Pulse Sequence Design for MR-Guided Interventions
Adrienne Campbell-Washburn1
1National Institutes of Health, United States

Synopsis

Keywords: Image acquisition: Sequences, Physics & Engineering: Interventional

During an MR-guided intervention, an invasive procedure is performed while the patient is inside the MRI system, and image guidance is used for real-time device navigation or to assess the therapeutic progress. The workflow for MRI-guided interventions differs from typical diagnostic MRI, and these procedures are more demanding on image acquisition and reconstruction times. Therefore, pulse sequences are designed differently. This talk will describe: interactive real-time imaging to enable the navigation of devices through complex anatomy, sequences to visualize interventional devices, thermometry methods to monitor thermal interventions, tracking motion for MRI-guided radiotherapy, and sequence compatibility with low-latency reconstruction.

Introduction

MR-Guided Interventions describes a family of exams in which an invasive procedure is performed while the patient is inside the MRI system. Image guidance is used intermittently intraprocedurally to assess the progress of the procedure, or for real-time guidance, for example to navigate a deice or monitor the effect of a therapy as it’s being applied. Image guidance improves the efficiency and safety of minimally invasive procedures. X-Ray, ultrasound, intracardiac echo, optical coherence tomography, and MRI all offer image guidance for procedures, with MRI being attractive for procedures that require soft tissue visualization or quantification of function/physiology. Examples of clinical MRI-guided procedures include MRI-guided biopsy (e.g., breast, prostate, liver), MRI-guided ablation (including radiofrequency-, laser-, cryo-ablation), invasive hemodynamic cardiac catheterization, intraoperative MRI, MRI-guided radiotherapy, and MRI-guided high-intensity focused ultrasound thermal therapy.

The requirements of MRI-guided intervention workflows are different than typical diagnostic MRI workflows. First, image-guided procedures are exceptionally demanding on imaging speed and low-latency image reconstruction, since imaging feedback is required for interventionists to make procedural decisions. The exact speed requirements depend on the procedure, varying from real-time imaging at up to 10 frames per second to several-minute imaging times. Second, real-time imaging often must be equipped with interactive capabilities. Throughout a procedure, interventionists may require on-the-fly changes to the slice position, image contrast, acceleration rate, and slice thickness. Therefore, it is essential that these parameters can be interactively modified during real-time imaging throughout the procedure. Third, interventional devices including catheters, guidewires, needles, bioptomes, etc, must be considered. Either the devices are visualized in relation to the surrounding tissue, or their artifacts must be mitigated to generate high quality images . This talk will specifically focus on pulse sequence design for MRI-guided precedures.

Interactive real-time imaging

For procedures that require “real-time imaging” (defined here as frame rates 3-10 frames per second), rapid parallel imaging is used to reconstruct real-time data for procedural navigation (e.g. GRAPPA, SENSE, TGRAPPA, TSENSE, etc.). Alternatively, non-Cartesian trajectories can offer improved frame rates and are attractive for 2D real-time imaging. Cardiac procedures tend to have the most demanding frame-rate requirements.

For certain procedures, interactive modification of parameters during real-time imaging is required [1]. Examples of interactive real-time imaging parameters are:

  • Imaging slice position and orientation: used to track device motion.
  • Slice thickness: increased to reduce the out-of-plane device motion during respiration and cardiac motion, or decreased to improve tissue visualization.
  • Image acceleration rate: increased during complex procedural steps that require rapid image update, typically implemented using a TGRAPPA image reconstruction.
  • Image contrast: Magnetization preparation pulses are used to modify image contrast in real-time.

Software that offers a graphical user interface to interactively change parameters during real-time imaging, to visualize devices, to annotate the procedure, and to retrospectively review cases is important for the success of real-time MRI-guided procedures.

Device visualization

There are several types of devices used during IMRI procedures, and sequences must be designed to visualize these devices. “Passive” devices are those that are visualized using the imaging signature from their inherent material properties, and so-called “active” devices contain receiver electronics and are visualized based on the spatial localization of the signal that they receive.

Passive device visualization includes imaging of susceptibility artifacts, for example to visualize metallic guidewires, metallic needles, metal-braided catheters, and polymer devices with iron oxide or stainless-steel markers. Different metals that are commonly used to manufacture medical devices have different imaging properties, for example nitinol and stainless steel 316 produces subtle artifacts, whereas stainless steel 304 produces substantial artifacts that, in severe cases, can destroy the surrounding MRI signal and are saturated even at low field strengths [2-4]. Another passive device visualization option is the use of a positive contrast agent, such as gadolinium. For example, gadolinium-filled balloon catheters are commonly used during human MRI-guided right heart catheterization. A saturation pulse can be used prior to real-time image acquisition to increase the contrast of a gadolinium-filled device compared to the background. One limitation of passive device visualization is the poor specificity, as susceptibility artifacts and regions of positive contrast may be similar to anatomical structures.

