4932

Quantitative susceptibility mapping for routine clinical use – An inline automated QSM reconstruction pipeline
Ashley Wilton Stewart1, Kieran O'Brien1,2, Jinsuh Kim3, Bénédicte Maréchal4,5,6, Fatimah Nasrallah7, Michael Kean8, Markus Barth1, and Steffen Bollmann1

1Centre for Advanced Imaging, University of Queensland, Brisbane, Australia, 2Siemens Helathcare Pty Ltd, Brisbane, Australia, 3Department of Radiology, University of Alabama Birmingham, Birmingham, AL, United States, 4Advanced Clinical Imaging Technology, Siemens Helathcare, Lausanne, Switzerland, 5Department of Radiology, CHUV, Lausanne, Switzerland, 6LTS5, EPFL, Lausanne, Switzerland, 7Queensland Brain Institute, University of Queensland, Brisbane, Australia, 8Royal Children's Hospital, Melbourne, Australia

Synopsis

Quantitative susceptibility mapping (QSM) is a post-processing technique for gradient-recalled-echo (GRE) phase data, which provides information about tissue composition complementary to common Susceptibility Weighted Imaging (SWI). To date, QSM’s multiple complex processing steps has limited its clinical application. In this work, we present an automated and robust inline QSM post-processing pipeline compatible with flow-compensated GRE and VIBE sequences. The QSM pipeline includes morphological and atlas-based segmentation, two different QSM algorithms and is compatible with SWI processing. Two clinical cases of QSM in Traumatic Brain Injury and Multiple Sclerosis are presented.

Target Audience

Clinicians & Researchers interested in translating quantitative susceptibility mapping (QSM) to the clinic.

Purpose

Quantitative susceptibility mapping (QSM) is a gradient-recalled echo (GRE) phase-based post-processing technique which is sensitive to the local susceptibility effects of biometals such as diamagnetic calcium and paramagnetic iron within tissue. Iron, a redox-active transition metal1, is of particular interest as it is implicated in the etiopathogenesis of many diseases such as Alzheimer’s, Parkinson’s, and Multiple Sclerosis1-4. Currently, QSM processing pipelines that are readily available run offline5-9, and are further hampered by conceptually simple but non-trivial masking steps, which propagate errors into the final maps. Currently, the clinical uptake of QSM is impeded by the required manual intervention and computational resources and the complexity and impracticality of offline solutions.

In this work, we present flow-compensated (FC) GRE and volumetric interpolated breath-hold examination (VIBE) prototype sequences that are compatible with two inline QSM post-processing pipelines. We further propose a robust atlas-based segmentation of the cranial cavity to automate the brain masking procedure and improve the reproducibility and integration of QSM into the clinical workflow.

Methods

Estimating a local susceptibility distribution from MRI signal phase involves multiple complex post-processing steps including: coil combination; phase unwrapping; background phase removal; and the solution to the ill-posed field-susceptibility inversion10. We have implemented single-step QSM7 and Total Generalized Variation (TGV) QSM8, summarized in Figure‑1, inline on the MR scanner.

QSM requires the masking step to distinguish between the unwanted background fields related to tissue-air interface from the desired local susceptibility effects of the tissue. Morphological threshold-based methods are commonly employed, but can result in sub-optimal masking results. To mitigate the propagation of these potential errors, we incorporated an atlas-based segmentation for the brain. Specifically, a manual mask of the cranial cavity is codified in a population template constructed using minimum deformation averaging12 (Figure-2). This template is deformed and aligned to the patient volume by estimating a free-form diffeomorphic displacement field using a fast multi-resolution iterative scheme that maximizes the local correlation between the template and the incoming scan12. The mask is automatically deformed to adapt to the reference volume, providing a subject-specific brain mask for the subsequent background removal steps.

The prototype QSM pipelines have been tested on MAGNETOM Prisma and MAGENTOM Skyra 3T MR scanners using 64-channel head/neck and 32-channel head coils (Siemens Healthcare, Erlangen, Germany). We optimized and compared a TA=4min52s 3D prototype FC-GRE QSM/SWI compatible whole-brain acquisition (GRAPPA acceleration=3) against a TA=2min43s 3D prototype FC-VIBE QSM whole-brain acquisition (CAIPIRINHA acceleration=3x2). Other image parameters: TR=25ms/TE=20ms/α=12deg; FOV=204x224x160mm3/matrix=204x224x160. The QSM/SWI compatible acquisition was applied clinically in traumatic brain injury and multiple sclerosis patients.

Results & Discussion

Figure-3 shows the influence of different population atlases generated based on a variety of echo-times on the brain extraction performance and consequently QSM reconstruction quality for the TGV algorithm. We achieve the best registration and segmentation results with an echo-time-matched population atlas, and a similar performance level using an atlas averaged across all available echo times (see Figure-3, zoomed sections).

