Alexander German1, Angelika Mennecke1, Jan Martin2, Jannis Hanspach1, Andrzej Liebert1, Jürgen Herrler1, Tristan Anselm Kuder3, Manuel Schmidt1, Armin Nagel1, Michael Uder1, Arnd Dörfler1, Jürgen Winkler1, Moritz Zaiss1,4, and Frederik Laun1
1University Hospital Erlangen, Erlangen, Germany, 2Lund University, Lund, Sweden, 3German Cancer Research Center, Heidelberg, Germany, 4Max Planck Institute for Biological Cybernetics, Tübingen, Germany
Synopsis
We acquired diffusion-weighted
and CEST data of the brain of 38 healthy volunteers. A MPRAGE and SWI based segmentation
into 102 brain regions revealed unique diffusion and chemical MR signals on
average. More importantly, we could infer these tissue classes form individual voxel
data using a neural network. The revival of this old paradigm for tissue
characterization from the 1990s points to the fact that unique MR signals of
different brain regions exist and can be used to determine the tissue type
voxel-wise. The approach as such is general and could unify the ever-growing
diversity of MR contrasts.
Introduction
Since the seminal works by
Brodmann and contemporaries, it is well-known that different brain regions
exhibit unique cytoarchitectonic and myeloarchitectonic features.1 Transferring the approach of classifying brain tissues
based on their intrinsic features to the realm of magnetic resonance (MR) is a
longstanding endeavor.2 We explore the feasibility of performing global brain
classification based on intrinsic MR features.Methods
We
recruited 38 participants who provided written informed consent to participate
in the study, which was approved by the local ethics committee. MRI acquisition
and processing scans were acquired using a 7T MRI scanner (Magnetom Terra,
Siemens Healthineers AG, Erlangen) with a 32-channel receive head coil and
8-channel parallel-transmit coil (ptx). Each scanning session consisted of a high-resolution
3D T1 MPRAGE scan and a gradient
echo sequence with ASPIRE coil combination to generate susceptibility map-weighted
images (SMWI).3 Subsequently, CEST MRI was acquired using a
snapshot-GRE CEST sequence at two different saturation B1 levels, 0.7
and 1.0 μT, presaturated at 56 different offsets, respectively.4 For the Gaussian
pre-saturation pulse train, multiple interleaved mode saturation (MIMOSA) was
applied.5 Individual CEST peaks
were fitted voxel-wise after coregistration using SPM.6 Diffusion-weighted images were acquired
using a spin-echo EPI sequence7 with linear, planar, and spherical b-tensors and b-values 0, 100, 500,
1000, 1500, and 2000 s/mm2. The linear and planar b-tensors were
rotated in 16 noncollinear diffusion directions. QTI parameters8 were fitted
voxel-wise after coregistration using elastix.9 A
segmentation of each brain into 102 anatomical regions was achieved by merging
the segmentations obtained with automated cortical reconstruction and
volumetric segmentation with the Freesurfer10 image analysis suite and
manual segmentation on MPRAGE and SMWI for each participant. The classification approach is
visualized in Fig. 1. A dataset was created from all participants except for
one young man (the test participant). For each participant individually, 10% of
the voxels of each anatomical region were randomly chosen from the MPRAGE
space. For each chosen voxel, the local 15 QTI parameters, 210 raw b-tensor
values, local four CEST parameters and 112 raw Z-spectrum values were extracted
from the images, coregistered and interpolated to MPRAGE space and resolution
using FSL11, and were saved to a
2D array with 6 million rows and 341 columns. The dataset was then shuffled,
split, and normalized and no longer contained spatial or neighboring information.
We defined a TensorFlow Keras dense neural network12 to predict one out of
the 102 one-hot-encoded anatomical brain regions. The model had 52 million
trainable parameters with 50% dropout. After training, the network was used to
perform voxel-wise prediction for the test participant. The gross accuracy was
calculated.
To investigate the
operating principle of DNN classification, we computed the normalized features
of brain tissues averaged over all study participant and the saliency13 vectors averaged
over the respective regions of the test participant.Results
Figure 2 shows the
segmentation results obtained with our gold-standard segmentation. Fig. 3 shows
the class-wise probability maps including the midbrain nuclei. Accurate
detection of small structures in the midbrain like substantia nigra and nucleus
ruber were particularly surprising. The confusion matrix of classification
(Fig. 4) shows robust results for most subcortical brain nuclei and slight
mixing-up between different cortical fields, white matter areas and between
cortical field and subjacent white matter area, respectively. The gross
accuracy was 60%.
The confusions matched with
explicit microstructural and chemical traits of brain regions, which showed
highly individual patterns in brain nuclei and only subtly distinct patterns in
the cortical fields and white matter areas (Fig. 5a). On average, each brain
region has a unique signature of QTI and CEST parameters computed from the
image raw data (Fig. 5a), which are just a subset of the input spectrum that
also contained the raw image data.
