Elsevier

NeuroImage

Volume 49, Issue 2, 15 January 2010, Pages 1289-1300
NeuroImage

An optimised framework for reconstructing and processing MR phase images

https://doi.org/10.1016/j.neuroimage.2009.09.071Get rights and content

Abstract

Phase contrast imaging holds great potential for in vivo biodistribution studies of paramagnetic molecules and materials. However, in vivo quantification of iron storage and other paramagnetic materials requires improvements in reconstruction and processing of MR complex images. To achieve this, we have developed a framework including (i) an optimal coil sensitivity smoothing filter for phase imaging determined at the maximal signal to noise ratio, (ii) a phase optimised and a complex image optimised reconstruction approach, and (iii) a magnitude and phase correlation test criterion to determine the low pass filter parameter for background phase removal. The method has been evaluated using 3T and 7T MRI data containing cortical regions, the basal ganglia including the caudate, and the midbrain including the substantia nigra. The optimised reconstruction improves phase image contrast and noise suppression compared with conventional reconstruction approaches, and the correlation test criterion provides an objective method for separation of the local phase signal from the background phase measurements. Phase values of several brain regions of interest have been calculated, including gray matter (− 1.23 Hz at 7T and − 0.55 Hz at 3T), caudate (− 3.8 Hz at 7T), and the substantia nigra (− 6.2 Hz at 7T).

Introduction

Magnetic resonance phase is known to be related to the underlying tissue susceptibility (Marques and Bowtell, 2005, Rauscher et al., 2005). Susceptibility weighted phase images are therefore potentially useful in the measurement of iron storage in the normal brain (Drayer et al., 1986, Haacke et al., 2005) as well as in the study of many neurodegenerative diseases, such as Huntington's disease, Alzheimer's disease, Parkinson's disease and multiple sclerosis (Bartzokis and Tishler, 2000, Kosta et al., 2006, Neema et al., 2007).

The advantages of MR phase images acquired with increased magnetic field strength and multiple receiver channels have been demonstrated and the possible applications of phase imaging have been discussed (Haacke et al., 2007, Duyn et al., 2007, Hammond et al., 2008, Yao et al., 2009). Although MR phase imaging shows great potential in several applications, there are a number of difficulties associated with quantitative analysis of phase images. Firstly, susceptibility values in many brain tissues are relatively small, and the induced phase signals of these regions of the brain therefore have low values. The reconstruction of phase images needs to be carefully optimised to address this issue. As discussed in this paper, current reconstruction methods are optimised only for magnitude images, and yield sub-optimal solutions for phase images. Furthermore, reconstructed phase images are wrapped into the (− π, π] range, and are corrupted with a background phase due to imperfect shimming. Therefore, phase unwrapping and background phase removal procedures are necessary. The background phase removal process requires filtering of reconstructed phase images. Consequently the contrast of the output phase images is dependent on the filter parameters used. Because of these issues in phase imaging, significant variations in phase measurements have been reported in the literature (Haacke et al., 2007, Hammond et al., 2008, Yao et al., 2009). These problems associated with phase reconstruction and processing methods are addressed in this paper.

The first step in generating phase images is image reconstruction from the k-space signals. Because multiple receive channels are now a standard MR scanner configuration, parallel image reconstruction is required. There are two widely used methods, SENSitivity Encoding (SENSE) (Pruessmann et al., 1999) and GeneRalized Autocalibrating Partially Parallel Acquisitions (GRAPPA) (Griswold et al., 2002). The GRAPPA method is used only for accelerated acquisitions (i.e. reduced k-space coverage). This method is in general an alternative to SENSE when the coil sensitivity information is not available. The SENSE method can be used for both the non-accelerated acquisition (i.e. full k-space coverage) and accelerated acquisition situations. SENSE can provide a near optimal reconstruction when the coil sensitivity profiles are known (Pruessmann et al., 1999). Recent extensions of SENSE are regularized SENSE methods (King and Angelos, 2001, Lin et al., 2004, Chen et al., 2007) and self-calibrated SENSE (Sodickson and McKenzie, 2001, Qian et al., 2004). These new extensions of the SENSE method can provide improved image signal to noise ratio (SNR) and reconstruction robustness. The SENSE method provides both magnitude and phase image reconstruction.

