Gian Franco Piredda1,2,3, Punith B. Venkategowda4, Piotr Radojewski5,6, Tom Hilbert1,2,3, Arun Joseph6,7,8, Gabriele Bonanno6,7,8, Roland Wiest5,6, Karl Egger9, Shan Yang9, Jean-Philippe Thiran2,3, Ricardo A. Corredor-Jerez1,2,3, Bénédicte Maréchal1,2,3, and Tobias Kober1,2,3
1Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, 2Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, 3LTS5, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, 4Siemens Healthcare Pvt. Ltd., Bangalore, India, 5Support Center for Advanced Neuroimaging, Institute for Diagnostic and Interventional Neuroradiology, Inselspital, Bern University, Bern, Switzerland, 6Translational Imaging Center, sitem-insel AG, Bern, Switzerland, 7Advanced Clinical Imaging Technology, Siemens Healthcare AG, Bern, Switzerland, 8Magnetic Resonance Methodology, Institute of Diagnostic and Interventional Neuroradiology, University of Bern, Bern, Switzerland, 9Department of Neuroradiology, Medical Center – University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany
Synopsis
Large spatial signal variations due to field
inhomogeneities complicate the application of automated brain morphometry at
7T. In this work, we propose to use transfer learning to adapt a template-based
segmentation algorithm to sub-millimeter ultra-high field applications. More specifically, a
convolutional neural network pre-trained on T1-weighted
scans to extract the total intracranial volume (TIV) from MP-RAGE acquisitions was re-trained to retrieve the TIV mask directly from MP2RAGE volumes.
The developed method proved to reliably deliver brain tissue masks and
volumetry at 7T.
Introduction
Automated brain morphometry from MR
images has been shown to deliver promising neuroimaging biomarkers to monitor
structural changes during aging and the progression of neurodegenerative
disease. Most of the tools available for automatically analyzing structural
brain MR images has been optimized for 3D T1-weighted
contrasts1–3, with the MP-RAGE being the most
commonly used sequence at 1.5T and 3T4. For clinical applications at 7T,
the MP2RAGE sequence is preferred as it provides a bias-free 3D T1-weighted
image (“UNI”)5. However, the salt-and-pepper
noise surrounding brain tissues in the UNI images affects automated
skull-stripping methods, often resulting in inaccurate masks of the total
intracranial volume (TIV). A
solution to this particular problem has been proposed by deriving a
MP-RAGE-like image (“combT1w”) through multiplying the MP2RAGE UNI
and GRE2 volumes6. However, at 7T, this solution is
not ideal since the GRE2 suffers from large spatial signal
variations due to B1+ and B1-
inhomogeneities that are difficult to correct and may result in wrong
segmentations. Alternatively, a simple modification to the MP2RAGE normalized
complex ratio was also proposed to suppress the noise in the UNI image7. However, such modification
requires the tuning of a correction factor which is impractical in a clinical
scenario7.
In this work, we propose a new
pipeline for automated brain morphometry at 7T from MP2RAGE acquisitions using
an in-house built prototype software, MorphoBox3, combined with a convolutional
neural network (CNN) that performs brain skull-stripping directly on the UNI
image. The developed pipeline was compared and validated against the original
algorithm both at 3T and 7T. Methods
Study population and MR protocol
Two cohorts of healthy individuals were
scanned at two different field strengths:
- 201$$$\,$$$subjects (123$$$\,$$$females, age$$$\,$$$=$$$\,$$$[20-64]$$$\,$$$y/o) at 3T (MAGNETOM Prisma, Siemens Healthcare,
Erlangen, Germany);
- 19$$$\,$$$subjects (11$$$\,$$$females, age$$$\,$$$=$$$\,$$$[15-72]$$$\,$$$y/o) at 7T (MAGNETOM Terra, Siemens
Healthcare, Erlangen, Germany).
Written informed consent was obtained prior
to each examination. Data was also collected from three patients (two females,
age$$$\,$$$=$$$\,$$$[25-67]$$$\,$$$y/o) who were scanned at both field strengths, in agreement with
the institutional regulations.
The MP2RAGE sequence was acquired in every
subject using the protocol parameters detailed in Table$$$\,$$$1.
Brain segmentation
The template-based method used by MorphoBox
for segmenting the TIV was replaced by a CNN-based
algorithm recently proposed by Venkategowda et al.
8.
The CNN was trained on more than 400 T
1-weighted scans to extract
the TIV from MP-RAGE acquisitions
8.
