Adaptive Real-time Control and Online Reconstruction of Free-Breathing Abdominal MRF

Andrew Dupuis¹, Yong Chen², Madison E Kretzler², Kelvin Chow³, Xinzhou Li⁴, Mark A Griswold^1,2, and Rasim Boyacioglu²
¹Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, ²Radiology, School of Medicine, Case Western Reserve University, Cleveland, OH, United States, ³Siemens Healthcare Ltd, Calgary, AB, Canada, ⁴Siemens Medical Solution, USA, St. Louis, MO, United States

Synopsis

Motivation: Abdominal MRF can require long breath-holds. Free-breathing MRF with retrospective binning and offline reconstruction does not guarantee adequate map quality.

Goal(s): Develop a self-binning, self-terminating MRF sequence for free-breathing abdominal imaging, providing real-time adjustments to the acquisition for improved image quality.

Approach: The proposed MRF sequence monitors respiratory states in real-time, adapts or extends the acquisition based on reconstruction feedback, and performs iterative conjugate gradient reconstruction online.

Results: The free-breathing approach shows promising results mitigating motion artifacts, producing similar maps to conventional breath-hold results. Sequence termination commands are processed within 0.55 seconds, while reconstruction is completed within 22 seconds.

Impact: This work enables free-breathing abdominal MRF scans with real-time control and online reconstruction. This approach potentially allows for shorter scans with improved map quality without breath-holds. New strategies for real-time control and adaptive MRF imaging can now be investigated.

Introduction

Magnetic Resonance Fingerprinting (MRF) [1] maps multiple quantitative tissue properties simultaneously. The MRF acquisition consists of a series of RF flip angles and phases, repetition times, preparation modules, and sampling of the magnetization with fast trajectories such as spiral or radial. A typical static MRF acquisition uses optimized parameters and order of the sampling trajectories to provide consistent results. However, there can be scenarios where real-time adjustment of MRF acquisition and sampling during the scan could be beneficial, such as motion during the scan, inadequate map quality from field inhomogeneities, or inefficiency of certain preparation modules.

In this work, we introduce near real time control of the MRF framework with feedback during the actual acquisition. Specifically, we introduce a novel self-binning, self-terminating MRF sequence tailored for free-breathing abdominal imaging leveraging vendor provided respiratory waveforms. The goal is to address the challenges of motion artifacts in abdominal MRI by providing an informed online reconstruction capable of dynamically altering an acquisition in real time to ensure adequate sampling of different respiratory states, while also returning iteratively-reconstructed quantitative maps to the scanner console within seconds.

Methods

Building on prior free-breathing binning efforts [2,3], the proposed research MRF sequence employs a self-binning mechanism to track the respiratory state in real-time using respiratory waveforms (BioMatrix Sensors, Siemens Healthineers AG, Erlangen, Germany). MRF raw data and respiratory waveforms from multiple repetitions are sent to the online reconstruction cluster (Azure, Microsoft Corporation, Redmond, USA) with the Framework for Image Reconstruction Environments (FIRE) research application interface [4] in real time. Each repetition of 1728 timepoints (readouts) was acquired over 16.2 seconds. Rather than image-based tracking [5] or respiratory triggering [6], the respiratory signal data from the first “dummy” repetition is used to estimate the thresholds necessary for binning. Individual readouts from subsequent repetitions are binned into 3 respiratory phases (end-inspiration, transitional, and end-expiration), resulting in three sets of quantitative maps. After each repetition, the reconstruction can choose to either extend or terminate the acquisition based on pre-defined criteria set by the user. In this work, the target is 95% of time points from a complete repetition sampled at each respiratory phase of interest. In cases where a specific combination of time point and respiratory phase are acquired multiple times across repetitions, data was averaged over all repetitions.

Once the data threshold has been reached, the aggregated acquisition data from each bin is compressed using through-time Singular Value Decomposition (SVD) [6] based on the simulated dictionary. An iterative conjugate gradient descent NUFFT reconstruction is followed by pattern matching and a quadratic interpolation approach [8,9]. The resulting quantitative maps for each respiratory bin are then returned to the scanner console for online visualization. Figure 1 details the acquisition and reconstruction pipeline used in this study.

A comparative imaging study was performed using both our free-breathing approach and conventional breath-holds to establish a ground-truth for both the end-inspiration and end-expiration respiratory states. Four volunteers (32±14 yrs, 2 male) were scanned on a 3T system (MAGNETOM Vida, VA50A, Siemens Healthineers AG, Erlangen, Germany) after giving informed consent under our institution’s IRB. Timing information was also collected to determine feedback delay and reconstruction time.

Results

Figure 2 illustrates the sequence termination latency of the proposed pipeline. Sequence termination commands were returned and processed within 0.55±0.02 seconds, while reconstructions were completed within 22.12±3.87 seconds of the acquisition’s completion. Figure 3 demonstrates the similarity between the free-breathing binned acquisition respiratory state maps versus the conventional breath-hold equivalents for the same MRF sequence design.