The weighted mean/average method for phase image reconstruction was originally introduced by Bernstein et al. (1994) and recently extended by Hammond et al. (2008). Compared with SENSE, the weighted mean method is easy to implement and has a lower computational requirement, but results in lower SNR phase images, as demonstrated by Lu et al. (2008). Furthermore, unlike SENSE, the weighted mean method cannot be used for accelerated data acquisition. To apply the SENSE method in phase imaging, the sensitivity estimation needs to be optimised for the phase reconstruction. The conventional SENSE method (Pruessmann et al., 1999) is only optimised for magnitude images because the widely used p-norm signal combination approach cannot correctly estimate the coil sensitivity profiles for the reconstruction of the phase images, resulting in reduced phase contrast. A more detailed discussion is given in the Theory section. A number of groups have noted that in dynamic phase imaging applications conventional SENSE cannot guarantee the absolute phase contrast in each imaging echo, and only preserves dynamic contrast amongst echoes (Ma et al., 2005, Brau et al., 2008). For better estimation of the coil sensitivities in SENSE phase image reconstruction, Yao et al. (2009) acquired an extra scan with a short TE to minimize the phase estimation error during the coil sensitivity calibration. However, whilst this approach may be feasible in phantom studies or for specific studies, it is difficult to apply in clinical applications. The use of coil sensitivity profiles from a different scan requires zero motion between scans, especially at high fields where a strong interaction can occur between the imaged object and the applied field. Furthermore, separate calibration scans require extra imaging time which is a prohibitive factor in some applications such as dynamic imaging.

To address these problems in phase image reconstruction, we have developed phase optimised and complex image optimised reconstruction approaches which outperform conventional SENSE reconstruction and the weighted mean reconstruction of phase images. The new reconstruction approaches are based on the self-calibrated SENSE idea (Sodickson and McKenzie, 2001, Qian et al., 2004). The self-calibration technique provides a robust and improved reconstruction without a separate calibration scan. The coil sensitivity functions for phase optimised reconstruction are obtained by maximising the SNR in the reconstructed phase images.

In this paper, we also address another major issue in phase image processing, namely background phase removal. This problem arises because phase images are significantly corrupted by a slowly varying background phase caused by the imperfect shimming of the main field. The background phase signal is normally removed using a low pass filter approach (Noll et al., 1991, Wang et al., 2000, Rauscher et al., 2008). However, the size of the filter alters the phase signals. Currently, the filter size chosen for the background phase removal is often heuristically determined, which gives rise to difficulties in reliably and reproducibly estimating phase signals. In this paper, a regional correlation test between the magnitude and phase is introduced to determine the low pass filter size for removing background phase variations while maximally retaining the local phase information.

In summary, we propose a framework to optimally reconstruct and process phase images. The optimised framework involves three novel approaches including: (i) selection of the coil sensitivity phase smoothing filter parameter by maximisation of the SNR in the reconstructed phase images, (ii) a phase optimised and a complex image optimised reconstruction approach, and (iii) a correlation test between reconstructed magnitude and phase images to determine the low pass filter parameters for background phase removal. Fig. 1 summarises the steps involved in the proposed framework.

High resolution 3T and 7T data acquired using T2⁎ weighted brain imaging were used to study the new method. Higher SNR and contrast in phase images were observed when compared with images using the conventional reconstruction methods, and phase values were measured after background phase removal using the filter size obtained with the magnitude and phase image correlation test.

Section snippets

Theory

This section provides details of the MR encoding, the sensitivity phase profile estimation and smoothing, the optimised image reconstruction, the phase unwrapping, and the background phase removal following the steps shown in Fig. 1.

MRI data acquisition and reconstruction parameters

MRI data were acquired on a 7T Siemens system (Siemens Medical Solutions, Erlangen, Germany) with an 8 channel transmit-receive head coil (Neuroscience Research Institute, Incheon, South Korea). One young normal volunteer participated in the study. Axial T2-weighted gradient echo (GRE) images were acquired with echo time (TE) = 21.6 ms, repetition time (TR) = 750 ms, flip angle = 30°, bandwidth = 30 Hz per pixel, slice thickness = 2 mm, FOV = 256 × 224 mm2 and matrix size = 1024 × 896. The spatial resolution

Phase sensitivity profile estimation and smoothing

To evaluate the influence of the sensitivity smoothing filter on the reconstruction, we measured phase values (Fig. 2d) and calculated the SNR in WM for a set of filters of varying size, the results of which are shown in Fig. 2e.

There were no significant phase value changes for smoothing filters from 0.75 to 7.5 mm. A wide range of filters (side length from 3.75 mm to 7.5 mm) produced high SNR, with the maximum SNR obtained for Ws = 6.25 mm. The sensitivity phase profiles were not effectively

Discussion

High field MRI enables the possibility of using structural phase images as biomarkers (Duyn et al., 2007). However, because of the difficulties associated with phase imaging, the quantification of phase signals in practice is difficult and great variability has been shown in the reported values. In this paper, we have developed an optimised framework for improved phase imaging.