Here, a transfer learning strategy was employed to achieve the same goal for
MP2RAGE UNI data. More specifically, the CNN algorithm pre-trained on MP-RAGE
data was re-trained to retrieve the TIV mask directly from UNI volumes. To that end, ground truth masks were
obtained by segmenting MP2RAGE 3T data with the original pipeline of MorphoBox
on the respective combT
1w image. The network was re-trained for 200
epochs using the Dice coefficient between the predicted and target masks as
loss function. Images from 181 of the subjects scanned at 3T were used for the
training.
After extracting the TIV, brain volumes were
segmented on the UNI images using the original MorphoBox pipeline
3.
Validation Volumes of brain regions estimated with the
original pipeline (i.e., TIV extraction on the combT
1w) and the new
proposed solution (i.e., TIV extraction directly on the UNI) were computed in
the remaining 20 testing datasets acquired at 3T and compared via Wilcoxon
rank-sum tests. P-values were corrected for multiple comparisons with the
Benjamini-Hochberg procedure.
Segmentation masks produced by the original
and new pipeline applied to the 7T data were visually compared. Masks obtained
in patients scanned at both field strengths were also visually compared.
Results
Representative sagittal slices of the
retrieved segmentation masks are shown in Figure$$$\,$$$1 for three subjects scanned
at 3T. At visual inspection, the new pipeline delivered reliable masks that are
comparable to the original pipeline. Quantitatively, no significant volume
differences were found in any of the segmented brain regions when comparing the
two pipelines (see Table$$$\,$$$2).
Segmentation masks obtained from 7T data
acquired in healthy subjects are shown in Figure$$$\,$$$2. In this case, TIV masks are
wrongly estimated by the original pipeline, especially in regions with large
intensity variations in the combT1w contrast (e.g., frontal lobe,
temporal lobe, and cerebellum). Conversely, TIV masks were correctly segmented
with the proposed pipeline.
In comparison to segmentation masks obtained
at 3T in the same patient, masks obtained from 7T data visually appear to
provide a better definition of tissue boundaries and small structures, as for
instance the cerebellum (see Figure$$$\,$$$3).Discussion and Conclusion
This work introduced a new method for
automated brain morphometry at 7T. Using transfer learning from a CNN algorithm
pre-trained on 3T MP-RAGE data, TIV masks were directly retrieved from 7T
MP2RAGE UNI data. The proposed strategy was quantitatively validated by
comparing the estimated volumes with those segmented from the original pipeline
at 3T showing no significant differences. Future work should focus on
validating the developed pipeline in a larger cohort of patients. Additionally,
the benefit of performing MP2RAGE acquisitions in combination with parallel
transmission at 7T should be investigated as it allows to reduce spatial signal
variations.
The higher resolution of 7T MRI combined
with the proposed automated morphometry may enable more accurate and
reproducible monitoring of brain volumes both in cross-sectional and
longitudinal studies.Acknowledgements
No acknowledgement found.References
1. Ashburner J. Computational anatomy with the SPM software. Magn Reson Imaging.
2009;27(8):1163-1174. doi:10.1016/j.mri.2009.01.006
2. Fischl B. FreeSurfer. Neuroimage. 2012;62(2):774-781.
doi:10.1016/j.neuroimage.2012.01.021
3. Schmitter
D, Roche A, Maréchal B, et al. An evaluation of volume-based morphometry for
prediction of mild cognitive impairment and Alzheimer’s disease. NeuroImage
Clin. 2015;7:7-17. doi:10.1016/j.nicl.2014.11.001
4. Mugler JP,
Brookeman JR. Three-dimensional magnetization-prepared rapid gradient-echo
imaging (3D MP RAGE). Magn Reson Med. 1990;15(1):152-157.
doi:10.1002/mrm.1910150117
5. Marques
JP, Kober T, Krueger G, van der Zwaag W, Van de Moortele P-F, Gruetter R.
MP2RAGE, a self bias-field corrected sequence for improved segmentation and
T1-mapping at high field. Neuroimage. 2010;49(2):1271-1281.
doi:10.1016/j.neuroimage.2009.10.002
6. Fujimoto
K, Polimeni JR, van der Kouwe AJW, et al. Quantitative comparison of cortical
surface reconstructions from MP2RAGE and multi-echo MPRAGE data at 3 and 7T. Neuroimage.
2014;90:60-73. doi:10.1016/j.neuroimage.2013.12.012
7. O’Brien
KR, Kober T, Hagmann P, et al. Robust T1-weighted structural brain imaging and
morphometry at 7T using MP2RAGE. PLoS One. 2014;9(6).
doi:10.1371/journal.pone.0099676
8. Venkategowda
PB, Kumaraswamy AK, Richiardi J, et al. Retrofitting a Brain Segmentation
Algorithm with Deep Learning Techniques: Validation and Experiments. In: Proceedings
of the International Society of Magnetic Resonance Imaging. 2020. Abstract number: 3566.