Active device visualization uses the signal collected from the device itself, and there are two common configurations: microcoils and loopless antennae. Microcoils are small solenoids added to devices that provide a “point-source” signal. Typically, 3 orthogonal projection images are acquired to localize the microcoils in 3D space. The location and orientation of the device (orientation estimated based on 2 microcoils) is displayed as an overlay on a pre-acquired 3D roadmap. Loopless antennae generate signal along the entire length of a device, which is overlaid on the anatomy images, allowing visualization of the shaft position and curvature, in addition to the tip location.

Thermometry

MRI thermometry is a standard method to assess tissue damage during ablative MRI-guided procedures [5, 6]. Several MRI parameters are temperature-dependent, including T1, T2, and Larmor frequency. Proton resonance frequency-based thermometry is the preferred method, due to the high temporal resolution and sensitivity. Temperature maps are derived from gradient echo images by measuring the phase evolution resulting from a temperature-dependent change in Larmor frequency.
PRF thermometry measures temperature-induced phase changes in spoiled GRE images, as follows: $${\Delta T = \frac{\Delta \Phi}{\gamma \alpha B_0 TE} }$$
where T is temperature, $$${\Phi}$$$ is phase, $$${\gamma}$$$ is the gyromagnetic ratio, $$${\alpha}$$$ is the thermal coefficient for aqueous tissue (typically $$${\alpha}$$$ = -0.01 ppm/°C). Notably, PRF shift is absent in fat, and therefore these techniques cannot be used in adipose tissue, and PRF thermometry is not suitable for frozen tissue, ie. cryoablation.

Typically, phase is measured relative to baseline images acquired prior to heating, however, if there is motion between acquisitions (eg. respiratory motion or bulk motion), artifacts are introduced to the temperature map. Referenceless thermometry is a popular method to eliminate the dependence on baseline images, where the non-heated surrounding tissue is used to estimate the expected phase within the heated region with the assumption that the phase will be smoothly varying. Alternatively, multibaseline methods use several images acquired across varying physiological motion states and cross correlation to select a similar baseline image for subtraction relative to the image acquired during heating.

Temporal resolution requirements depend on application, but ranges from seconds to hundreds of milliseconds. For example, a temporal resolution of approximately 3s and spatial resolution of 1mm2 is used for static organs [6], HIFU procedures have targeted 10 frames per second (100ms temporal resolution) [7], and cardiac applications target <1s temporal resolution [8]. As with all IMRI application, latency between acquisition and temperature map display is minimized, with a target of <100ms latency to avoid unintended tissue damage.

MRI-guided radiotherapy

MRI offers value for several stages of radiotherapy procedures. MRI can be used for procedural planning, for imaging between fractional doses of radiation (interfraction), and during radiation delivery (intrafraction). New hybrid systems that combine MRI and linear accelerator (MR-linac) radiotherapy systems have enabled improved image-guided radiotherapy.

Organ motion needs to be carefully tracked for MRI-guided radiotherapy applications. Approaches to pair some pre-acquired data with real-time imaging during radiation delivery have been proposed to enable volumetric real-time imaging. For example, MR Signature Matching (MRSIGMA) uses a dictionary of pre-acquired motion states paired with fast acquisitions and dictionary matching during real-time radiation delivery [9]. Alternatively, MR-MOTUS (Model-based Reconstruction of MOTion from Undersampled Signals) augments a precomputed motion model with minimal k-space data acquired in real-time [10, 11].

Low latency reconstruction

While this talk will primarily focus on pulse sequences, low-latency reconstruction will be briefly included, as low latency image reconstruction is critical for the IMRI workflow and has implications on data sampling choices. Latency refers to the time between the end of k-space acquisition and the display of that image. Procedural decisions are made based on the images generated during IMRI procedures. For example, interventionists may need to navigate a catheter device to a specific location using real-time imaging, or to assess the position of a needle relative to a lesion for ablation, or may adjust radiotherapy dose based on intraprocedural imaging. These examples illustrate the need for rapid image reconstruction where the IMRI procedure cannot proceed without the relevant imaging information. The tolerable reconstruction latency depends on the application, but if the latency exceeds the acquisition time, a lag is introduced, and the images cannot be safely used to guide the procedural steps.

