Isabelle Heukensfeldt Jansen1, Sangtae Ahn1, Rafi Brada2, Michael Rotman2, and Christopher J. Hardy1
1GE Global Research Center, Niskayuna, NY, United States, 2GE Global Research Center, Herzliya, Israel
Synopsis
We introduce ISHMAPS,
a method for detecting and adapting to patient motion in real time during an MR
scan. The method uses a neural network trained on motion-corrupted data to
detect and score motion using as little as 6% of k-space. Once motion is
detected, multiple separate complex sub-images from different motion states can
be reconstructed and combined into a motion-free image, or the scan can adaptively
re-acquire sections of k-space taken before motion occurred.
Purpose
Patient motion is one
of the biggest sources of inefficiency in clinical MRI, often requiring
re-scans or even second visits by the patient1. Most current approaches to
motion correction require either some sort of hardware for monitoring the
motion (adding to cost and patient setup time), or navigator sequences (which
take time away from the imaging sequence). We created a scoring system to train
a convolutional network to detect motion in sub-images reconstructed from
partial k-space to detect the timing as well as severity of motion while a scan
is in progress. The ISHMAPS (Inter-Shot Motion Artifact Predictive Scoring) network
was able to detect motion within a window of as little as 6% of k-space.Methods
The ISHMAPS neural
network was trained to continuously identify and score motion severity during
an MRI scan. Various acquisition protocols were considered including multi-shot
fast spin echo (FSE), with echo-train length (ETL) ranging from 8 to 23, as
exemplified in Fig. 1. Two neighboring shots, each comprising ETL k-space lines,
were fed into the neural network, to predict whether motion had occurred
between shots, and the window was then slid by one shot to assess the next
pairing. The network was trained with use of a dataset of simulated in-plane
rigid-body motion. The dataset was generated starting with a series of
motion-free head images. For each image, a second image was created by randomly
translating (by up to 10 pixels in any direction) and/or rotating (by angles
between +/- 10 degrees centered around the back of the head) the initial image.
Both original and moved images were multiplied by coil sensitivity maps and
Fourier transformed into k-space. K-space
lines from the first shot of a pair were filled with pre-motion data, and the
second shot with post-motion. Zero filling and Fourier transformation back into
the image domain created a multi-coil complex “sub-image” with motion artifacts
(Fig. 2). For each sub-image, the corresponding motion-free sub-image was created,
and a motion corruption score calculated based on entropy-of-the-difference
between the two. This metric takes normalized difference images, multiplies by
the log of the difference, and averages the value over all pixels and coils.
The convolutional neural network shown in Fig. 3 was then trained to predict
the motion-corruption score by comparing it to the labeled score in the cost
function. The ISHMAPS network was trained on 1713 datasets using random
translations and rotations, all combinations of adjacent shot pairings, and a
variety of ETLs. The network was validated and tested on 151 and 183 datasets,
respectively. The network was then used to generate continuous motion scores
from a subject moving his head in specified patterns during scanning. Relatively
artifact-free images were reconstructed using an iterative reconstruction
algorithm similar to [2], with the motion model constrained using the motion
timing information generated by ISHMAPS. In other cases, adaptive scanning was
performed by selectively reacquiring regions of k-space after detecting motion events.Results
Figure 4a,b shows the
results of the network prediction against the ground-truth scores. An ROC curve
was generated for motion detection by this method, with the area under the
curve found to be 0.98. With choice of an appropriate threshold (dashed line,
Fig. 4), the network prediction becomes a classifier determining whether
significant motion occurred in the sub-image or not. An example of volunteer head
motion is shown in Fig. 4c, with the resulting continuous output of the network
shown in Fig. 4d, and a motionless scan in Fig. 4e. Figure 5a-d shows an
example of a motion-free image reconstructed using motion timings derived from
the network. Figure 5e-g shows an example of adaptive scanning based on network
detection of motion events during scanning.Discussion
Training a network to
predict a motion score rather than just training for motion classification enables
a more nuanced approach to ameliorating motion’s effects. By simultaneously
determining if/when motion has occurred and the severity of motion artifacts,
algorithms can use this information to make decisions on appropriate steps to reduce
or remove the artifacts, including reacquiring portions of k-space to
compensate, combining multiple motion-free sub-images created from regions
acquired between detected motion using a neural network or similar
reconstruction, or, in case of severe motion, restarting the scan after
instructing or sedating the patient.Conclusion
We created a scoring
system to train a convolutional network to detect motion artifacts in
sub-images reconstructed from partial k-space, to detect the timing as well as
severity of motion while a scan is in progress. Using a score based on entropy
between a motion-corrupted and ground-truth sub-image allows the network to
predict the degree of motion corruption. The network was able to predict motion
from a sub-image within 8 phase-encodes (3%) of the motion occurring.Acknowledgements
No acknowledgement found.References
1. J Andre,
et al. Toward quantifying the prevalence, severity, and cost associated with
patient motion during clinical MR examinations. JACR, 2015;12:689695.
2. L Cordero-Grande, et
al. Sensitivity encoding for aligned multishot magnetic resonance
reconstruction. IEEE Trans Comput Imaging, 2016;2(3):266-280.