Debra McGivney1, Yun Jiang1, Dan Ma1, Chaitra Badve2, Vikas Gulani1, and Mark Griswold1
1Radiology, Case Western Reserve University, Cleveland, OH, United States, 2Radiology, University Hospitals, Cleveland, OH, United States
Synopsis
A
Bayesian methodology has been previously applied to MRF reconstructions to
analyze subvoxel T1 and T2 distributions. The multidimensional results from
this algorithm are difficult to visualize. We propose to apply this Bayesian approach
in the brain to create T1 and T2 Gaussian distributions to represent various tissue
types. Using these distributions as prior information, the Bayesian methodology
is applied over the brain with a smaller dictionary. Results from this Bayesian
approach with a smaller dictionary are weighted by the Gaussian probabilities
and summed to create tissue probability maps in normal volunteers and a brain
tumor patient.
Purpose
The unique signal evolution structure in magnetic resonance
fingerprinting (MRF)1 offers an opportunity to separate multiple
components within a single voxel. Recent work on the Bayesian methodology for
partial volume analysis in MRF2 shows that mixed signal evolutions can be
separated with minimal prior information using a full MRF dictionary of
simulated signal evolutions. This is especially important when one would like
to segment an organ into specific tissue types. In this methodology, neither
the number or types of subvoxel tissues need to be specified in advance, and
the result is a distribution of tissue property values to explain the voxel
composition, rather than one single set of effective tissue property values. One
issue with this approach is the increased dimensionality of the solution
over the traditional MRF tissue property maps, which makes visualization challenging.
Additionally, when two tissues within a voxel have similar relaxation properties, for example, in white matter and gray matter, an approach that
uses the full dictionary, such as this one2, may not be able to resolve these tissues
clearly. Here we present a modification of the Bayesian methodology to provide
tissue property probability maps, using prior information from targeted
ROI analysis within the original Bayesian methodology for partial volume in normal
volunteers and a brain tumor patient. Methods
Two normal volunteers were scanned in an IRB-approved study at 3T using
MRF-FISP3. Regions of interest (ROIs) were
chosen in each volunteer for targeted subvoxel analysis using the full dictionary with the Bayesian approach. For the normal volunteers, regions were
selected in pure white matter, gray matter, and CSF. The regions for one
volunteer are shown in Figure 1. The T1, T2 results were combined across all
regions, and a Gaussian mixture model was applied. Three Gaussians were
selected by considering the maximum mixing probabilities of the distributions,
which resulted in one distribution to model each tissue type. Using this ROI
analysis as prior information, a smaller dictionary was formed, containing only
the dictionary entries corresponding to the signal evolutions from the T1, T2
values that fell within the predefined distributions. Additionally, the maximum
probability for each dictionary entry is saved, which indicates the distribution
to which the entry is most likely to belong. This smaller dictionary for one
volunteer contained 426 T1 and T2 combinations (compared to the full dictionary
containing 5970) and was subsequently used in a multicomponent Bayesian
analysis over the whole 2D slice. Results from this second analysis were again
multidimensional. Tissue property probability maps were computed for the tissue
types by computing a weighted sum of the the Bayesian result from the small dictionary
multiplied with the precomputed tissue probabilities. Additionally, the method
was applied to one patient who presented with a pathology-proved glioblastoma
(GBM)4. This patient was scanned at 3T with MRF-bSSFP1. To account for pathological tissues, regions
within the solid tumor and peritumoral white matter were also included in the
initial ROI analysis. Results
Figure
2 shows the Gaussian distributions as 95% credibility ellipses, formed from Bayesian
subvoxel analysis on the ROIs from Figure 1. The distributions were computed by
applying the expectation-maximization model for Gaussian mixture models. The
distributions with the three largest final mixing probabilities were retained
in the case of the normal volunteers, resulting in one distribution each to
describe the three tissue types. Note that the standard deviation of the
ellipses is impacted, in part, by the varying step size in the dictionary. For the patient, 10 Gaussians were generated,
due to the question of the true T1 and T2 of the pathological tissues, and
three were discarded due to low mixing probabilities. Probability maps for the
normal volunteers are shown in Figures 3 and 4, and for the patient in Figure
5.Discussion and Conclusion
We present one possible solution to the visualization and interpretation
of the results from a blind multicomponent analysis in MRF, such as the
Bayesian analysis2 presented previously. We believe that this will
have specific use in segmenting regions of the body into specific tissue types.
One advantage of this approach is a significant savings in computation time,
since after the initial ROI analysis utilizing the full dictionary, a smaller dictionary
containing about one tenth of the number of entries can be used to compute the
tissue property probability maps over the entire slice. A disadvantage of the
approach is the inclusion of additional prior information into the Bayesian
approach, which will clearly bias the solution based on the chosen ROIs. However,
we believe that the SNR and clear definition of these tissues in the brain
could lead to multiple applications, including surgical planning and the
diagnosis of small gray matter lesions. Acknowledgements
The
authors would like to acknowledge funding from Siemens and NIH grants
1R01EB016728-01A1 and 5R01EB017219-02.References
1. Ma, D. et al.
Magnetic resonance fingerprinting. Nature 495, 187–192 (2013).
2. McGivney, D. et al. Bayesian estimation of multicomponent relaxation parameters in magnetic resonance fingerprinting. Magn. Reson. Med. (accepted).
3. Jiang, Y., Ma, D., Seiberlich, N.,
Gulani, V. & Griswold, M. MR fingerprinting using fast imaging with steady
state precession (FISP) with spiral readout. Magn. Reson. Med. 74,
1621–1631 (2015).
4. Badve, C. et al. MR
fingerprinting of adult brain tumors: Initial experience. Am. J.
Neuroradiol. 38, 492-499 (2017).