Dahan Kim1,2, Laura Eisenmenger3, and Kevin M. Johnson3,4
1Department of Medical Physics, University of Wisconsin, Madison, WI, United States, 2Department of Physics, University of Wisconsin, Madison, WI, United States, 3Department of Radiology, University of Wisconsin, Madison, WI, United States, 4Department of Medical Physics, University of Wisconsin, Middleton, WI, United States
Synopsis
4D-flow MRI suffers from long scan time due to a minimum of
four velocity encodings necessary to solve for three velocity components and
the reference background phase. We examine the feasibility of using machine
learning (ML) to determine the background phase and hence three velocity
components from only three flow encodings. The results show that ML is capable
of estimating three-directional velocities from three flow encodings with high
accuracy (1.5%-3.8% velocity underestimation) and high precision (R2=0.975).
These findings indicate that 4D-flow MRI can be accelerated without requiring a
dedicated reference scan, with a scan time reduction of 25%.
Introduction
4D-flow
MRI suffers from long scan times resulting from necessary multiple velocity
encodings and cardiac gating. At minimum, four encodings are required to solve
voxel-wise for three velocity components and a reference background phase.1
The background phase is spatially complex in spoiled gradient echo (GRE) based
4D-flow; however, bSSFP-based flow imaging without phase reference has been
achieved in the neck using background phase fitting.2 This has not
been extended to spoiled-GRE imaging or other complex organs. Yet, recent
progress3 in the related problem of phase estimation in quantitative
susceptibility mapping (QSM) suggests machine learning may be capable of
learning spatially complex phase fields. Given this, a convolutional neural
network (CNN) has potential to estimate the background phase from three flow
encodings, without requiring a dedicated fourth reference scan. Here, we
examine the optimal strategies for using a CNN to accurately determine the
background phase and thus the three velocity components from only three flow
encodings, potentially reducing the scan time of 4D-flow MRI by 25%. In
particular, we investigate strategies to ensure unbiased estimation in vessels
and interrogate the errors from the method. Methods
Data
was included from subjects undergoing whole-brain 4D-flow4 with
4-point reference encoding and 0.7mm isotropic resolution. 3-directional
velocity components were determined using all 4 encodings (called ‘4-point velocities’)
using standard reconstruction and eddy-current phase corrections. 112 subjects
were used for training, while 28 subjects were used for validation and testing.
A fully 3D CNN with U-net5 architecture was used which takes
randomly selected 64x64x64 sub-volumes. The complex-valued images of the 3 flow
encodings were used as an input for the CNN, which learned to estimate 3
velocity components (called ‘3-point velocities’). We investigated two weighted
least squares loss functions: one in which the error was weighted by magnitude and
another in which the error was weighted by the pseudo complex-difference image
(CD). The magnitude weight equalizes the velocity noise errors, which is higher
in areas of low signal (e.g. air). The
pseudo complex difference additionally weights the data by velocity magnitude
aiding in the class imbalance between vessels and background tissue. Following
training, 3-directional components of the 3-point and 4-point velocities were
compared voxel-wise within vessels to determine their correlation and agreement.
The comparison was also performed outside the vessels to examine possible areas
of higher discrepancies. Results
The standard
4-point velocities and ML-estimated 3-point velocities showed high agreement,
indicating the CNN successfully estimated the background phase and 3 velocity
components from three flow encodings only (Figure 1). There were high
correlations (~0.98) between 3-pt and 4-pt velocities regardless of the choice
of loss function, indicating the high precision of ML-estimated velocities. Accuracy
of the 3-point velocities was dependent on the choice of loss function. With
magnitude weighting alone, velocities were slightly underestimated
(slope=0.925). This finding was concomitant with an observed reduction in the
background signal noise levels and a small degree of blurring in the velocity
images (Figure 2). When the training loss was replaced with CD-weighted least
squares, the underestimation was significantly reduced from 7.5% to 2.9% (slope=0.971).
CD-weighted images also show similar background noise levels and resolution to
the input 4-point velocity images. The
difference map between the 3-pt and 4-pt velocities show that the two
velocities are in high agreement within vessels but tend to show small discrepancies immediately
outside large vessels, such as internal carotid arteries (ICA) and middle cerebral
arteries (MCA) (Figure 3). However, the velocity discrepancies did not
translate to discrepancies in flow-derived angiogram, as the velocity error was
suppressed by the magnitude image during angiogram generation (Figure 4). Discussion
Our results
demonstrate that CNNs can estimate velocities in 4D-flow using only three flow
encodings without utilizing the reference encoding. These results have
potential implications that 4D-flow MRI can be performed without a dedicated
reference encoding, resulting in 25% reduction in scan time. While the
ML-estimated 3-pt velocities show slight underestimation in velocities and
hence blurring in the velocity images, such errors were significantly reduced
with greater weighting of vessels during training. With an optimized choice of
vessel weighting, the discrepancy between the 3-point and 4-point velocities
can potentially be reduced within the velocity noise of the 4D-flow MRI.Acknowledgements
No acknowledgement found.References
1. Johnson et al. MRM 2012 63(2): 349-355.
2. Ebbers et al. MRM 2001; 45(5):872-879.
3. Rasmussen et al. bioRxiv 278036.
4. Gu et al. AJNR 2005 26(4):743-9
5. Ronneberger et al. arXiv: 1505.04597 7