in the image. While shortening radio-frequency (RF) pulse duration alleviates this problem the resulting elevated RF peak power and specific absorption rate (SAR) in practice impede such a solution. In this work, an approach is described to correct the artifacts by applying quadratic phase-modulated RF excitation and iteratively solving an inverse problem formulated from the signal model of ZTE imaging. A simple pulse sequence is also developed to measure the excitation profile of the RF pulse. Results from simulations, phantom and in vivo studies, demonstrate the effectiveness of the method in correcting image artifacts caused by inhomogeneous excitation. The proposed method may contribute toward establishing ZTE MRI as a routine 3D pulse sequence for imaging protons and other nuclei with quasi solid-state behavior on clinical scanners.

Improved B0 -distortion correction in diffusion MRI using interlaced q-space sampling and constrained reconstruction

PURPOSE

To enable high-quality correction of susceptibility-induced geometric distortion artifacts in diffusion magnetic resonance imaging (MRI) images without increasing scan time.

THEORY AND METHODS

A new method for distortion correction is proposed based on subsampling a generalized version of the state-of-the-art reversed-gradient distortion correction method. Rather than acquire each q-space sample multiple times with different distortions (as in the conventional reversed-gradient method), we sample each q-space point once with an interlaced sampling scheme that measures different distortions at different q-space locations. Distortion correction is achieved using a novel constrained reconstruction formulation that leverages the smoothness of diffusion data in q-space.

RESULTS

The effectiveness of the proposed method is demonstrated with simulated and in vivo diffusion MRI data. The proposed method is substantially faster than the reversed-gradient method, and can also provide smaller intensity errors in the corrected images and smaller errors in derived quantitative diffusion parameters.

CONCLUSION

The proposed method enables state-of-the-art distortion correction performance without increasing data acquisition time.