One of the primary contributions of the current paper is the development of two new reconstruction approaches, phase

Conclusions

In this paper, we have developed a novel reconstruction and processing framework that includes (i) the selection of a Gaussian low pass smoothing filter for sensitivity phase estimation based on the maximisation of signal to noise ratio in phase images, (ii) new phase and complex image optimised reconstruction approaches, and (iii) an approach for recovering the local phase signal through maximising the correlation between magnitude and phase signals. High resolution 3T and 7T data have been

Acknowledgments

This work was supported by the Australian National Health Medical Research Council (Grant # 400317), the Australian Research Council (Grant # LX0774759), and the Australia Korea Foundation. This work was partly supported by the Ministry of Education, Science and Technology (MEST), Republic of Korea, and the Korea Science and Engineering Foundation (KOSEF) (Grant no. 2009-0065597).

References (34)

  • BernsteinM.A. et al.

    Reconstructions of phase contrast, phased array multicoil data

    Magn. Reson. Med.

    (1994)
  • Brau, A., Beatty, P., McKenzie, C., Yu, H., Shimakawa, A., Reeder, S., J.H., B., 2008. The impact of parallel imaging...
  • ChenZ. et al.

    Weighted H(infinity) optimization approach to parallel MR image reconstruction

    Conf. Proc. IEEE Eng. Med. Biol. Soc.

    (2007)
  • CukurT. et al.

    Multiple-profile homogeneous image combination: application to phase-cycled SSFP and multicoil imaging

    Magn. Reson. Med.

    (2008)
  • DrayerB. et al.

    MRI of brain iron

    AJR Am. J. Roentgenol.

    (1986)
  • DuynJ.H. et al.

    Highfield MRI of brain cortical substructure based on signal phase

    Proc. Natl. Acad. Sci. U. S. A.

    (2007)
  • GelmanN. et al.

    MR imaging of human brain at 3.0 T: preliminary report on transverse relaxation rates and relation to estimated iron content

    Radiology

    (1999)
  • Cited by (27)

    • Domain knowledge augmentation of parallel MR image reconstruction using deep learning

      2021, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Lastly, the experiments in this work focus on magnitude images for several contrasts and anatomical regions. The application to phase imaging (Chen et al., 2010a) including susceptibility contrasts would require the networks to be extended into the complex domain, such as using the method from Lee et al. (Lee et al., 2018). In this work, the idea of using knowledge of parallel imaging was implemented for the commonly used Unet architecture.

    • Foundations of MRI phase imaging and processing for Quantitative Susceptibility Mapping (QSM)

      2016, Zeitschrift fur Medizinische Physik
      Citation Excerpt :

      In the case of highly localized receive-coil sensitivity profiles (e.g. at ultra-high field strengths [17]) or high phase noise level (e.g. due to low magnitude signal) the associated assumptions or computations may break down, resulting in destructive interferences of the signal from different channels [143]. These destructive interferences are associated with increased phase noise and often also with severe phase singularities, i.e. wraps that terminate within the object instead of at the object boundaries [20,139,140,143]. Though novel strategies have been introduced recently to process phase with singularities and mitigate associated artifacts in subsequent processing steps [111,157], phase singularities are a nuisance and can hamper quantitative processing in the course of QSM [157].

    • Imaging of Iron

      2013, International Review of Neurobiology
      Citation Excerpt :

      A limitation of this method with regards to iron measurement is that phase values are influenced not only by the iron content but also by the difference in the magnetic susceptibility of the surrounding tissue. Background field low-spatial-frequency effects from air/tissue and tissue/bone interfaces lead to unwanted phase information and must be removed by a high-pass phase filter (Chen et al., 2010; Hammond et al., 2008). Symmetric central 64 × 64 filter was recommended as it removes low-spatial-frequency components, but it may cause also some loss of the phase contrast for the structures of interest (Wang et al., 2000).

    • The contribution of myelin to magnetic susceptibility-weighted contrasts in high-field MRI of the brain

      2012, NeuroImage
      Citation Excerpt :

      This discrepancy may originate from differences in the age and type of animals used. In addition, different ROIs were used and the high-pass filtering included here in the processing significantly affects the frequency contrast (Chen et al., 2010). In the current study, small increases in gray matter T2* were observed in cuprizone-treated mice.

    • Quantitative imaging of intrinsic magnetic tissue properties using MRI signal phase: An approach to in vivo brain iron metabolism?

      2011, NeuroImage
      Citation Excerpt :

      Since IFS-based QSM can only detect iron in a magnetic state, underestimation of iron could be caused by the presence of weakly magnetic or non-magnetic iron such as heme, transferrin or iron-sulfur cluster proteins. Connor et al. (1990), for example, reported a high relative abundance of transferrin in the cerebral cortex compared to the basal ganglia, with higher concentration in WM than in GM. Finally, there are also several other possible reasons; for example, the voxel susceptibility may also be affected by paramagnetic ions, such as copper or manganese (Popescu et al., 2009) or variations of the amount of heme iron, e.g., due to regional differences in the deoxygenated blood volume or in the oxygen extraction fraction (Haacke et al., 2004; Petridou et al., 2010).

    View all citing articles on Scopus
    View full text