Highly undersampled acquisitions that require prolonged computationally intense image reconstructions (e.g., compressed sensing) are often incompatible with the IMRI environment. Computational hardware is increasingly available and affordable, which has been important for the deployment of advanced image reconstruction methods within the IMRI workflow. For example, graphical processing units (GPUs) plus the advent of open application programming interfaces for commodity GPUs, eg. CUDA, was critical in making this hardware accessible for MRI applications [12, 13]. Iterative reconstructions, such as non-Cartesian parallel imaging and compressed sensing benefit the most from GPU hardware, and have been used for real-time non-Cartesian imaging in the IMRI environment.

Reconstruction algorithms can be specifically optimized to minimize latency by design and implementation details. For example, the choice of solver, precomputation of matrices used repeatedly during iterative reconstruction, and choices of regularizers can all have an impact on computation time. NLINV (non-linear inversion) is a specific reconstruction method designed for low latency [14]. NLINV combines undersampled spoiled gradient echo radial imaging with an iterative reconstruction. Computation is minimized by combining operations as simpler steps, and pre-computing where possible. The NLINV approach has been applied to MRI-guided endomyocardial biopsy with an 42ms temporal resolution.

Most recently, deep learning has also been deployed for rapid image reconstruction. Importantly, training of neural networks can be time-consuming and requires substantial computational power, but once trained, model inference is rapid and uses only modest computational hardware. Deep learning has been used specifically for real-time cardiac imaging and device navigation with low latency and high image quality [15].

Conclusion

MRI-guided procedures impose demands on image acquisition and reconstruction times which inform pulse sequence decisions. The workflow of MRI-guided interventions differs from diagnostic imaging, and the pulse sequences are adapted to this workflow. Several tools exist to enable MR image guidance of invasive procedures, as discussed during this talk.

Acknowledgements

No acknowledgement found.

References

  1. Campbell-Washburn, A.E., et al., Magnetic Resonance Sequences and Rapid Acquisition for MR-Guided Interventions. Magn Reson Imaging Clin N Am, 2015. 23(4): p. 669-79.
  2. Peeters, J., et al., Development and testing of passive tracking markers for different field strengths and tracking speeds. Physics in Medicine & Biology, 2006. 51(6): p. N127.
  3. Peeters, J.M., E.E. van Faassen, and C.J. Bakker, Magnetic resonance imaging of microstructure transition in stainless steel. Magnetic resonance imaging, 2006. 24(5): p. 663-672.
  4. Basar, B., et al., Susceptibility artifacts from metallic markers and cardiac catheterization devices on a high-performance 0.55 T MRI system. Magn Reson Imaging, 2021. 77: p. 14-20.
  5. Rieke, V. and K. Butts Pauly, MR thermometry. J Magn Reson Imaging, 2008. 27(2): p. 376-90.
  6. Winter, L., et al., Magnetic resonance thermometry: Methodology, pitfalls and practical solutions. Int J Hyperthermia, 2016. 32(1): p. 63-75.
  7. Ries, M., et al., Real-time 3D target tracking in MRI guided focused ultrasound ablations in moving tissues. Magn Reson Med, 2010. 64(6): p. 1704-12.
  8. Yon, M., et al., Continuous cardiac thermometry via simultaneous catheter tracking and undersampled radial golden angle acquisition for radiofrequency ablation monitoring. Sci Rep, 2022. 12(1): p. 4006.
  9. Feng, L., N. Tyagi, and R. Otazo, MRSIGMA: Magnetic Resonance SIGnature MAtching for real-time volumetric imaging. Magn Reson Med, 2020. 84(3): p. 1280-1292
  10. Huttinga, N.R.F., et al., MR-MOTUS: model-based non-rigid motion estimation for MR-guided radiotherapy using a reference image and minimal k-space data. Phys Med Biol, 2020. 65(1): p. 015004.
  11. Terpstra, M.L., et al., Real-time 3D motion estimation from undersampled MRI using multi-resolution neural networks. Med Phys, 2021. 48(11): p. 6597-6613.
  12. Hansen, M.S., D. Atkinson, and T.S. Sorensen, Cartesian SENSE and k-t SENSE reconstruction using commodity graphics hardware. Magn Reson Med, 2008. 59(3): p. 463-8.
  13. Sørensen, T.S., et al., Real-time reconstruction of sensitivity encoded radial magnetic resonance imaging using a graphics processing unit. IEEE Trans Med Imaging, 2009. 28(12): p. 1974-85.
  14. Uecker, M., S. Zhang, and J. Frahm, Nonlinear inverse reconstruction for real-time MRI of the human heart using undersampled radial FLASH. Magn Reson Med, 2010. 63(6): p. 1456-62.
  15. Jaubert, O., et al., Real-time deep artifact suppression using recurrent U-Nets for low-latency cardiac MRI. Magn Reson Med, 2021. 86(4): p. 1904-1916.
Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)