Figure-4 shows the potential of QSM over SWI in a traumatic brain injury patient using the implemented Total Generalized Variation (TGV) QSM algorithm. SWI phase images show artefacts that interfere with the clear differentiation of calcium or iron causing the hypo-intensity in the SWI magnitude images. On the other hand QSM provides unambiguous definition.

Figure-5 shows that QSM provides complimentary information that can potentially help discriminate acute MS lesions, and characterize the inflammatory process in chronic non-enhancing lesions when the BBB has recovered. For this clinical case we utilized the single-step (SS) QSM algorithm.

Conclusions

We have developed a robust and fully automated prototype capable of producing inline QSM and SWI images. Optimizing the atlas-based segmentation reduced artefacts in the QSM processing that can be attributed to incorrect background field removal. This is a critical step to enable the successful validation of this promising technique in a large range of clinical applications.

Acknowledgements

SB acknowledges funding from UQ Postdoctoral Research Fellowship grant. MB acknowledges funding from ARC Future Fellowship grant FT140100865. FN acknowledges funding from the Motor Accident Insurance Commission (grant number: 2014000857). The authors acknowledge the facilities and scientific and technical assistance of the National Imaging Facility, a National Collaborative Research Infrastructure Strategy (NCRIS) capability, at the Centre for Advanced Imaging, The University of Queensland.

References

  1. Vladimir V (2018) Magnetic, Ferroelectric, and Multiferroic Metal Oxides in Metal Oxides, Elsevier
  2. Duyn, J. (2013) MR susceptibility imaging. J. Magn. Reson.
  3. Haacke, E. M. et al. (2014). Quantitative susceptibility mapping: current status and future directions. Magn. Reson. Imaging
  4. Wang, Y. et al. (2017), Clinical quantitative susceptibility mapping (QSM): Biometal imaging and its emerging roles in patient care. J. Magn. Reson. Imaging
  5. Liu, T et al. (2012) Accuracy of the Morphology Enabled Dipole Inversion (MEDI) Algorithm for Quantitative Susceptibility Mapping in MRI. IEEE Trans. Med. Imaging
  6. Liu, C. (2010) Susceptibility Tensor Imaging. Magn. Reson. Med. Off. J. Soc. Magn. Reson. Med. Soc. Magn. Reson. Med.
  7. Chatnuntawech, I. et al. (2016) Single-step quantitative susceptibility mapping with variational penalties. NMR Biomed
  8. Langkammer, C; et. al. (2015) Fast Quantitative Susceptibility Mapping using 3D EPI and Total Generalized Variation. NeuroImage.
  9. Li, W., Wu, B. & Liu, C. (2011) Quantitative susceptibility mapping of human brain reflects spatial variation in tissue composition. NeuroImage
  10. A. Deistung et al (2017) Overview of quantitative susceptibility Mapping NMR Biomed
  11. Jellus V. and Kannengiesser S. (2014) Adaptive Coil Combination Using a Body Coil Scan as Phase Reference Proceeding of ISMRM Milan
  12. Janke, A. L. & Ullmann, J. F. P. (2015). Robust methods to create ex vivo minimum deformation atlases for brain mapping. Methods

Figures

Figure-1 shows the individual steps of both implemented pipelines. The Singe-step QSM7 was written in C++, and Total Generalized Variation (TGV) QSM8 was written in in C++ using CUDA to utilise the 2 Tesla K10 GPU accelerators (2x1536 cores, 8GB, NVIDIA) available on Siemens MR scanners.

Figure-2 (A) illustrates the construction of a template, and mask, that represents population anatomy by means of minimum deformation averaging12. Atlas-based segmentation is performed by registering the population template to the participant’s incoming scan and propagating the mask into the participant’s space (B).

Figure-3 illustrates that the echo time of the atlas used for brain extraction affects the segmentation quality and that we can minimize artifacts (yellow arrows) using either echo-time matched atlases or as a compromise an atlas including all available echo times across the echo time range.

Figure-4 shows image examples from a patient with traumatic brain injury, a microbleed is present in two different slices. Whereas the QSM/SWI FC GRE clearly provides confirmation of the microbleeds, the filtered Phase, SWI MIP and SWI are inconclusive due to a residual artefact from the high-pass filtering. The QSM reconstructions where performed inline using the Total Generalized Variation (TGV) algorithm in Siemens ICE.

Figure-5 provides image examples of 3 MS lesions from one patient with different signal characteristics between QSM/SWI FC-GRE acquisition and the post-Gd+ Flair and MPRAGE. The differences in signal may be reflective of the age of the lesion and its current inflammation state. The QSM reconstructions where performed inline using the single-step (SS) algorithm in Siemens ICE.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
4932