Synthetic Quantitative MRI through Relaxometry Modelling for Improved Brain Segmentation
Martina F Callaghan1, Siawoosh Mohammadi1,2, and Nikolaus Weiskopf1,3

1Wellcome Trust Centre for Neuroimaging, UCL Institute of Neurology, London, United Kingdom, 2Department of Systems Neuroscience, University Medical Center Hamburg-Eppendorf, Hamburg, Germany, 3Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany

Synopsis

Here we exploit the inter-dependence of quantitative MRI (qMRI) parameters via relaxometry modelling to generate synthetic quantitative maps of magnetisation transfer saturation. The utility of the new concept of synthetic quantitative data is demonstrated by improving image segmentation of deep gray matter structures for neuroimaging applications.

Purpose

qMRI aims to produce measurements directly related to tissue microstructure with high diagnostic and research value that are independent of scanner and acquisition protocol (1). As part of a multi-parameter mapping (MPM) protocol (2), maps of the effective transverse relaxation rate (R2*) can be calculated by performing a fit to the signal decay of a multi-echo FLASH volume. Maps of the longitudinal relaxation rate (R1) can be calculated from two high resolution 3D FLASH acquisitions with different excitation flip angles, and calibration data mapping the participant-specific transmit field. An additional magnetisation transfer-weighted volume is required in order to calculate a map of magnetisation transfer saturation (MT) (3). The acquisition of this extra data increases the exam duration and thus the risk of data corruption due to participant motion. However MT maps have greater specificity to macromolecular content and have been shown to improve the segmentation of deep gray matter (GM) structures as compared to standard T1-weighted data (4). Here we used a linear relaxometry model of R1 (5) to synthesise an MT map from a subset of the MPM protocol in which no MT-weighted data were acquired. We then assessed whether the established improvement in segmenting deep GM structures was maintained with synthetic MT maps.

Methods

MPM data were acquired on a 3T Tim Trio (Siemens Healthcare) system on a group of 30 healthy volunteers (13 male, age range 18-25 years, mean 21.6 years, std. dev. 1.9 years). The mean values for the linear relaxometry model coefficients reported in (5) were used to calculate synthetic MT maps using only R1 and R2* maps. Deep GM structures were identified by segmenting the T1-weighted volume, the R1 map and the synthetic MT map and the performance was compared using a voxel-wise t-test on the GM tissue probabilities with a significance threshold of p<0.05 after small volume (central brain) and family-wise error corrections for multiple comparisons. All data processing was performed within the SPM12 framework (Wellcome Trust Centre for Neuroimaging, London).

Results

The GM probability of the synthetic MT map was found to be significantly higher than that of the T1-weighted data in the left pulvinar nucleus, thalamus and pallidum; within the brainstem, particularly the substantia nigra (SN) and the pons. It was significantly higher than that of the R1 map in the pallidum and focally within the right SN. The GM probability of the MT map was significantly higher than that of the synthetic MT map in distributed GM regions encompassing the pallidum, extending into the right putamen, in the SN, gyrus rectus and in multiple white matter regions.

Conclusions

The utility of this novel concept of synthesising quantitative MR data has been demonstrated by maintaining improved segmentation performance of deep GM structures while reducing the overall scan time. However, the measured MT map remains the optimum choice for segmentation of deep GM structures.

Acknowledgements

This work was supported by the Wellcome Trust and the ERC.

References

1. Weiskopf N, et al. Curr. Opin. Neurol. 2015;28:313–22

2. Weiskopf N, et al.Front. Neurosci. 2013;7:1–11

3. Helms G, et al. Magn. Reson. Med. 2008;60:1396–407

4. Helms G, et al. Neuroimage 2009;47:194–8

5. Callaghan MF, et al. Magn. Reson. Med. 2015;73:1309–1314.

Figures

Fig.1: Regions showing higher (red) or lower (blue) GM probability when using a synthetic MT map for segmentation as compared to: (A) a T1-weighted image, (B) a quantitative R1 map and (C) an MT map within the ROI outlined in black, displayed at p<0.001 uncorrected for multiple comparisons.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1160