Borjan Gagoski1,2, Junshen Xu3, Paul Wighton4, Dylan Tisdall5, Robert Frost2,4, Sayeri Lala6, Wei-Ching Lo7, Polina Golland8,9, Andre van der Kouwe2,4, Elfar Adalsteinsson8,10, and P. Ellen Grant1,2
1Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children's Hospital, Boston, MA, United States, 2Department of Radiology, Harvard Medical School, Boston, MA, United States, 3(co-first author) Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, 4Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, United States, 5Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, United States, 6Department of Electrical Engineering, Princeton University, Princeton, NJ, United States, 7Siemens Medical Solutions USA, Inc, Charlestown, MA, United States, 8Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, 9Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, Cambridge, MA, United States, 10Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States
Synopsis
Fetal brain MRI
suffers from unpredictable and unconstrained fetal motion that not only causes
severe image artifacts even with single-shot FSE readouts, but also results in
slice-to-slice variations of the imaging plane and long scanning sessions, as
the MR technologist “chases” the fetal head in an attempt to acquire
artifact-free orthogonal images. In this work, we have implemented a
closed-loop pipeline that automatically detects and reacquires HASTE images
that were degraded by fetal motion, without any interaction from the MRI
technologist. The presented methods demonstrate the basic infrastructure needed
for successful prospective automated fetal brain motion correction.
Introduction/Purpose
Fetal brain MRI
suffers from unpredictable fetal motion, limiting the types of MR acquisition
schemes practical for use in clinical settings to fast single-shot encoding
techniques, such as HASTE. Nevertheless, even when using HASTE, fetal MRI sessions can be
significantly lengthened, as fetal motion during the HASTE acquisition not only
degrades image quality, but also introduces slice-to-slice variations of the
imaging plane, resulting in double-oblique slices which are hard to interpret
clinically. Therefore, in a typical fetal MRI session, the MRI technologist
“chases” the fetal head, in an attempt to obtain images in the standard three
orthogonal planes, resulting in repetition of the entire HASTE acquisition if
sufficient slices in the stack are motion degraded or became double oblique.
Thus, the current fetal imaging workflow depends on rapid assessment of the
image stack quality and determination if it should be repeated. In this work,
we designed and implemented a prototype closed-loop acquisition/reconstruction pipeline
that automatically, without human interaction, detects and re-acquires only the slices that have been
degraded by fetal motion, and not the whole stack.Methods
We have previously shown the feasibility of
using a convolutional neural network (CNN) engine that performs an automatic
image quality assessment (IQA) of fetal brain HASTE images, and outputs
a quality score1. In this
work we developed and implemented a pipeline that runs the IQA CNN engine on a
GPU (NVIDIA 1050Ti) equipped computer, connected to the scanner’s internal
network via 1GB Ethernet hub, for efficient communication/feedback between this
computer and the scanner's computer running our custom MRI sequence and reconstruction. This setup is similar to those used for real-time neurofeedback in
fMRI experiments2–4 and
prospective motion correction in neuroanatomical MRI5,6. The proposed methods involved modifying the
HASTE sequence acquisition and reconstruction, as well as generating Python
scripts that in real time receive the HASTE images, run the CNN IQA engine, and
send the IQA score of each slice back to the sequence. Figure 1 shows a schematic of the
overall acquisition/reconstruction engine, described below.
HASTE acquisition modification: We used an enhanced HASTE sequence, called vNav-HASTE,
which embeds low resolution EPI volumetric navigators (EPI-vNavs)7
within the TR to obtain a low resolution volume of the fetal head (5mm3
voxels, 3D EPI readout, TA=0.7s) before every acquired HASTE slice8. For this project, we further modified this
sequence to enable: 1. socket connection to the GPU computer for seamless
transfer of the IQA scores upon sequence request during run-time; 2.
reacquisition loop that starts seamlessly (without any human interaction) right
after the prescribed HASTE stack is acquired, and reacquires a user defined
number of slices, NREACQ, with the NREACQ worst IQA
scores.
HASTE reconstruction modification: The online reconstruction
was modified such that HASTE images are sent to the GPU computer via a socket
connection as soon as they are reconstructed. Additionally, besides the
original HASTE stack, the reconstruction engine outputs the NREACQ
reacquired images in a separate image series.
IQA scores calculation: We trained a VGG-16 network (Figure 2a) to
classify HASTE images as diagnostic or non-diagnostic. The network is first
pre-trained on Imagenet dataset and then fine-tuned on a fetal HASTE dataset
(4432/1557 diagnostic/nondiagnostic images). To address the problem of class
imbalance in MRI dataset, we adopted weighted binary cross entropy as loss
function during fine-tuning. The network was evaluated on a separate test set
with 1329 slices, see ROC curve in Figure 2b. With an image with size of
256x256, each IQA score is computed in 30 ms on the GPU computer used.