Fig. 5b shows the saliency
vector for every region of the test dataset. For the precentral cortex (Fig.1
and region 48 in Fig. 5b), the DNN relied predominantly on magnetization
transfer and very high offsets and the water pool resonance. Concordant
saliency peaks at -4 ppm in both z-spectra shows that the DNN has discovered
offset domains of the intramolecular NOE as a feature.Discussion
Extending earlier works on
global brain classification,2 we included novel high-dimensional contrasts into
the input data space. With the single-voxel classification approach, the spatial
coherence of prediction results in adjacent voxels serves as an inherent metric
of reliability (such as in Fig. 2). Current state-of-art atlas-based
classifications approaches outperform the accuracy we observed.14 However, it is tempting to consider our MR-based
patterns as analogous to histological tissue fingerprints. Potential confounders
can be B0 or B1 inhomogeneities, which increase at 7T
compared to lower field strengths. In the CEST preparation, we addressed the B1
inhomogeneity by means of the MIMOSA approach. In addition, B0 and B1
correction was performed employing acquired field maps. The approach as such is
general and can be extended to unify the ever-growing diversity of MR
contrasts.Conclusion
We show that mapping from
voxel-intrinsic MR data to the brain tissue class to which the data belongs is
possible. This indicates that unique MR signals of different brain regions
exist. Acknowledgements
We thank Jasmin Burczyk, Molecular Neurology, for her support in providing contact to patients and Matthias Schäfer, Computer Science 7, for providing computational resources. This work was supported by the Interdisciplinary Center for Clinical Research Erlangen (IZKF-Erlangen, ELAN DR-17-12-14-1). AG received a MD thesis fellowship by the IZKF Erlangen. JW is a member of the research training group 2162 “Neurodevelopment and Vulnerabilityof the Central Nervous System” funded by the DFG–270949263/GRK2162. The work was further supported by a Heisenberg professorship from the DFG to FBL (DFG LA2804/6-1 and 2804/12-1).References
- Brodmann
K. Vergleichende Lokalisationslehre der Großhirnrinde : in ihren Prinzipien
dargestellt auf Grund des Zellenbaues. Leipzig: Johann
Ambrosius Barth. 1909;X, 324 S. : Ill.
- Ozkan M, Dawant B M
& Maciunas R J. Neural-network-based segmentation of multi-modal medical
images: a comparative and prospective study. IEEE Transactions on Medical
Imaging. 1993;12:534-544.
- Eckstein
K, Dymerska B, Bachrata B et al. Computationally Efficient Combination of
Multi-channel Phase Data From Multi-echo Acquisitions (ASPIRE). Magnetic
resonance in medicine. 2018;79:2996-3006.
- Zaiss
M, Ehses P & Scheffler K. Snapshot-CEST: Optimizing
spiral-centric-reordered gradient echo acquisition for fast and robust 3D CEST
MRI at 9.4 T. NMR in Biomedicine. 2018;31:e3879.
- Liebert
A, Zaiss M, Gumbrecht R et al. Multiple interleaved mode saturation (MIMOSA)
for B1+ inhomogeneity mitigation in chemical exchange saturation transfer.
Magnetic resonance in medicine. 2019;82:693-705.
- Friston
K J. Imaging neuroscience: Principles or maps? Proceedings of the National
Academy of Sciences. 1998;95:796-802.
- Martin J, Endt S,
Wetscherek A et al. Contrast-to-noise ratio analysis of microscopic diffusion
anisotropy indices in q-space trajectory imaging. Z Med Phys. 2020;30:4-16.
- Westin
C-F, Knutsson H, Pasternak O et al. Q-space trajectory imaging for
multidimensional diffusion MRI of the human brain. NeuroImage.
2016;135:345-362.
- Nilsson M,
Szczepankiewicz F, van Westen D & Hansson O. Extrapolation-Based References
Improve Motion and Eddy-Current Correction of High B-Value DWI Data:
Application in Parkinson's Disease Dementia. PloS one. 2015;10:e0141825.
- Fischl
B. FreeSurfer. NeuroImage. 2012;62:774-781.
- Jenkinson M &
Smith S. A global optimisation method for robust affine registration of brain
images. Med Image Anal. 2001;5:143-156.
- Abadi
M, Agarwal A, Barham P et al. TensorFlow: Large-scale machine learning on
heterogeneous systems. 2015.
- Simonyan
K, Vedaldi A & Zisserman A. Deep Inside Convolutional Networks: Visualising
Image Classification Models and Saliency Maps. arXiv:1312.6034 (2013). <https://ui.adsabs.harvard.edu/abs/2013arXiv1312.6034S>.
- Iglesias
J E & Sabuncu M R. Multi-atlas segmentation of biomedical images: A survey.
Med Image Anal. 2015;24:205-219.