Discussion

In this study we present the capability of real-time control of the MRF framework. Decisions for the running MRF sequence have been made and feedback was provided during the scan to alter the sequence. This closed-loop control makes real-time adapted, flexible and targeted scans possible, potentially shortening scans, or improving image quality. As a first example, the preliminary results of our study indicate significant promise for the self-binning self-terminating MRF sequence in mitigating motion artifacts during free-breathing abdominal imaging. The observed reduction in motion-induced image degradation from this simple binning approach suggests that further optimization of the data aggregation logic may further improve map quality.

Additionally, integration of acquisition and reconstruction to dynamically extend or terminate the acquisition yielded a wide range of scan durations necessary to accurately sample the desired respiratory states. This indicates that a one-size-fits-all acquisition that relies on retrospective reconstruction may often over- or under-acquire data for different respiratory states in ways that cannot be remedied, once a patient has left the scanner.

Acknowledgements

No acknowledgement found.

References

[1] Ma, D., Gulani, V., Seiberlich, N., Liu, K., Sunshine, J. L., Duerk, J. L., & Griswold, M. A. (2013). Magnetic resonance fingerprinting. Nature, 495(7440), 187-192.
[2] Huang, S.S., Boyacioglu, R., Bolding, R., MacAskill, C., Chen, Y. and Griswold, M.A. (2021), Free-Breathing Abdominal Magnetic Resonance Fingerprinting Using a Pilot Tone Navigator. J Magn Reson Imaging, 54: 1138-1151. https://doi.org/10.1002/jmri.27673
[3] Yong Chen, Yun Jiang, Shivani Pahwa, Dan Ma, Lan Lu, Michael D. Twieg, Katherine L. Wright, Nicole Seiberlich, Mark A. Griswold, and Vikas Gulani (2016). MR Fingerprinting for Rapid Quantitative Abdominal Imaging. Radiology 279:1, 278-286 https://doi.org/10.1148/radiol.2016152037
[4] Chow K, Kellman P, Xue H. Prototyping Image Reconstruction and Analysis with FIRE. In: proc. SCMR. Virtual Scientific Sessions; 2021. p. 838972
[5] Thesen, Stefan, et al. "Prospective acquisition correction for head motion with image‐based tracking for real‐time fMRI." Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine 44.3 (2000): 457-465.
[6] Hirokawa, Yuusuke, et al. "MRI artifact reduction and quality improvement in the upper abdomen with PROPELLER and prospective acquisition correction (PACE) technique." American Journal of Roentgenology 191.4 (2008): 1154-1158.
[7] McGivney D F, Pierre E, Ma D, et al. SVD Compression for Magnetic Resonance Fingerprinting in the Time Domain. IEEE Transactions on Medical Imaging. 2014; 33:2311-2322.
[8] McGivney D, Boyacioglu R, Jiang Y, Wang C, Ma D, Griswold MA. Towards Continuous Dictionary Resolution in MR Fingerprinting using a Quadratic Inner Product Model. ISMRM Annual Meeting, 2019.
[9] Boyacioglu, R., Wang, C., Ma, D., McGivney, D. F., Yu, X., & Griswold, M. A. (2021). 3D magnetic resonance fingerprinting with quadratic RF phase. Magnetic Resonance in Medicine, 85(4), 2084-2094.

Figures

Figure 1: An integrated approach for data acquisition and online reconstruction was used. First, a “dummy” repetition is performed to reach steady state and train the reconstruction’s respiratory bins. Subsequent acquisitions are then binned based on the trained respiratory thresholds, tracking the percentage of completed readouts for each bin. Once a threshold (95%) is reached for all bins, a termination signal is sent to the scanner, terminating acquisition at the end of the current repetition.

Figure 2: The delay of the feedback system was measured for 55 repetitions (3 subjects, 3 scans each, variable repetitions) of the MRF sequence using timestamped logs on the scanner. The average delay was 0.55+/-0.02 seconds. A corresponding “cutoff” time before which a termination must be requested before the end of a repetition based on this average delay is visualized on this figure as a yellow box. The delays seen over the first 2.5s (250 readouts) are noticeably longer due to the overhead of initializing a repetition on both the scanner and reconstruction end.

Figure 3: Respiratory waveform from a 6-repetition exam visualizing the respiration training “dummy” scans (highlighted in green) followed by the Inspiration, Transition, and Expiration phases separated by the trained threshold values. In the above dataset, the detected mean respiratory signal during training was 0.58+/-0.32, yielding cutoff thresholds at 0.38 for inspiration and 0.77 for expiration

Figure 4: T1 and T2 quantitative map comparisons between breath-hold and free-breathing binned acquisitions using the same MRF sequence parameters. Liver and kidneys are blurry in the no binning case T1 and T2 maps. Lines were drawn to illustrate the positional match between the breath-hold and free-breathing acquisitions.