Phantom development and in vivo fetal scans were
performed on a 3T Skyra scanner (Siemens Healthcare, Erlangen, Germany) using spine and
body flex receive arrays. The vNav-HASTE-reacq sequence (TE/TR=119ms/1.8s, FOV=33x33cm2, 1.3x1.3x3mm3 voxels,
PF=5/8, RGRAPPA=2) was run on two pregnant mothers who signed
informed consent forms approved by BCH’s IRB.Results
Figure 3 shows images
from vNav-HASTE-reacq scan with N=10 slices and NREACQ=6 slices to
reacquire, showing that our sequence correctly reacquired the slices at locations
where the 6 slices with the lowest IQA scores (shown in red) were originally
acquired. Figure 4 shows 4 images from 3 separate scans, where the originally
acquired slices were motion degraded, and the re-acquired ones were not. The
IQA scores shown above the images range from 0 to 1 (0=poor, 1=excellent quality).Discussion/Conclusion
We have demonstrated
a closed-loop acquisition/reconstruction pipeline that automatically detects
and reacquires motion degraded fetal brain HASTE images. The proposed
infrastructure has all the necessary infrastructural capabilities for robust
real-time, prospective fetal head motion correction. Specifically, our future work
includes developing a computational engine capable of efficiently and reliably
determining the fetal head pose from the low resolution EPI-vNavs obtained
before each HASTE slice. The communication framework described here can easily
be adapted to include this information for appropriate reorientation of the FOV
of the next HASTE slice. Moreover, we intend to further improve the accuracy
performance of the IQA CNN engine. This work is part of our greater effort and
ultimate goal of developing an intelligent system that “chases” the fetal head
in real-time to obtain high quality, high resolution, diagnostic HASTE images
of the entire brain in the least amount of time.Acknowledgements
NIH R01 EB017337, U01 HD087211, R01HD100009,
R00HD074649, R01HD099846, R01HD093578, R01HD085813, The NVIDIA Corporation.References
1 Lala S, Singh N, Gagoski B, Turk E,
Grant PE, Golland P, Adalsteinsson E. A Deep Learning Approach for Image
Quality Assessment of Fetal Brain MRI. 27th Annual
Proceedings of the International Society for Magnetic Resonance in Medicine 2017.
2 Hinds O, Ghosh S, Thompson TW, Yoo
JJ, Whitfield-Gabrieli S, Triantafyllou C, et al. Computing moment-to-moment BOLD
activation for real-time neurofeedback. Neuroimage 2011;54:361–8.
3 Lee J, Wighton P, Cauley SF,
Setsompop K, van der Kouwe A, Loggia ML, et al. Application of Simultaneous
Multi-Slice (SMS) Imaging to Real-time fMRI for Improved Neurofeedback Signal
Fidelity. Real-Time Functional Imaging and Neurofeedback
(rtFIN) Conference 2015.
4 Wighton P, Karahanoglu FI, Tisdall,
van der Kouwe A. Slice-by-Slice prospective motion correction in EPI using a
markerless motion tracking system. International
Society for Magnetic Resonance in Medicine’s (ISMRM) Workshop on Motion
Correction in MRI 2017.
5 Gilman J, Wighton P, Curran MT, Lee
S, Thompson T, De Los Angeles CS, et al. Modulation of Visual Attention of
Blended Faces and Scenes in the FFA and PPA. Real-Time
Functional Imaging and Neurofeedback (rtFIN) Conference 2015.
6 Frost R, Wighton P, Karahanoğlu FI,
Robertson RL, Grant PE, Fischl B, et al. Markerless high-frequency
prospective motion correction for neuroanatomical MRI. Magn Reson Med 2019;82:126–44.
7 Tisdall MD, Hess AT, Reuter M,
Meintjes EM, Fischl B, van der Kouwe AJW. Volumetric navigators for prospective
motion correction and selective reacquisition in neuroanatomical MRI. Magn Reson Med 2012;68:389–99.
8 Gagoski B, McDaniel P, van der Kouwe
A, Bhat H, Wald L, Adalsteinsson E, et al. HASTE Imaging with EPI Volumetric
Navigators for Real-Time Fetal Head Motion detection. 24th Annual
Proceedings of the International Society for Magnetic Resonance in Medicine 2016.